US20130325148A1 - Efficient quadratic programming (qp) solver for process control and optimization - Google Patents
Efficient quadratic programming (qp) solver for process control and optimization Download PDFInfo
- Publication number
- US20130325148A1 US20130325148A1 US13/485,594 US201213485594A US2013325148A1 US 20130325148 A1 US20130325148 A1 US 20130325148A1 US 201213485594 A US201213485594 A US 201213485594A US 2013325148 A1 US2013325148 A1 US 2013325148A1
- Authority
- US
- United States
- Prior art keywords
- solution
- computer readable
- constraint
- matrix
- violated
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000005457 optimization Methods 0.000 title description 18
- 238000004886 process control Methods 0.000 title description 10
- 238000000034 method Methods 0.000 claims abstract description 111
- 230000008569 process Effects 0.000 claims abstract description 58
- 230000000295 complement effect Effects 0.000 claims abstract description 27
- 239000011159 matrix material Substances 0.000 claims description 35
- 238000004519 manufacturing process Methods 0.000 claims description 13
- 238000004590 computer program Methods 0.000 claims description 9
- 238000003860 storage Methods 0.000 claims description 7
- 230000001419 dependent effect Effects 0.000 claims description 4
- 230000000977 initiatory effect Effects 0.000 claims 1
- 239000000123 paper Substances 0.000 description 51
- 239000000243 solution Substances 0.000 description 48
- 238000003491 array Methods 0.000 description 15
- 238000004422 calculation algorithm Methods 0.000 description 9
- 239000000047 product Substances 0.000 description 8
- 239000013598 vector Substances 0.000 description 6
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 6
- 230000009977 dual effect Effects 0.000 description 5
- 230000006870 function Effects 0.000 description 5
- 238000004891 communication Methods 0.000 description 4
- 238000010790 dilution Methods 0.000 description 4
- 239000012895 dilution Substances 0.000 description 4
- 239000000463 material Substances 0.000 description 4
- 239000000725 suspension Substances 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 238000010438 heat treatment Methods 0.000 description 3
- 230000006698 induction Effects 0.000 description 3
- 238000005259 measurement Methods 0.000 description 3
- 230000004075 alteration Effects 0.000 description 2
- 230000003190 augmentative effect Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000001035 drying Methods 0.000 description 2
- 239000000945 filler Substances 0.000 description 2
- 229920002522 Wood fibre Polymers 0.000 description 1
- 239000000654 additive Substances 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 239000011111 cardboard Substances 0.000 description 1
- 239000011248 coating agent Substances 0.000 description 1
- 238000000576 coating method Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 239000000975 dye Substances 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 239000000835 fiber Substances 0.000 description 1
- 239000012467 final product Substances 0.000 description 1
- 239000011888 foil Substances 0.000 description 1
- 238000009472 formulation Methods 0.000 description 1
- 238000012804 iterative process Methods 0.000 description 1
- 230000014759 maintenance of location Effects 0.000 description 1
- 239000002184 metal Substances 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 239000011087 paperboard Substances 0.000 description 1
- 239000004033 plastic Substances 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 239000004753 textile Substances 0.000 description 1
- 238000012876 topography Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
- 239000002025 wood fiber Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
- G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
- G05B13/04—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators
- G05B13/048—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators using a predictor
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
Definitions
- This disclosure relates generally to control and optimization systems. More specifically, this disclosure relates to an efficient quadratic programming (QP) solver for process control and optimization.
- QP quadratic programming
- Model predictive control is a popular technique for controlling multi-input and multi-output processes, such as industrial manufacturing processes. MPC uses a model to predict how one or more controlled process variables are expected to behave in the future. Changes can then be made to one or more manipulated process variables in order to alter the controlled process variable(s). Ideally, each controlled process variable is thereby maintained within a desired range.
- An MPC controller often implements an online quadratic programming (QP) solver for solving an optimization problem related to a controlled process.
- QP quadratic programming
- process processes may involve hundreds of manipulated process variables and thousands of controlled process variables (many with active limits and rate constraints).
- control intervals can be relatively short, such as ten to twenty seconds.
- generic and custom QP solvers have been developed, they often suffer from various shortcomings.
- This disclosure provides an efficient quadratic programming (QP) solver for process control and optimization.
- QP quadratic programming
- a method in a first embodiment, includes identifying an initial solution to a quadratic programming (QP) problem associated with a process. The method also includes performing an iterative procedure having one or more iterations. Each iteration includes determining whether any constraint associated with the process is violated in the solution. If so, each iteration also includes selecting a violated constraint, determining a step direction and a step length associated with the selected violated constraint, and updating the solution based on the step direction and the step length. Determining the step direction and the step length includes using a Schur complement based on an active set of constraints associated with the solution. The Schur complement is nonsingular during all iterations of the iterative procedure except when the active set is empty.
- QP quadratic programming
- an apparatus in a second embodiment, includes at least one memory configured to store an initial solution to a quadratic programming (QP) problem associated with a process.
- the apparatus also includes at least one processing device configured to perform an iterative procedure having one or more iterations. During each iteration, the at least one processing device is configured to determine whether any constraint associated with the process is violated in the solution. If so, the at least one processing device is also configured to select a violated constraint, determine a step direction and a step length associated with the selected violated constraint, and update the solution based on the step direction and the step length.
- the at least one processing device is configured to determine the step direction and the step length using a Schur complement based on an active set of constraints associated with the solution. The Schur complement is nonsingular during all iterations of the iterative procedure except when the active set is empty.
- a computer readable medium embodies a computer program.
- the computer program includes computer readable program code for identifying an initial solution to a quadratic programming (QP) problem associated with a process.
- the computer program also includes computer readable program code for performing an iterative procedure having one or more iterations. Each iteration includes determining whether any constraint associated with the process is violated in the solution. If so, each iteration also includes selecting a violated constraint, determining a step direction and a step length associated with the selected violated constraint, and updating the solution based on the step direction and the step length.
- QP quadratic programming
- the computer readable program code for determining the step direction and the step length includes computer readable program code for using a Schur complement based on an active set of constraints associated with the solution.
- the Schur complement is nonsingular during all iterations of the iterative procedure except when the active set is empty.
- FIG. 1 illustrates an example sheet manufacturing or processing system according to this disclosure
- FIGS. 2 through 4 illustrate an example method for efficient quadratic programming (QP) solving for process control and optimization according to this disclosure
- FIG. 5 illustrates an example system using an efficient QP solver for process control and optimization according to this disclosure.
- FIGS. 1 through 5 discussed below, and the various embodiments used to describe the principles of the present invention in this patent document are by way of illustration only and should not be construed in any way to limit the scope of the invention. Those skilled in the art will understand that the principles of the invention may be implemented in any type of suitably arranged device or system.
- a control or optimization problem in a given industry is solved using a general-purpose quadratic programming (QP) solver or a customized QP solver.
- QP quadratic programming
- general-purpose QP solvers often have various shortcomings.
- general-purpose QP solvers are intended to solve a broader class of problems and may not achieve performance that a customized QP solver could achieve.
- general-purpose QP solvers may not take advantage of a specific problem's structure and features for a given industry, which might otherwise significantly reduce the computation time needed to identify a problem solution.
- Customized QP solvers also often have various shortcomings, such as a poorly organized algorithm structure or defects that can result in complete control failure. A poorly organized structure makes the solver less reliable and makes it harder to diagnose problems (such as infeasible solutions).
- the QP solver disclosed here is based on a dual-feasible active-set algorithm, a Schur complement method, and a warm start strategy.
- the Schur complement is nonsingular throughout its iterations, which makes the QP solver very reliable numerically and helps to avoid control failures.
- the QP solver can also identify problem solutions much faster than conventional QP solvers.
- the efficient QP solver disclosed here solves a control or optimization problem by taking advantage of the problem's structure and features, which can vary depending on the specific application of the QP solver. By doing this, the QP solver can speed up solution computations and enable a controller to control a process closer to the process' physical limits, resulting in increased process efficiency and product quality.
- the QP solver is used for cross direction (CD) control in a paper-making process.
- CD cross direction
- the QP solver can generally be used in any other suitable device or system having at least one array of actuators that are controlled using a solution to a QP problem, such as any large-scale spatially-distributed system.
- FIG. 1 illustrates an example sheet manufacturing or processing system 100 according to this disclosure.
- the system 100 includes a paper machine 102 , a control system 104 , and a network 106 .
- the paper machine 102 includes various components used to produce a paper product, namely a paper sheet 108 that is collected at a reel 110 .
