CN111781821B - Parameter self-tuning method of SISO (SISO) compact-format model-free controller based on Attention mechanism cyclic neural network - Google Patents
Parameter self-tuning method of SISO (SISO) compact-format model-free controller based on Attention mechanism cyclic neural network Download PDFInfo
- Publication number
- CN111781821B CN111781821B CN202010560142.9A CN202010560142A CN111781821B CN 111781821 B CN111781821 B CN 111781821B CN 202010560142 A CN202010560142 A CN 202010560142A CN 111781821 B CN111781821 B CN 111781821B
- Authority
- CN
- China
- Prior art keywords
- siso
- controller
- format model
- neural network
- input
- 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.)
- Active
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 70
- 238000000034 method Methods 0.000 title claims abstract description 38
- 125000004122 cyclic group Chemical group 0.000 title claims abstract description 24
- 230000007246 mechanism Effects 0.000 title claims abstract description 16
- URWAJWIAIPFPJE-YFMIWBNJSA-N sisomycin Chemical compound O1C[C@@](O)(C)[C@H](NC)[C@@H](O)[C@H]1O[C@@H]1[C@@H](O)[C@H](O[C@@H]2[C@@H](CC=C(CN)O2)N)[C@@H](N)C[C@H]1N URWAJWIAIPFPJE-YFMIWBNJSA-N 0.000 title claims abstract 39
- 230000000306 recurrent effect Effects 0.000 claims abstract description 43
- 238000004364 calculation method Methods 0.000 claims abstract description 32
- 238000011478 gradient descent method Methods 0.000 claims abstract description 8
- 230000006870 function Effects 0.000 claims description 48
- 230000009897 systematic effect Effects 0.000 claims description 21
- 230000008569 process Effects 0.000 claims description 11
- 230000004913 activation Effects 0.000 claims description 6
- 238000004519 manufacturing process Methods 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 2
- 238000005070 sampling Methods 0.000 claims description 2
- 238000012549 training Methods 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 14
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 10
- 101100532097 Vitis rotundifolia RUN1 gene Proteins 0.000 description 8
- 230000008859 change Effects 0.000 description 6
- 238000012360 testing method Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 5
- 238000002474 experimental method Methods 0.000 description 4
- 238000009825 accumulation Methods 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000000513 principal component analysis Methods 0.000 description 2
- 238000012216 screening Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 238000012800 visualization Methods 0.000 description 2
- 101100284769 Drosophila melanogaster hemo gene Proteins 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 239000004568 cement Substances 0.000 description 1
- 238000003889 chemical engineering Methods 0.000 description 1
- 230000001276 controlling effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 239000007788 liquid Substances 0.000 description 1
- 238000005272 metallurgy Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000007670 refining Methods 0.000 description 1
- 230000001105 regulatory effect Effects 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
- G05B13/0205—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric not using a model or a simulator of the controlled system
- G05B13/024—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric not using a model or a simulator of the controlled system in which a parameter or coefficient is automatically adjusted to optimise the performance
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Automation & Control Theory (AREA)
- Feedback Control In General (AREA)
Abstract
The invention discloses a parameter self-setting method of an SISO (SISO) compact form model-free controller based on an Attention mechanism cyclic neural network, which comprises the steps of firstly, utilizing the Attention mechanism to screen important information of an original input set and calculate to generate input of the cyclic neural network, carrying out forward calculation on the cyclic neural network to output parameters to be set of the SISO compact form model-free controller, adopting a control algorithm to calculate to obtain control input of a controlled object, taking a minimized system error function value as a target, adopting a gradient descent method, combining the control input with gradient information of each parameter to be set, utilizing a chain rule to carry out system error back propagation calculation, updating an ownership coefficient of the cyclic neural network, and realizing the parameter self-setting of the controller based on the cyclic neural network. The SISO compact-format model-free controller provided by the invention is based on the parameter self-setting method of the recurrent neural network of the Attention mechanism, can capture the important characteristics of input information, overcomes the difficulty of the online setting of the controller parameters, and has good control effect on the SISO system.
Description
Technical Field
The invention belongs to the field of automatic control, and particularly relates to a parameter self-tuning method of an SISO (SISO) compact-format model-free controller based on an Attention mechanism cyclic neural network.