- the control system 104 monitors and controls the operation of the paper machine 102 , which may help to maintain or increase the quality of the paper sheet 108 produced by the paper machine 102 .
- the paper machine 102 includes at least one headbox 112 , which distributes a pulp suspension uniformly across the machine onto a continuous moving wire screen or mesh 113 .
- the pulp suspension entering the headbox 112 may contain, for example, 0.2-3% wood fibers, fillers, and/or other materials, with the remainder of the suspension being water.
- the headbox 112 may include an array of dilution actuators, which distributes dilution water into the pulp suspension across the sheet. The dilution water may be used to help ensure that the resulting paper sheet 108 has a more uniform basis weight across the sheet 108 .
- Arrays of drainage elements 114 such as vacuum boxes, remove as much water as possible to initiate the formation of the sheet 108 .
- An array of steam actuators 116 produces hot steam that penetrates the paper sheet 108 and releases the latent heat of the steam into the paper sheet 108 , thereby increasing the temperature of the paper sheet 108 in sections across the sheet. The increase in temperature may allow for easier removal of remaining water from the paper sheet 108 .
- An array of rewet shower actuators 118 adds small droplets of water (which may be air atomized) onto the surface of the paper sheet 108 . The array of rewet shower actuators 118 may be used to control the moisture profile of the paper sheet 108 , reduce or prevent over-drying of the paper sheet 108 , or correct any dry streaks in the paper sheet 108 .
- the paper sheet 108 is then often passed through a calender having several nips of counter-rotating rolls.
- Arrays of induction heating actuators 120 heat the shell surfaces of various ones of these rolls. As each roll surface locally heats up, the roll diameter is locally expanded and hence increases nip pressure, which in turn locally compresses the paper sheet 108 .
- the arrays of induction heating actuators 120 may therefore be used to control the caliper (thickness) profile of the paper sheet 108 .
- the nips of a calender may also be equipped with other actuator arrays, such as arrays of air showers or steam showers, which may be used to control the gloss profile or smoothness profile of the paper sheet.
- a thick stock flow actuator 122 controls the consistency of incoming stock received at the headbox 112 .
- a steam flow actuator 124 controls the amount of heat transferred to the paper sheet 108 from drying cylinders.
- the actuators 122 - 124 could, for example, represent valves controlling the flow of stock and steam, respectively. These actuators may be used for controlling the dry weight and moisture of the paper sheet 108 .
- Additional components could be used to further process the paper sheet 108 , such as a supercalender (for improving the paper sheet's thickness, smoothness, and gloss) or one or more coating stations (each applying a layer of coatant to a surface of the paper to improve the smoothness and printability of the paper sheet).
- additional flow actuators may be used to control the proportions of different types of pulp and filler material in the thick stock and to control the amounts of various additives (such as retention aid or dyes) that are mixed into the stock.
- one or more properties of the paper sheet 108 may be continuously or repeatedly measured.
- the sheet properties can be measured at one or various stages in the manufacturing process. This information may then be used to adjust the paper machine 102 , such as by adjusting various actuators within the paper machine 102 . This may help to compensate for any variations of the sheet properties from desired targets, which may help to ensure the quality of the sheet 108 .
- the paper machine 102 includes one or more sensor arrays 126 - 128 , each of which may include one or more sensors.
- Each sensor array 126 - 128 is capable of measuring one or more characteristics of the paper sheet 108 .
- each sensor array 126 - 128 could include sensors for measuring the moisture, basis weight, caliper, coat weight, anisotropy, color, gloss, sheen, haze, fiber orientation, surface features (such as roughness, topography, or orientation distributions of surface features), or any other or additional characteristics of the paper sheet 108 .
- Each sensor array 126 - 128 includes any suitable structure or structures for measuring or detecting one or more characteristics of the paper sheet 108 .
- the sensors in a sensor array 126 - 128 could be stationary or scanning sensors. Stationary sensors could be deployed in one or a few locations across the sheet 108 , or they could be deployed at multiple locations across the whole width of the sheet 108 such that substantially the entire sheet width is measured.
- a scanning set of sensors could include any number of moving sensors.
- the control system 104 receives measurement data from the sensor arrays 126 - 128 and uses the data to control the paper machine 102 .
- the control system 104 may use the measurement data to adjust any of the actuators or other components of the paper machine 102 .
- the control system 104 includes any suitable structure for controlling the operation of at least part of the paper machine 102 , such as one or more computing devices.
- the control system 104 includes at least one processing device 130 and at least one memory 132 storing instructions and data used, generated, or collected by the processing device(s) 130 .
- the control system 104 also includes at least one network interface 134 for communicating over one or more networks, such as an Ethernet network, an electrical signal network, or any other or additional type(s) of network(s).
- the control system 104 could include a single computing device with these components, or multiple distributed computing devices could include multiple instances of these components.
- the network 106 is coupled to the control system 104 and various components of the paper machine 102 (such as the actuators and sensor arrays).
- the network 106 facilitates communication between components of the system 100 .
- the network 106 represents any suitable network or combination of networks facilitating communication between components in the system 100 .
- the network 106 could, for example, represent a wired or wireless Ethernet network, an electrical signal network (such as a HART or FOUNDATION FIELDBUS network), a pneumatic control signal network, or any other or additional network(s).
- various actuator arrays represent CD actuator arrays, such as the dilution actuators, steam actuators, rewet shower actuators, and induction heating actuators. These actuator arrays control characteristics of the sheet 108 in the cross direction across the width of the sheet 108 (as opposed to the machine direction along the length of the sheet 108 ).
- a collection of actuators in one or more arrays may be said to represent an actuator “beam.” There is typically some overlap between adjacent actuators, meaning one actuator in an array often affects a sheet characteristic in its own associated zone and in adjacent zones associated with other actuators.
- the paper-making process is a large-scale two-dimensional system with spatial and dynamic components.
- the control system 104 often continuously monitors and controls this process to ensure that the quality of the final product meets desired specifications.
- CD control can be used to minimize the variance in paper quality in the cross direction perpendicular to the sheet's travel.
- Model predictive control (MPC) or other advanced control techniques can be used by the control system 104 to perform CD control in the system 100 .
- MPC controller could compute a sequence of control actions by minimizing a performance index while incorporating input constraints (such as CD actuator limits) and output constraints (such as paper quality specifications).
- This type of CD control is often implemented using an online QP solver, which can be executed or otherwise implemented by the control system 104 .
- an efficient QP solver having a dual-feasible active-set algorithm can be used to solve the control or optimization problem associated with CD control of paper-making processes.
- the new QP solver disclosed here can be significantly faster than conventional solvers for complex problems (such as about five to about twenty times faster or more). This decreases the amount of time needed to generate a control solution, which could be particularly helpful when conventional solvers cannot compute control solutions within a control interval's length.
- the new QP solver can be implemented using any suitable device(s) in the system 100 , such as within the control system 104 or as a stand-alone device.
- the QP solver is implemented using software, such as a real-time language like C or C++, executed by one or more processing systems.
- FIG. 1 illustrates one example of a sheet manufacturing or processing system 100
- various changes may be made to FIG. 1 .
- other systems could be used to produce paper products or other products.
- the production system 100 could include any number of paper machines or other production machinery having any suitable structure, and the system 100 could include any number of control systems.
- FIG. 1 illustrates one operational environment in which QP solving functionality can be used. This functionality could be used in any other suitable system.
- FIGS. 2 through 4 illustrate an example method for efficient QP solving for process control and optimization according to this disclosure.
- FIG. 2 illustrates an example method 200 for efficient QP solving
- FIGS. 3 and 4 illustrate example methods 300 and 400 that occur during the method 200 of FIG. 2 .
- the methods 200 - 400 are described as being performed by the control system 104 in the system 100 of FIG. 1 .
- the methods 200 - 400 could be performed by any other suitable device(s) and in any other suitable systems.
- a CD process can be modeled as a linear time-invariant system with multiple inputs and multiple outputs.
- An example process could be realized in the following state space form:
- X k ⁇ R Nx , ⁇ U k ⁇ R Nu , and Y k ⁇ R Ny are respectively the state variables, the control moves, and the controlled variables at time k.
- A, B, and C are respectively the state transition matrix, the input matrix, and the output matrix with compatible dimensions.
- N u is equal to
- ⁇ j 1 n u ⁇ n j ,
- n u is the number of actuator beams and n j is the number of actuator arrays installed on the j th (1 ⁇ j ⁇ n u ) beam.
- ⁇ ⁇ ⁇ U k [ ⁇ ⁇ ⁇ u k 1 ⁇ ⁇ ⁇ u k 2 ⁇ ⁇ ⁇ ⁇ u k n u ]
- ⁇ ⁇ u k j [ ⁇ ⁇ ⁇ u k j , 1 ⁇ ⁇ ⁇ u k j , 2 ⁇ ⁇ ⁇ ⁇ u k j ] ( 3 )
- MPC for CD control uses a prediction model to obtain estimations of controlled variables in a specified time horizon.
- H U denotes the control horizon
- H Y denotes the prediction horizon (1 ⁇ H u ⁇ H y ).
- Equation (1) a prediction model can be derived as follows:
- CD-MPC the following cost function can be defined for obtaining a control move:
- U sp and Y sp are dimension-compatible vectors containing setpoints of the corresponding variables at different time instants.
- Q 1 , Q 2 , Q 3 , and Q 4 are weighting matrices, which can be generated by a CD-MPC tuning algorithm.
- CD-MPC can also involve different physical constraints on actuators that make the problem in Equation (6) more difficult and time consuming to solve.
- actuator setpoints often have upper and lower limits, such as:
- each actuator beam there can be a bend limit for two adjacent actuator zones, which can be expressed as:
- the average of the actuator setpoints can be kept in a specified range, such as:
- the CD-MPC problem can be formulated as a linear inequality-constrained QP problem, such as:
- G ⁇ (H U •N U ) ⁇ (H U •N U ) is a symmetric, strictly positive definite Hessian matrix.
- g ⁇ H U •N U is the gradient vector
- a c ⁇ N c ⁇ (H U •N U ) is the constraint matrix (where N c equals the number of different types of constraints)
- b c ⁇ N c is the right-hand side constraint vector.
- G and A c can be sparse matrices with large dimensions. Moreover, the number of rows in A c can be much larger than the number of columns (meaning N c >>H U •N U ), and some constraints are dependant on each other.
- FIG. 2 illustrates an example method 200 for efficient QP solving.
- the method 200 can be used, for example, to solve the QP problem in Equation (14) using a dual-feasible active-set algorithm.
- the following notations and definitions are used:
- input data is received and analyzed at step 202 .
- This could include, for example, the control system 104 receiving measurement data from one or more of the sensor arrays 126 - 128 .
- This could also include the control system 104 receiving or generating a Hessian matrix and actuator constraint matrices associated with the process being controlled.
- the input data could involve any suitable process being controlled, such as a paper-making process for forming a paper sheet 108 .
- the Hessian matrix often has a band diagonal structure (a sparse structure), but sometimes it may have a dense structure.
- the sparsity of a matrix can be determined in any suitable manner.
- the sparse version does not factorize the Hessian matrix, while the dense version can use a Cholesky factorization of the Hessian matrix when computing step directions using a Schur complement as described below.
- the dense version can still exploit the sparsity of the constraint matrix even when the Hessian is not sparse.
- the solution points for two consecutive QP problems are often relatively close to each other. Therefore, it can be useful to use the optimal solution to a previous QP problem as the initial guess for a solution to a current QP problem. This can significantly reduce the number of iterations needed to identify the optimal solution to the current QP problem.
- a “warm” start occurs at step 212 . Otherwise, a “cold” start occurs at step 214 .
- FIG. 3 illustrates an example method 300 for startup using a warm or cold start.
- a determination is made whether an initial guess is non-zero at step 302 .
- An initial guess of zero indicates that a prior solution is not available, and a cold start occurs.
- an initial solution is set by solving a non-constrained QP problem (a QP problem without any active constraints) at step 306 . This can be expressed as:
- a non-zero initial guess at step 302 indicates that a prior solution is available, and a warm start may occur.
- a determination is made whether the prior solution has any active constraints at step 304 . If there are not any active constraints, the process moves to step 306 and proceeds as described above. If there are any active constraints, a warm start occurs, and the optimal solution to a previous QP problem can be used in various ways. In some embodiments, for instance, the optimal solution to the previous QP problem is used to identify active constraints and initialize an active set and other related variables at step 314 .
- an iterative process using a dual-feasible active-set algorithm then occurs, where any violated constraints are brought into the active constraint set of the solution.
- a determination is made whether there are any violated constraints at step 216 . If not, the method 200 ends as there are no additional constraints to be brought into the active constraint set of the solution.
- the optimal solution denoted as ⁇ U* has been found, which can be expressed as:
- one of the violated constraints is selected at step 218 , one or more step directions are determined at step 220 , the step length is determined at step 222 , and the solution is updated at step 224 .
- the violated constraint that is selected can be denoted A + and b + .
- a step direction ⁇ right arrow over (U) ⁇ (k) and a step length ⁇ are computed to define the next iteration ⁇ U (k) + ⁇ right arrow over (U) ⁇ (k) .
- a step direction ⁇ right arrow over ( ⁇ ) ⁇ (k) and a step length ⁇ are computed to define the next iteration ⁇ (k) + ⁇ right arrow over ( ⁇ ) ⁇ (k) .
- the step lengths can be calculated as shown in FIG. 4 .
- a determination can be made whether the active set is empty at step 402 . If the active set is not empty, full and partial step lengths can be determined at step 404 . If the active set is empty, only the full step length is determined at step 406 .
- step directions and full step lengths are computed by solving an augmented Karush-Kuhn-Tucker (KKT) system.
- KKT Karush-Kuhn-Tucker
- Equation (18) After some manipulation of Equation (18), the solution of ⁇ f can be obtained, and the equations to calculate ⁇ right arrow over (U) ⁇ (k) and ⁇ right arrow over ( ⁇ ) ⁇ (k) can be expressed as:
- a full step length need not be taken since some dual variables are possibly negative if the full step length is taken. In such a situation, dual feasibility is not guaranteed. For this reason, a maximum step length that maintains all dual variables non-negative can be determined, and this is referred to as a partial step length. This can be expressed as:
- the partial step length could be calculated as:
- ⁇ p min i ⁇ ⁇ - ⁇ i ( k ) ⁇ ⁇ i ( k )
- ⁇ f and ⁇ p can be used to determine how to update the problem solution. If both ⁇ f and ⁇ p equal infinity ( ⁇ ), it can imply that the current solution in infeasible. If only ⁇ f equals infinity, the selected violated constraint can be linearly dependent on at least one of the constraints in A w . In that case, a partial solution update can occur as follows:
- G ⁇ 1 can be calculated in the initial step and stored for reuse, which can help speed up the computations.
- the Schur complement can also be augmented when a constraint is added/dropped.
- a new constraint is added to A w , and one more step can be added to augment the Schur complement as follows:
- the Schur complement can be changed by deleting the corresponding row and column of S c .
- FIGS. 2 through 4 illustrate examples of methods for efficient QP solving for process control and optimization
- various changes may be made to FIGS. 2 through 4 .
- steps in each figure could overlap, occur in parallel, occur in a different order, or occur any number of times.
- the QP problem being solved relates to the control of cross-direction actuators in a paper-making process.
- the QP solver described above could be used to solve a QP problem associated with any other suitable device or system.
- FIG. 5 illustrates an example system 500 using an efficient QP solver for process control and optimization according to this disclosure.
- the system 500 here generically represents any suitable system that uses QP problem solving for control.
- a QP solver 502 is used to solve QP problems associated with control of a spatially distributed actuating array 504 , whose individual actuators 506 are distributed in space.
- the QP solver 502 could represent a stand-alone device or be integrated into another device or system, such as a control system.
- the QP solver 502 could be implemented in any suitable manner, such as using at least one processing device, at least one memory, and at least one network interface.
- the array 504 represents any suitable collection of actuators 506 in any suitable configuration.
- the array 504 could represent a one-dimensional array of actuators or a multi-dimensional array of actuators.
- the actuators 506 represent any suitable actuators for performing one or more functions within a larger system.
- the QP solver 502 can be used with any large-scale system where one actuator 506 interacts with or affects adjacent actuators. In these types of systems, controlling or optimizing the actuators 506 can involve the generation and solution of a QP problem.
- the QP solver 502 can perform the same types of operations described above to solve the QP problem. Note that the specific equations used to solve the QP problem can vary depending, for example, on the type of actuators 506 being controlled. Moreover, certain operations described above could be omitted, such as the determination whether input data is sparse. Depending on the type of actuators 506 being controlled, there may be only sparse data matrices or only dense data matrices involved in the calculations.
- system 500 could represent any suitable device or system that uses multiple actuators with overlapping effects.
- Example systems include the paper-making system described above or similar sheet-making systems.
- Other example systems could include large telescopes where the actuators 506 control mirrors in the telescopes or any other large-scale spatially-distributed systems.