Background
A SISO (Single Input and Single Output) system is widely used in controlled objects such as reactors, rectifying towers, machines, devices, apparatuses, production lines, workshops, factories, and the like in industries such as oil refining, chemical engineering, thermal power, machinery, electricity, petrochemical industry, pharmaceutical industry, food, paper making, water treatment, metallurgy, cement, rubber, and the like. With the continuous improvement of the technological level, the industrial devices are increasingly large and complex, so that the production process presents more and more strong nonlinearity, time-varying characteristics and the like, and the traditional controller represented by the PID is often difficult to achieve an ideal control effect when controlling a complex controlled object with the strong nonlinearity, the time-varying characteristics and the like. The model-free controller is a novel control model based on data driving, has a good control effect on an unknown nonlinear time-varying system, and therefore has a good application prospect.
Existing implementations of a modeless controller for a SISO system include a SISO compact-format modeless controller. The SISO compact format model-free controller only relies on input and output data measured by a SISO controlled object in real time to analyze and design the controller, has the advantages of simple implementation method, small calculation load, strong robustness and the like, and can well control an unknown nonlinear time-varying SISO system. The theoretical basis of the SISO compact-format model-free controller is proposed by Houzhong and Jinshangtai in the 'model-free adaptive control-theory and application' (scientific publishing agency, 2013, page 55) of the Hemo, and the control algorithm is as follows:
wherein u (k) is the system control input at time k; e (k) is the system error at time k; phi is ac(k) The estimated value of the pseudo-partial derivative of the SISO system at the moment k; λ is a penalty factor and ρ is a step factor.
However, the SISO compact-format modeless controller also faces a difficult problem in the actual use process, and before the controller is put into use, the numerical values of parameters such as the penalty factor λ and the step factor ρ are manually set in advance depending on professional prior knowledge, and meanwhile, the modeless controller does not realize the online self-tuning of the parameters such as the penalty factor λ and the step factor ρ in the actual use process. The lack of effective setting means of the controller parameters not only causes time and labor consumption in the using and debugging process of the SISO compact-format model-free controller, but also can seriously affect the control effect of the SISO compact-format model-free controller sometimes, thereby limiting the wide application of the SISO compact-format model-free controller.
In order to break the bottleneck restricting the popularization and application of the SISO compact format model-free controller, the SISO compact format model-free controller also needs to solve the problem of online self-tuning parameters in the actual application process.
Disclosure of Invention
The invention aims to provide a parameter self-tuning method of a SISO (SISO) compact form model-free controller based on an Attention mechanism cyclic neural network, so as to solve the problem of parameter on-line self-tuning of the SISO compact form model-free controller.
To this end, the above object of the present invention is achieved by the following technical solution, comprising the steps of:
step (1): parameters of the SISO compact format model-free controller comprise a penalty factor lambda and a step factor rho; determining parameters to be set of a SISO (system-in-process) compact-format model-free controller, wherein the parameters to be set of the SISO compact-format model-free controller are part or all of the parameters of the SISO compact-format model-free controller and comprise any one or combination of a punishment factor lambda and a step factor rho; determining the number of input layer nodes, the number of hidden layer units and the number of output layer nodes of a cyclic neural network, wherein the number of the output layer nodes is not less than the number of parameters to be set of the SISO compact-format model-free controller; initializing all weight coefficients to be trained and learned in a hidden layer and an output layer of the recurrent neural network; determining a learning rate parameter value of the recurrent neural network back propagation; initializing partial derivatives in the gradient information set;
step (2): recording the current moment as k moment, and calculating by adopting a system error calculation function to obtain a system error of the k moment based on a system output expected value and a system output actual value, and recording as e (k);
and (3): recording any one or any combination of the system error and the function set thereof, the system output expected value and the system output actual value obtained by the calculation in the step (2) as a system error set, wherein the system error set is taken as a first part of an original input set; respectively taking partial derivatives of parameters to be set of each SISO tight-format model-free controller at corresponding time in the previous time control input u (k-1), the previous two time control input u (k-2) and the previous three time control input u (k-3) in the gradient information set as a second part of the original input set; the first part of the original input set and the second part of the original input set jointly form an original input set;
and (4): performing similarity calculation on the original input set in the step (3) and the hidden layer value of the recurrent neural network at the last moment, marking as a similarity score, inputting the similarity score into a normalization exponential function, namely a softmax function, and calculating to obtain a weight coefficient of each input in the recurrent neural network input set; multiplying a weight set formed by the weight coefficients with the original input set in the step (3) to obtain a new input set of the recurrent neural network at the current moment;
and (5): based on the new input set in the step (4), the recurrent neural network carries out forward calculation, firstly, the product of the input at the current moment and the weight coefficient to be trained and the product of the hidden layer state value at the previous moment and the weight coefficient to be trained are calculated, the results of the two products are summed and input to an activation function, and the hidden layer value at the current moment is obtained; multiplying the hidden layer value at the current moment by a weight coefficient and obtaining the value of a final output layer through an activation function, wherein the output value is used as the value of the parameter to be set of the SISO compact-format model-free controller;
and (6): calculating to obtain a control input u (k) of the SISO tight format model-free controller at the time k for the controlled object by adopting a control algorithm of the SISO tight format model-free controller based on the system error e (k) obtained in the step (2) and the value of the parameter to be set of the SISO tight format model-free controller obtained in the step (5);
and (7): based on the control input u (k) obtained in the step (6), calculating partial derivatives of the control input u (k) at the time k for the parameters to be set of each SISO compact-format model-free controller, wherein the specific calculation formula is as follows:
when the parameters to be set of the SISO compact-format model-free controller comprise a penalty factor lambda, the partial derivative of the control input u (k) at the moment k for the penalty factor lambda is as follows:
when the parameter to be set of the SISO compact-format model-free controller contains a step factor rho, the partial derivative of the control input u (k) at the k moment with respect to the step factor rho is as follows:
wherein phi (k) is a pseudo gradient estimation value at the k moment;
and (8): the method comprises the steps of taking the minimization of a value of a system error function as a target, adopting a gradient descent method based on a chain rule to carry out backward propagation calculation on a cyclic neural network, updating all weight coefficients to be trained and learned of the cyclic neural network, and taking the updated weight coefficients as the weight coefficients for carrying out forward calculation on the cyclic neural network at the later moment; the gradient descent method formula is as follows:
wherein w is a weight coefficient to be trained and learned; j (w) is a systematic error function with respect to w; alpha is the learning rate of the recurrent neural network training and is a real number between 0 and 1; in the back propagation calculation process, when all weight coefficients to be trained and learned of the recurrent neural network are updated, the partial derivatives of the parameters to be set of each SISO tight format model-free controller at the time k are respectively used according to the control input u (k) obtained in the step (7);
and (9): updating partial derivatives in the gradient information set, namely: recording partial derivatives of the control input u (k-2) at the first two moments in the corresponding moments of the parameters to be set of each SISO tight format model-free controller as partial derivatives of the control input u (k-3) at the first three moments in the corresponding moments of each SISO tight format model-free controller; recording partial derivatives of the control input u (k-1) at the previous moment at the corresponding moment of the parameters to be set of each SISO tight format model-free controller respectively as partial derivatives of the control input u (k-2) at the previous two moments at the corresponding moment of each SISO tight format model-free controller respectively; recording partial derivatives of the parameters to be set of each SISO tight-format model-free controller at the time k obtained in the step (7) at the current time control input u (k) as partial derivatives of the parameters to be set of each SISO tight-format model-free controller at the corresponding time at the previous time control input u (k-1), namely: when the parameters to be set of the SISO tight format model-free controller contain penalty factor lambda, the partial derivative at the k moment
Is recorded as partial derivative of previous time
When the parameters to be set of the SISO compact-format model-free controller contain step length factors rho, the partial derivative at the k moment
Is recorded as partial derivative of previous time
Step (10): and (e) after the control input u (k) acts on the controlled object, obtaining a system output actual value of the controlled object at the later moment, returning to the step (2), and repeating the steps (2) to (10).
A Recurrent Neural Network (RNN) is a time recursive Neural Network, and a Recurrent Neural Network structure includes an input layer, a hidden layer, and an output layer. All parameters of the traditional neural network are independent, and neurons at adjacent moments are independent, so that information sharing cannot be achieved. But the calculation process of the recurrent neural network synchronously calculates the hidden layer state at the previous moment and the input of the current moment, so that the current moment of the network can utilize the related information of the data at the previous moment, and the network operation structure ensures that the recurrent neural network has the learning capacity of the traditional feedforward neural network and the memory capacity of storing time sequence information. The SISO compact-format model-free controller can generate different time series information such as system control input and output, system errors and the like at each moment, setting of the controller parameters is closely connected with the system variables, and the cyclic neural network can well capture the time series characteristics of the input information and analyze and learn the internal connection between the parameters to be set and the time series information. Therefore, the problem of online self-tuning of parameters of the SISO compact-format model-free controller can be solved by utilizing the cyclic neural network.