- FIG. 5 illustrates one example of a system 500 using an efficient QP solver 502 for process control and optimization
- the system 500 could include any number of actuator arrays, each containing any number of actuators.
- various functions described above are implemented or supported by a computer program that is formed from computer readable program code and that is embodied in a computer readable medium.
- computer readable program code includes any type of computer code, including source code, object code, and executable code.
- computer readable medium includes any type of medium capable of being accessed by a computer, such as read only memory (ROM), random access memory (RAM), a hard disk drive, a compact disc (CD), a digital video disc (DVD), or any other type of memory.
- the term “couple” and its derivatives refer to any direct or indirect communication between two or more elements, whether or not those elements are in physical contact with one another.
- application and “program” refer to one or more computer programs, software components, sets of instructions, procedures, functions, objects, classes, instances, related data, or a portion thereof adapted for implementation in a suitable computer code (including source code, object code, or executable code).
- transmit and “communicate,” as well as derivatives thereof, encompass both direct and indirect communication.
- the term “or” is inclusive, meaning and/or.
- phrases “associated with” and “associated therewith,” as well as derivatives thereof, may mean to include, be included within, interconnect with, contain, be contained within, connect to or with, couple to or with, be communicable with, cooperate with, interleave, juxtapose, be proximate to, be bound to or with, have, have a property of, have a relationship to or with, or the like.
Abstract
Description
- This disclosure relates generally to control and optimization systems. More specifically, this disclosure relates to an efficient quadratic programming (QP) solver for process control and optimization.
- Model predictive control (MPC) is a popular technique for controlling multi-input and multi-output processes, such as industrial manufacturing processes. MPC uses a model to predict how one or more controlled process variables are expected to behave in the future. Changes can then be made to one or more manipulated process variables in order to alter the controlled process variable(s). Ideally, each controlled process variable is thereby maintained within a desired range.
- An MPC controller often implements an online quadratic programming (QP) solver for solving an optimization problem related to a controlled process. However, the efficient execution of an optimization routine often poses challenges in various circumstances. For example, process processes may involve hundreds of manipulated process variables and thousands of controlled process variables (many with active limits and rate constraints). Also, control intervals can be relatively short, such as ten to twenty seconds. While generic and custom QP solvers have been developed, they often suffer from various shortcomings.
- This disclosure provides an efficient quadratic programming (QP) solver for process control and optimization.
- In a first embodiment, a method includes identifying an initial solution to a quadratic programming (QP) problem associated with a process. The method also includes performing an iterative procedure having one or more iterations. Each iteration includes determining whether any constraint associated with the process is violated in the solution. If so, each iteration also includes selecting a violated constraint, determining a step direction and a step length associated with the selected violated constraint, and updating the solution based on the step direction and the step length. Determining the step direction and the step length includes using a Schur complement based on an active set of constraints associated with the solution. The Schur complement is nonsingular during all iterations of the iterative procedure except when the active set is empty.
- In a second embodiment, an apparatus includes at least one memory configured to store an initial solution to a quadratic programming (QP) problem associated with a process. The apparatus also includes at least one processing device configured to perform an iterative procedure having one or more iterations. During each iteration, the at least one processing device is configured to determine whether any constraint associated with the process is violated in the solution. If so, the at least one processing device is also configured to select a violated constraint, determine a step direction and a step length associated with the selected violated constraint, and update the solution based on the step direction and the step length. The at least one processing device is configured to determine the step direction and the step length using a Schur complement based on an active set of constraints associated with the solution. The Schur complement is nonsingular during all iterations of the iterative procedure except when the active set is empty.
- In a third embodiment, a computer readable medium embodies a computer program. The computer program includes computer readable program code for identifying an initial solution to a quadratic programming (QP) problem associated with a process. The computer program also includes computer readable program code for performing an iterative procedure having one or more iterations. Each iteration includes determining whether any constraint associated with the process is violated in the solution. If so, each iteration also includes selecting a violated constraint, determining a step direction and a step length associated with the selected violated constraint, and updating the solution based on the step direction and the step length. The computer readable program code for determining the step direction and the step length includes computer readable program code for using a Schur complement based on an active set of constraints associated with the solution. The Schur complement is nonsingular during all iterations of the iterative procedure except when the active set is empty.
- Other technical features may be readily apparent to one skilled in the art from the following figures, descriptions, and claims.
- For a more complete understanding of this disclosure, reference is now made to the following description, taken in conjunction with the accompanying drawings, in which:
-
FIG. 1 illustrates an example sheet manufacturing or processing system according to this disclosure; -
FIGS. 2 through 4 illustrate an example method for efficient quadratic programming (QP) solving for process control and optimization according to this disclosure; and -
FIG. 5 illustrates an example system using an efficient QP solver for process control and optimization according to this disclosure. -
FIGS. 1 through 5 , discussed below, and the various embodiments used to describe the principles of the present invention in this patent document are by way of illustration only and should not be construed in any way to limit the scope of the invention. Those skilled in the art will understand that the principles of the invention may be implemented in any type of suitably arranged device or system. - Conventionally, a control or optimization problem in a given industry is solved using a general-purpose quadratic programming (QP) solver or a customized QP solver. However, general-purpose QP solvers often have various shortcomings. For example, general-purpose QP solvers are intended to solve a broader class of problems and may not achieve performance that a customized QP solver could achieve. Also, general-purpose QP solvers may not take advantage of a specific problem's structure and features for a given industry, which might otherwise significantly reduce the computation time needed to identify a problem solution. Customized QP solvers also often have various shortcomings, such as a poorly organized algorithm structure or defects that can result in complete control failure. A poorly organized structure makes the solver less reliable and makes it harder to diagnose problems (such as infeasible solutions).
- This disclosure describes an efficient QP solver for process control and optimization. The QP solver disclosed here is based on a dual-feasible active-set algorithm, a Schur complement method, and a warm start strategy. The Schur complement is nonsingular throughout its iterations, which makes the QP solver very reliable numerically and helps to avoid control failures. The QP solver can also identify problem solutions much faster than conventional QP solvers.
- Different from general-purpose QP solvers, the efficient QP solver disclosed here solves a control or optimization problem by taking advantage of the problem's structure and features, which can vary depending on the specific application of the QP solver. By doing this, the QP solver can speed up solution computations and enable a controller to control a process closer to the process' physical limits, resulting in increased process efficiency and product quality.
- In particular embodiments described below, the QP solver is used for cross direction (CD) control in a paper-making process. However, this represents one example use of the QP solver. The QP solver can generally be used in any other suitable device or system having at least one array of actuators that are controlled using a solution to a QP problem, such as any large-scale spatially-distributed system.
-
FIG. 1 illustrates an example sheet manufacturing orprocessing system 100 according to this disclosure. In this example, thesystem 100 includes apaper machine 102, acontrol system 104, and anetwork 106. Thepaper machine 102 includes various components used to produce a paper product, namely apaper sheet 108 that is collected at areel 110. Thecontrol system 104 monitors and controls the operation of thepaper machine 102, which may help to maintain or increase the quality of thepaper sheet 108 produced by thepaper machine 102. - In this example, the
paper machine 102 includes at least oneheadbox 112, which distributes a pulp suspension uniformly across the machine onto a continuous moving wire screen ormesh 113. The pulp suspension entering theheadbox 112 may contain, for example, 0.2-3% wood fibers, fillers, and/or other materials, with the remainder of the suspension being water. Theheadbox 112 may include an array of dilution actuators, which distributes dilution water into the pulp suspension across the sheet. The dilution water may be used to help ensure that the resultingpaper sheet 108 has a more uniform basis weight across thesheet 108. - Arrays of
drainage elements 114, such as vacuum boxes, remove as much water as possible to initiate the formation of thesheet 108. An array ofsteam actuators 116 produces hot steam that penetrates thepaper sheet 108 and releases the latent heat of the steam into thepaper sheet 108, thereby increasing the temperature of thepaper sheet 108 in sections across the sheet. The increase in temperature may allow for easier removal of remaining water from thepaper sheet 108. An array ofrewet shower actuators 118 adds small droplets of water (which may be air atomized) onto the surface of thepaper sheet 108. The array ofrewet shower actuators 118 may be used to control the moisture profile of thepaper sheet 108, reduce or prevent over-drying of thepaper sheet 108, or correct any dry streaks in thepaper sheet 108. - The
paper sheet 108 is then often passed through a calender having several nips of counter-rotating rolls. Arrays ofinduction heating actuators 120 heat the shell surfaces of various ones of these rolls. As each roll surface locally heats up, the roll diameter is locally expanded and hence increases nip pressure, which in turn locally compresses thepaper sheet 108. The arrays ofinduction heating actuators 120 may therefore be used to control the caliper (thickness) profile of thepaper sheet 108. The nips of a calender may also be equipped with other actuator arrays, such as arrays of air showers or steam showers, which may be used to control the gloss profile or smoothness profile of the paper sheet. - Two additional actuators 122-124 are shown in
FIG. 1 . A thickstock flow actuator 122 controls the consistency of incoming stock received at theheadbox 112. Asteam flow actuator 124 controls the amount of heat transferred to thepaper sheet 108 from drying cylinders. The actuators 122-124 could, for example, represent valves controlling the flow of stock and steam, respectively. These actuators may be used for controlling the dry weight and moisture of thepaper sheet 108. - Additional components could be used to further process the
paper sheet 108, such as a supercalender (for improving the paper sheet's thickness, smoothness, and gloss) or one or more coating stations (each applying a layer of coatant to a surface of the paper to improve the smoothness and printability of the paper sheet). Similarly, additional flow actuators may be used to control the proportions of different types of pulp and filler material in the thick stock and to control the amounts of various additives (such as retention aid or dyes) that are mixed into the stock. - This represents a brief description of one type of
paper machine 102 that may be used to produce a paper product. Additional details regarding this type ofpaper machine 102 are well-known in the art and are not needed for an understanding of this disclosure. Also, this represents one specific type ofpaper machine 102 that may be used in thesystem 100. Other machines or devices could be used that include any other or additional components for producing a paper product. In addition, this disclosure is not limited to use with systems for producing paper products and could be used with systems that process a paper product or with systems that produce or process other items or materials (such as multi-layer paperboard, cardboard, plastic, textiles, metal foil or sheets, or other or additional materials that are manufactured or processed as moving sheets). - In order to control the paper-making process, one or more properties of the
paper sheet 108 may be continuously or repeatedly measured. The sheet properties can be measured at one or various stages in the manufacturing process. This information may then be used to adjust thepaper machine 102, such as by adjusting various actuators within thepaper machine 102. This may help to compensate for any variations of the sheet properties from desired targets, which may help to ensure the quality of thesheet 108. - As shown in
FIG. 1 , thepaper machine 102 includes one or more sensor arrays 126-128, each of which may include one or more sensors. Each sensor array 126-128 is capable of measuring one or more characteristics of thepaper sheet 108. For example, each sensor array 126-128 could include sensors for measuring the moisture, basis weight, caliper, coat weight, anisotropy, color, gloss, sheen, haze, fiber orientation, surface features (such as roughness, topography, or orientation distributions of surface features), or any other or additional characteristics of thepaper sheet 108. - Each sensor array 126-128 includes any suitable structure or structures for measuring or detecting one or more characteristics of the
paper sheet 108. The sensors in a sensor array 126-128 could be stationary or scanning sensors. Stationary sensors could be deployed in one or a few locations across thesheet 108, or they could be deployed at multiple locations across the whole width of thesheet 108 such that substantially the entire sheet width is measured. A scanning set of sensors could include any number of moving sensors. - The
control system 104 receives measurement data from the sensor arrays 126-128 and uses the data to control thepaper machine 102. For example, thecontrol system 104 may use the measurement data to adjust any of the actuators or other components of thepaper machine 102. Thecontrol system 104 includes any suitable structure for controlling the operation of at least part of thepaper machine 102, such as one or more computing devices. In this example, thecontrol system 104 includes at least oneprocessing device 130 and at least onememory 132 storing instructions and data used, generated, or collected by the processing device(s) 130. Thecontrol system 104 also includes at least onenetwork interface 134 for communicating over one or more networks, such as an Ethernet network, an electrical signal network, or any other or additional type(s) of network(s). Thecontrol system 104 could include a single computing device with these components, or multiple distributed computing devices could include multiple instances of these components. - The
network 106 is coupled to thecontrol system 104 and various components of the paper machine 102 (such as the actuators and sensor arrays). Thenetwork 106 facilitates communication between components of thesystem 100. Thenetwork 106 represents any suitable network or combination of networks facilitating communication between components in thesystem 100. Thenetwork 106 could, for example, represent a wired or wireless Ethernet network, an electrical signal network (such as a HART or FOUNDATION FIELDBUS network), a pneumatic control signal network, or any other or additional network(s). - In the
system 100 ofFIG. 1 , various actuator arrays represent CD actuator arrays, such as the dilution actuators, steam actuators, rewet shower actuators, and induction heating actuators. These actuator arrays control characteristics of thesheet 108 in the cross direction across the width of the sheet 108 (as opposed to the machine direction along the length of the sheet 108). A collection of actuators in one or more arrays may be said to represent an actuator “beam.” There is typically some overlap between adjacent actuators, meaning one actuator in an array often affects a sheet characteristic in its own associated zone and in adjacent zones associated with other actuators. - The paper-making process is a large-scale two-dimensional system with spatial and dynamic components. The
control system 104 often continuously monitors and controls this process to ensure that the quality of the final product meets desired specifications. CD control can be used to minimize the variance in paper quality in the cross direction perpendicular to the sheet's travel. Model predictive control (MPC) or other advanced control techniques can be used by thecontrol system 104 to perform CD control in thesystem 100. For example, at each of multiple sampling intervals, an MPC controller could compute a sequence of control actions by minimizing a performance index while incorporating input constraints (such as CD actuator limits) and output constraints (such as paper quality specifications). This type of CD control is often implemented using an online QP solver, which can be executed or otherwise implemented by thecontrol system 104. As described below, an efficient QP solver having a dual-feasible active-set algorithm can be used to solve the control or optimization problem associated with CD control of paper-making processes. - The new QP solver disclosed here can be significantly faster than conventional solvers for complex problems (such as about five to about twenty times faster or more). This decreases the amount of time needed to generate a control solution, which could be particularly helpful when conventional solvers cannot compute control solutions within a control interval's length. The new QP solver can be implemented using any suitable device(s) in the
system 100, such as within thecontrol system 104 or as a stand-alone device. In particular embodiments, the QP solver is implemented using software, such as a real-time language like C or C++, executed by one or more processing systems. - Although
FIG. 1 illustrates one example of a sheet manufacturing orprocessing system 100, various changes may be made toFIG. 1 . For example, other systems could be used to produce paper products or other products. Also, while shown as including asingle paper machine 102 with various components and asingle control system 104, theproduction system 100 could include any number of paper machines or other production machinery having any suitable structure, and thesystem 100 could include any number of control systems. In addition,FIG. 1 illustrates one operational environment in which QP solving functionality can be used. This functionality could be used in any other suitable system. -
FIGS. 2 through 4 illustrate an example method for efficient QP solving for process control and optimization according to this disclosure. In particular,FIG. 2 illustrates anexample method 200 for efficient QP solving, andFIGS. 3 and 4 illustrateexample methods 300 and 400 that occur during themethod 200 ofFIG. 2 . For ease of explanation, the methods 200-400 are described as being performed by thecontrol system 104 in thesystem 100 ofFIG. 1 . However, the methods 200-400 could be performed by any other suitable device(s) and in any other suitable systems. - Prior to a discussion of the methods 200-400, consider the following system. A CD process can be modeled as a linear time-invariant system with multiple inputs and multiple outputs. An example process could be realized in the following state space form:
-
X k+1 =AX k +BΔU k (1) -
Y k =CX k (2) - where XkεRNx, ΔUkεRNu, and YkεRNy are respectively the state variables, the control moves, and the controlled variables at time k. Also, A, B, and C are respectively the state transition matrix, the input matrix, and the output matrix with compatible dimensions. Nu is equal to
-
- where nu is the number of actuator beams and nj is the number of actuator arrays installed on the jth (1≦j≦nu) beam. With this, the following can be obtained:
-
- It is noted that using ΔUk instead of Uk can simplify the predictions in each MPC iterations. The actual input at time k satisfies Uk=Uk−1+Uk, and uk j=uk−1 j+Δuk j.
- MPC for CD control (CD-MPC for short) uses a prediction model to obtain estimations of controlled variables in a specified time horizon. Here, HU denotes the control horizon, and HY denotes the prediction horizon (1≦Hu≦Hy). Based on Equation (1), a prediction model can be derived as follows:
-
- with t=1, 2, . . . , Hy. Here, it is assumed that (A, C) is observable, and {circumflex over (X)}k is an estimate of Xk.
- In CD-MPC, the following cost function can be defined for obtaining a control move:
-
- Usp and Ysp are dimension-compatible vectors containing setpoints of the corresponding variables at different time instants. Also, Q1, Q2, Q3, and Q4 are weighting matrices, which can be generated by a CD-MPC tuning algorithm.
- CD-MPC can also involve different physical constraints on actuators that make the problem in Equation (6) more difficult and time consuming to solve. For example, actuator setpoints often have upper and lower limits, such as:
-
u bound j ≦u k j ≦u bound −j ,j=1, 2, . . . , n u (8) - Also, for each actuator beam, there can be a bend limit for two adjacent actuator zones, which can be expressed as:
-
|u k j,2 −u k j,1 |≦u bend 1 −j (9) -
|u k j,nj −u k j,nj −1 |≦u bend 1 −j ,j=1, 2, . . . , n u (10) -
|u k j,l−2u k j,l−1 +u k j,l−2 |≦u bend 2 −j ,l=3, 4, . . . , n j (11) - Further, in each actuator beam, the average of the actuator setpoints can be kept in a specified range, such as:
-
- In addition, a large magnitude of Δuk j may not be physically feasible, which can be expressed as:
-
|Δu k j |≦u delta −j ,j=1, 2, . . . , n u (13) - By including these constraints and replacing all variables by ΔU, the CD-MPC problem can be formulated as a linear inequality-constrained QP problem, such as:
-
- In this problem, G and Ac can be sparse matrices with large dimensions. Moreover, the number of rows in Ac can be much larger than the number of columns (meaning Nc>>HU•NU), and some constraints are dependant on each other.
- With this example problem formulation in mind, reference is now made to
FIG. 2 , which illustrates anexample method 200 for efficient QP solving. Themethod 200 can be used, for example, to solve the QP problem in Equation (14) using a dual-feasible active-set algorithm. In this discussion ofFIG. 2 , the following notations and definitions are used: - Notations:
-
- ΔU(k), λ(k): primal and dual variables at iteration k
- Aw, bw: active constraint set
- A+, b+: violated constraint that should be added
- A_, b_: active constraint that should be dropped
- (•)i: ith row of a matrix or vector (•)
- Δ{right arrow over (U)}(k), {right arrow over (λ)}(k): step directions of ΔU(k), λ(k)
- Sc: Schur complement
- τf, τp: full and partial step lengths
- ∥•∥: Euclidean norm
- Definitions:
-
- Active constraint: Any constraint subject to AiΔU(k)−bi=0
- Inactive constraint: Any constraint subject to AiΔU(k)−bi<0
- Violated constraint: Any constraint subject to AiΔU(k)>bi
- As shown in
FIG. 2 , input data is received and analyzed atstep 202. This could include, for example, thecontrol system 104 receiving measurement data from one or more of the sensor arrays 126-128. This could also include thecontrol system 104 receiving or generating a Hessian matrix and actuator constraint matrices associated with the process being controlled. The input data could involve any suitable process being controlled, such as a paper-making process for forming apaper sheet 108. - A determination is made whether certain data associated with the process is sparse data at
step 204. This could include, for example, thecontrol system 104 determining whether the Hessian and actuator constraint matrices are sparse matrices. Sparse matrices can use less storage and can be manipulated more efficiently, so this determination can be used to determine which types of matrix functions are invoked during themethod 200. The Hessian matrix often has a band diagonal structure (a sparse structure), but sometimes it may have a dense structure. The sparsity of a matrix can be determined in any suitable manner. Once a determination is made, a sparse QP solver can be invoked atstep 206, or a dense QP solver can be invoked atstep 208. In particular embodiments, the sparse version does not factorize the Hessian matrix, while the dense version can use a Cholesky factorization of the Hessian matrix when computing step directions using a Schur complement as described below. The dense version can still exploit the sparsity of the constraint matrix even when the Hessian is not sparse. - A determination is made whether an initial guess of the problem solution is available at
step 210. In MPC and other advanced control applications, the solution points for two consecutive QP problems are often relatively close to each other. Therefore, it can be useful to use the optimal solution to a previous QP problem as the initial guess for a solution to a current QP problem. This can significantly reduce the number of iterations needed to identify the optimal solution to the current QP problem. If an initial guess for the solution to a current QP problem is available, a “warm” start occurs atstep 212. Otherwise, a “cold” start occurs atstep 214. -
FIG. 3 illustrates anexample method 300 for startup using a warm or cold start. A determination is made whether an initial guess is non-zero atstep 302. An initial guess of zero indicates that a prior solution is not available, and a cold start occurs. In a cold start, an initial solution is set by solving a non-constrained QP problem (a QP problem without any active constraints) atstep 306. This can be expressed as: -
(ΔU (0),λ(0)):=(−G −1 g,0) (15) -
A w (0)=0 (16) - A determination is made whether there are any active constraints in the initial solution at
step 308. If there are not any active constraints, an optimal solution to the QP problem is returned atstep 310. If there are any active constraints, the active set Aw and the Schur complement are set to empty sets atstep 312. - A non-zero initial guess at
step 302 indicates that a prior solution is available, and a warm start may occur. A determination is made whether the prior solution has any active constraints atstep 304. If there are not any active constraints, the process moves to step 306 and proceeds as described above. If there are any active constraints, a warm start occurs, and the optimal solution to a previous QP problem can be used in various ways. In some embodiments, for instance, the optimal solution to the previous QP problem is used to identify active constraints and initialize an active set and other related variables atstep 314. - Returning to
FIG. 2 , an iterative process using a dual-feasible active-set algorithm then occurs, where any violated constraints are brought into the active constraint set of the solution. A determination is made whether there are any violated constraints atstep 216. If not, themethod 200 ends as there are no additional constraints to be brought into the active constraint set of the solution. At this point, the optimal solution denoted as ΔU* has been found, which can be expressed as: -
ΔU*→ΔU k (17) - Otherwise, one of the violated constraints is selected at
step 218, one or more step directions are determined atstep 220, the step length is determined atstep 222, and the solution is updated atstep 224. Here, the violated constraint that is selected can be denoted A+ and b+. In an active set method, at a current point ΔU(k), a step direction Δ{right arrow over (U)}(k) and a step length τ are computed to define the next iteration ΔU(k)+τΔ{right arrow over (U)}(k). Similarly, at a current point λ(k), a step direction {right arrow over (λ)}(k) and a step length τ are computed to define the next iteration λ(k)+τ{right arrow over (λ)}(k). - In some embodiments, the step lengths can be calculated as shown in
FIG. 4 . A determination can be made whether the active set is empty atstep 402. If the active set is not empty, full and partial step lengths can be determined atstep 404. If the active set is empty, only the full step length is determined atstep 406. In particular embodiments, step directions and full step lengths are computed by solving an augmented Karush-Kuhn-Tucker (KKT) system. An example KKT system can be expressed as: -
- After some manipulation of Equation (18), the solution of τf can be obtained, and the equations to calculate Δ{right arrow over (U)}(k) and {right arrow over (λ)}(k) can be expressed as:
-
- Notice that a full step length need not be taken since some dual variables are possibly negative if the full step length is taken. In such a situation, dual feasibility is not guaranteed. For this reason, a maximum step length that maintains all dual variables non-negative can be determined, and this is referred to as a partial step length. This can be expressed as:
-
λi (k)+τp{right arrow over (λ)}i (k)≧0 (21) - Therefore, the partial step length could be calculated as:
-
- Once the full and partial step lengths are calculated, the values of τf and τp can be used to determine how to update the problem solution. If both τf and τp equal infinity (∞), it can imply that the current solution in infeasible. If only τf equals infinity, the selected violated constraint can be linearly dependent on at least one of the constraints in Aw. In that case, a partial solution update can occur as follows:
- (a) drop the constraint A_and b_from Aw and bw, where the row index of the dropped constraint is equal to:
-
- (b) set λ(k+1)=λ(k)+τpλ′(k) and delete the dual variable associated with the dropped constraint; and
- (c) set k=k+1 and repeat steps (a)-(b) without searching for a new violated constraint.
- If τp≦τf<∞, the following can be set:
-
ΔU (k+1) =ΔU (k)+τp Δ{right arrow over (U)} (k) (24) - in addition to steps (a)-(c) above. If τf≦τp, a full solution update can occur as follows:
- (a′) add the violated constraint to Aw and bw;
- (b′) set:
-
- (c′) set k=k+1 and repeat starting with a new violated constraint search.
- In the algorithm described above, one of the more time-consuming portions of the algorithm is the computation of Δ{right arrow over (U)}(k) and {right arrow over (λ)}(k) in Equations (19)-(20). Since this is an equilibrium system, it can be solved in various ways, such as by triangular factorization or Gaussian elimination. Example solutions can thus be explicitly given by:
-
Δ{right arrow over (U)}(k) =−G −1 [A + T +A w T{right arrow over (λ)}(k)] (26) -
{right arrow over (λ)}(k) =S c −1 A w G −1 A + T (27) - where Sc=−AWG−1AW T is called the Schur complement. Comparing Equations (26)-(27) with Equations (19)-(20), Sc −1 can be determined, which has a much smaller dimension than the matrix:
-
- Furthermore, G−1 can be calculated in the initial step and stored for reuse, which can help speed up the computations.
- The Schur complement can also be augmented when a constraint is added/dropped. In a full update, a new constraint is added to Aw, and one more step can be added to augment the Schur complement as follows:
-
- In a partial update, a constraint is dropped from Aw. Accordingly, the Schur complement can be changed by deleting the corresponding row and column of Sc.
- Regarding the Schur complement used here, it is nonsingular throughout the iteration steps, which can help to improve the reliability of the QP solver. At iteration k of the process, if AW has a full row rank and A+ is linearly dependent on the rows of Aw, A+ may not be added to the active constraint set. From Equation (19), it can be established that:
-
- which implies:
-
A w G −1 A w T{right arrow over (λ)}(k) =−A w G −1 A + T (31) - Since Aw has full row rank, {right arrow over (λ)}(k) has a unique solution. On the other hand, if A+ is linearly dependent on the rows of AW, there exists a vector z0 satisfying:
-
A w T z 0 =−A + T (32) - It turns out that z0 is a solution to Equation (31). Thus:
-
- is the solution of Equation (19). Since ΔU(k)=0 implies τf=∞, A+ is not added to Aw.
- Moreover, if the Hessian matrix G in Equation (14) is strictly positive definite, the Schur complement Sc can always be nonsingular during the iteration steps. To establish that Sc is nonsingular, it suffices to show that Aw always has full row rank because G is strictly positive definite. In the initial step (cold start), Aw=0 means no constraint is included in the active set. When the first constraint is added, Aw becomes full row rank. Once Aw has full row rank, newly added constraints are independent with the existing constraints in Aw. Thus, Aw always has full row rank except during the initial step, so Sc is nonsingular throughout its iterations.
- Another time-consuming portion of this active set algorithm is solving the underlying KKT system at optimization iterations. In some embodiments, the Schur complement is always invertible, which can be computationally cheaper and reliable than conventional systems.
- To summarize, some embodiments of the proposed QP solver disclosed here contain (among others) the following features:
-
- The QP solver exploits the sparsity of the problem data (such as the Hessian and constraints matrices) to speed up computations and reduce memory requirements of the solver.
- The Schur complement used by the QP solver is guaranteed to be nonsingular.
- The QP solver employs a warm startup to take advantage of the initial guess for the optimal solution to reduce the number of optimization iterations.
- All constraints can be uniformly treated as inequalities, which helps to simplify calculations.
- Although
FIGS. 2 through 4 illustrate examples of methods for efficient QP solving for process control and optimization, various changes may be made toFIGS. 2 through 4 . For example, while shown as a series of steps, various steps in each figure could overlap, occur in parallel, occur in a different order, or occur any number of times. - Note that in the description above, it has been assumed that the QP problem being solved relates to the control of cross-direction actuators in a paper-making process. However, the QP solver described above could be used to solve a QP problem associated with any other suitable device or system.
-
FIG. 5 illustrates anexample system 500 using an efficient QP solver for process control and optimization according to this disclosure. Thesystem 500 here generically represents any suitable system that uses QP problem solving for control. - In the
system 500, aQP solver 502 is used to solve QP problems associated with control of a spatially distributedactuating array 504, whoseindividual actuators 506 are distributed in space. TheQP solver 502 could represent a stand-alone device or be integrated into another device or system, such as a control system. TheQP solver 502 could be implemented in any suitable manner, such as using at least one processing device, at least one memory, and at least one network interface. - The
array 504 represents any suitable collection ofactuators 506 in any suitable configuration. Thearray 504 could represent a one-dimensional array of actuators or a multi-dimensional array of actuators. Theactuators 506 represent any suitable actuators for performing one or more functions within a larger system. - In some embodiments, the
QP solver 502 can be used with any large-scale system where oneactuator 506 interacts with or affects adjacent actuators. In these types of systems, controlling or optimizing theactuators 506 can involve the generation and solution of a QP problem. TheQP solver 502 can perform the same types of operations described above to solve the QP problem. Note that the specific equations used to solve the QP problem can vary depending, for example, on the type ofactuators 506 being controlled. Moreover, certain operations described above could be omitted, such as the determination whether input data is sparse. Depending on the type ofactuators 506 being controlled, there may be only sparse data matrices or only dense data matrices involved in the calculations. - Note that the
system 500 could represent any suitable device or system that uses multiple actuators with overlapping effects. Example systems include the paper-making system described above or similar sheet-making systems. Other example systems could include large telescopes where theactuators 506 control mirrors in the telescopes or any other large-scale spatially-distributed systems. - Although
FIG. 5 illustrates one example of asystem 500 using anefficient QP solver 502 for process control and optimization, various changes may be made toFIG. 5 . For example, thesystem 500 could include any number of actuator arrays, each containing any number of actuators. - In some embodiments, various functions described above are implemented or supported by a computer program that is formed from computer readable program code and that is embodied in a computer readable medium. The phrase “computer readable program code” includes any type of computer code, including source code, object code, and executable code. The phrase “computer readable medium” includes any type of medium capable of being accessed by a computer, such as read only memory (ROM), random access memory (RAM), a hard disk drive, a compact disc (CD), a digital video disc (DVD), or any other type of memory.
- It may be advantageous to set forth definitions of certain words and phrases used throughout this patent document. The term “couple” and its derivatives refer to any direct or indirect communication between two or more elements, whether or not those elements are in physical contact with one another. The terms “application” and “program” refer to one or more computer programs, software components, sets of instructions, procedures, functions, objects, classes, instances, related data, or a portion thereof adapted for implementation in a suitable computer code (including source code, object code, or executable code). The terms “transmit,” “receive,” and “communicate,” as well as derivatives thereof, encompass both direct and indirect communication. The terms “include” and “comprise,” as well as derivatives thereof, mean inclusion without limitation. The term “or” is inclusive, meaning and/or. The phrases “associated with” and “associated therewith,” as well as derivatives thereof, may mean to include, be included within, interconnect with, contain, be contained within, connect to or with, couple to or with, be communicable with, cooperate with, interleave, juxtapose, be proximate to, be bound to or with, have, have a property of, have a relationship to or with, or the like.
- While this disclosure has described certain embodiments and generally associated methods, alterations and permutations of these embodiments and methods will be apparent to those skilled in the art. Accordingly, the above description of example embodiments does not define or constrain this disclosure. Other changes, substitutions, and alterations are also possible without departing from the spirit and scope of this disclosure, as defined by the following claims.
Claims (23)
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US13/485,594 US8600525B1 (en) | 2012-05-31 | 2012-05-31 | Efficient quadratic programming (QP) solver for process control and optimization |
CA2814455A CA2814455A1 (en) | 2012-05-31 | 2013-05-01 | Efficient quadratic programming (qp) solver for process and optimization |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US13/485,594 US8600525B1 (en) | 2012-05-31 | 2012-05-31 | Efficient quadratic programming (QP) solver for process control and optimization |
Publications (2)
Publication Number | Publication Date |
---|---|
US8600525B1 US8600525B1 (en) | 2013-12-03 |
US20130325148A1 true US20130325148A1 (en) | 2013-12-05 |
Family
ID=49640831
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US13/485,594 Active US8600525B1 (en) | 2012-05-31 | 2012-05-31 | Efficient quadratic programming (QP) solver for process control and optimization |
Country Status (2)
Country | Link |
---|---|
US (1) | US8600525B1 (en) |
CA (1) | CA2814455A1 (en) |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140372112A1 (en) * | 2013-06-18 | 2014-12-18 | Microsoft Corporation | Restructuring deep neural network acoustic models |
US9589565B2 (en) | 2013-06-21 | 2017-03-07 | Microsoft Technology Licensing, Llc | Environmentally aware dialog policies and response generation |
US9614724B2 (en) | 2014-04-21 | 2017-04-04 | Microsoft Technology Licensing, Llc | Session-based device configuration |
US9697200B2 (en) | 2013-06-21 | 2017-07-04 | Microsoft Technology Licensing, Llc | Building conversational understanding systems using a toolset |
US9717006B2 (en) | 2014-06-23 | 2017-07-25 | Microsoft Technology Licensing, Llc | Device quarantine in a wireless network |
US20170316336A1 (en) * | 2014-11-21 | 2017-11-02 | Microsoft Technology Licensing, Llc | Method for efficient implementation of diagonal operators over clifford+t basis |
US9874914B2 (en) | 2014-05-19 | 2018-01-23 | Microsoft Technology Licensing, Llc | Power management contracts for accessory devices |
US10111099B2 (en) | 2014-05-12 | 2018-10-23 | Microsoft Technology Licensing, Llc | Distributing content in managed wireless distribution networks |
US10497367B2 (en) | 2014-03-27 | 2019-12-03 | Microsoft Technology Licensing, Llc | Flexible schema for language model customization |
US10613490B2 (en) | 2018-02-05 | 2020-04-07 | Mitsubishi Electric Research Laboratories, Inc. | Method and apparatus for preconditioned predictive control |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9618912B2 (en) * | 2013-05-17 | 2017-04-11 | Mitsubishi Electric Research Laboratories, Inc. | MPC controller using parallel quadratic programming |
US9760534B2 (en) | 2014-02-20 | 2017-09-12 | Mitsubishi Electric Research Laboratories, Inc. | Optimal parameter selection and acceleration in ADMM for multi-stage stochastic convex quadratic programs |
US9753892B2 (en) | 2014-02-20 | 2017-09-05 | Mitsubishi Electric Research Laboratories, Inc. | Method for solving quadratic programs for convex sets with linear equalities by an alternating direction method of multipliers with optimized step sizes |
CN104850008B (en) * | 2015-03-17 | 2017-08-01 | 厦门奥普拓自控科技有限公司 | A kind of multivariable Control of grate firing boiler and the method for real-time optimization |
US10190522B2 (en) * | 2016-06-17 | 2019-01-29 | Toyota Motor Engineering & Manufacturing North America, Inc. | Hybrid partial and full step quadratic solver for model predictive control of diesel engine air path flow and methods of use |
US10969749B2 (en) * | 2017-08-22 | 2021-04-06 | Honeywell Limited | Application of model predictive control (MPC)-based forced ramping of process input variables and process output reference trajectory design over a prediction horizon for MPC-based paper machine grade change control |
Family Cites Families (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5185715A (en) * | 1990-03-30 | 1993-02-09 | Hughes Aircraft Company | Data processing systems and methods for linear programming |
US5940791A (en) * | 1997-05-09 | 1999-08-17 | Washington University | Method and apparatus for speech analysis and synthesis using lattice ladder notch filters |
US20010032029A1 (en) * | 1999-07-01 | 2001-10-18 | Stuart Kauffman | System and method for infrastructure design |
JP2005509135A (en) * | 2001-04-03 | 2005-04-07 | ザ、テクサス、エイ、アンド、エム、ユーニヴァーサティ、システィム | A method for characterizing particles in suspension systems by frequency domain photon propagation measurements |
US20070297332A1 (en) * | 2006-06-22 | 2007-12-27 | James Andrew Broberg | Distributed resource allocation in stream processing systems |
US20080072182A1 (en) * | 2006-09-19 | 2008-03-20 | The Regents Of The University Of California | Structured and parameterized model order reduction |
US7650263B2 (en) * | 2006-09-26 | 2010-01-19 | Strider Labs, Inc. | Method for fast computation of optimal contact forces |
US20090271443A1 (en) * | 2008-04-29 | 2009-10-29 | International Business Machines Corporation | Method for detecting data attribute dependencies |
-
2012
- 2012-05-31 US US13/485,594 patent/US8600525B1/en active Active
-
2013
- 2013-05-01 CA CA2814455A patent/CA2814455A1/en not_active Abandoned
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9728184B2 (en) * | 2013-06-18 | 2017-08-08 | Microsoft Technology Licensing, Llc | Restructuring deep neural network acoustic models |
US20140372112A1 (en) * | 2013-06-18 | 2014-12-18 | Microsoft Corporation | Restructuring deep neural network acoustic models |
US10304448B2 (en) | 2013-06-21 | 2019-05-28 | Microsoft Technology Licensing, Llc | Environmentally aware dialog policies and response generation |
US9697200B2 (en) | 2013-06-21 | 2017-07-04 | Microsoft Technology Licensing, Llc | Building conversational understanding systems using a toolset |
US9589565B2 (en) | 2013-06-21 | 2017-03-07 | Microsoft Technology Licensing, Llc | Environmentally aware dialog policies and response generation |
US10572602B2 (en) | 2013-06-21 | 2020-02-25 | Microsoft Technology Licensing, Llc | Building conversational understanding systems using a toolset |
US10497367B2 (en) | 2014-03-27 | 2019-12-03 | Microsoft Technology Licensing, Llc | Flexible schema for language model customization |
US9614724B2 (en) | 2014-04-21 | 2017-04-04 | Microsoft Technology Licensing, Llc | Session-based device configuration |
US10111099B2 (en) | 2014-05-12 | 2018-10-23 | Microsoft Technology Licensing, Llc | Distributing content in managed wireless distribution networks |
US9874914B2 (en) | 2014-05-19 | 2018-01-23 | Microsoft Technology Licensing, Llc | Power management contracts for accessory devices |
US9717006B2 (en) | 2014-06-23 | 2017-07-25 | Microsoft Technology Licensing, Llc | Device quarantine in a wireless network |
US20170316336A1 (en) * | 2014-11-21 | 2017-11-02 | Microsoft Technology Licensing, Llc | Method for efficient implementation of diagonal operators over clifford+t basis |
US10366339B2 (en) * | 2014-11-21 | 2019-07-30 | Microsoft Technology Licensing, Llc | Method for efficient implementation of diagonal operators over clifford+T basis |
US10613490B2 (en) | 2018-02-05 | 2020-04-07 | Mitsubishi Electric Research Laboratories, Inc. | Method and apparatus for preconditioned predictive control |
Also Published As
Publication number | Publication date |
---|---|
CA2814455A1 (en) | 2013-11-30 |
US8600525B1 (en) | 2013-12-03 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US8600525B1 (en) | Efficient quadratic programming (QP) solver for process control and optimization | |
US7496413B2 (en) | Apparatus and method for coordinating controllers to control a paper machine or other machine | |
US7650195B2 (en) | Automated tuning of large-scale multivariable model predictive controllers for spatially-distributed processes | |
US9971318B2 (en) | Method and apparatus for robust tuning of model-based process controllers used with uncertain multiple-input, multiple-output (MIMO) processes | |
CA2647716C (en) | Fast performance prediction of multivariable model predictive controller for paper machine cross-directional processes | |
US10429800B2 (en) | Layered approach to economic optimization and model-based control of paper machines and other systems | |
CN107831737B (en) | Apparatus, method and non-transitory computer readable medium for measurement and control | |
US10309059B2 (en) | Method of designing model predictive control for cross directional flat sheet manufacturing processes to guarantee temporal robust stability and performance | |
US10890882B2 (en) | Method and apparatus for designing model-based control having spatial robustness for multiple array cross-direction (CD) web manufacturing or processing systems or other systems | |
CN110622079B (en) | Method and apparatus for designing model-based control with spatial robustness for a multi-array cross-direction (CD) web manufacturing or processing system or other system | |
CN110637261B (en) | Method and apparatus for designing model-based controls | |
US20170045419A1 (en) | System and method for step test-free machine modeling using statistical information about multiple web manufacturing or processing systems | |
Mäenpää | Robust model predictive control for cross-directional processes | |
WO2016191849A1 (en) | Method and apparatus for robust tuning of model-based process controllers used with uncertain multiple-input, multiple-output (mimo) processes |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AS | Assignment |
Owner name: HONEYWELL ASCA INC., CANADA Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:MUSTAFA, GHULAM;WANG, JIADONG;CHEN, TONGWEN;AND OTHERS;REEL/FRAME:028299/0472 Effective date: 20120530 |
|
STCF | Information on status: patent grant |
Free format text: PATENTED CASE |
|
AS | Assignment |
Owner name: HONEYWELL LIMITED, CANADA Free format text: CHANGE OF NAME;ASSIGNOR:HONEYWELL ASCA INC.;REEL/FRAME:039309/0806 Effective date: 20150701 |
|
FPAY | Fee payment |
Year of fee payment: 4 |
|
MAFP | Maintenance fee payment |
Free format text: PAYMENT OF MAINTENANCE FEE, 8TH YEAR, LARGE ENTITY (ORIGINAL EVENT CODE: M1552); ENTITY STATUS OF PATENT OWNER: LARGE ENTITY Year of fee payment: 8 |