Meanwhile, when the input information amount of the modeless controller is too large, the variable in the input set is too complicated, the recurrent neural network faces the difficult problems of heavy calculation burden, poor convergence and the like, the reason for this is that the recurrent neural network performs undifferentiated operation on the input data, and in the actual working condition, the input information of the modeless controller is not completely related to the parameter to be set, so that the proportion of irrelevant information in the input needs to be reduced. The Attention (Attention) mechanism can realize screening and screening of the importance of input information, the importance of each input variable is determined by calculating the relevance of the input at the current moment and a hidden layer at the last moment of the recurrent neural network, and compared with the traditional statistical learning methods such as PCA (principal component analysis), the Attention (Attention) mechanism has the advantages of simple calculation method, strong visualization function and the like, and is suitable for helping the recurrent neural network to improve the operation efficiency in the actual working condition.
While adopting the above technical scheme, the present invention can also adopt or combine the following further technical schemes:
the independent variables of the system error calculation function in the step (2) comprise a system output expected value and a system output actual value.
The systematic error calculation function employs e (k) y*(k) -y (k), wherein y*(k) The system output expected value is set for the time k, and y (k) is the system output actual value obtained by sampling at the time k; or using e (k) ═ y*(k +1) -y (k), wherein y*(k +1) is the system output expected value at the moment of k + 1; or using e (k) ═ y (k) — y*(k) (ii) a Or using e (k) ═ y (k) — y*(k+1)。
The system error and the function set thereof in the step (3) include the system error e (k) at the time k, and the accumulation of the system errors at the time k and all previous times, that is, the accumulation of the system errors
Any one or any combination of first order backward differences e (k) -e (k-1) of the k-time systematic error e (k), second order backward differences e (k) -2e (k-1) + e (k-2) of the k-time systematic error e (k), and high order backward differences of the k-time systematic error e (k).
The independent variable of the system error function in the step (8) comprises any one or any combination of a system error, a system output expected value and a system output actual value.
The system error function in the step (10) is ae2(k)+bΔu2(k) Where e (k) is the systematic error, Δ u (k) -u (k-1), and a and b are constants greater than or equal to 0.
The controlled object comprises a reactor, a rectifying tower, a machine, equipment, a device, a production line, a workshop and a factory.
The hardware platform for operating the control method comprises any one or any combination of a digital signal processing controller, an embedded system controller, a programmable logic controller, an industrial control computer, a single chip microcomputer controller, a microprocessor controller, a field programmable gate array controller, a distributed control system, a field bus control system, an industrial Internet of things control system and an industrial Internet control system.
The parameter self-setting method of the SISO tight format model-free controller based on the Attention mechanism recurrent neural network can realize good control effect and effectively overcome the problem that the punishment factor lambda and the step factor rho of the SISO tight format model-free controller need to be time-consuming and labor-consuming to set in industrial control.
Drawings
FIG. 1 is a functional block diagram of the present invention;
FIG. 2 is a schematic diagram of a three-tank SISO system;
FIG. 3 is a diagram of the control effect of the output of the SISO system of the three-water-tank system;
FIG. 4 is a control input curve of a three-tank SISO system;
FIG. 5 is a variation curve of penalty factor λ for control input of a three-tank SISO system;
FIG. 6 is a variation curve of step factor ρ inputted by the control of the SISO system of the three-tank SISO system;
FIG. 7 is a weight visualization of the control inputs of a three-tank SISO system.
Detailed Description
The invention is further described with reference to the following figures and specific examples.
Fig. 1 shows a schematic block diagram of the present invention. For SISO systems with a single input and a single output, adoptControlling by using a SISO compact format model-free controller; parameters of the SISO compact format model-free controller comprise a penalty factor lambda and a step factor rho; determining parameters to be set of a SISO (system-in-process) compact-format model-free controller, wherein the parameters to be set of the SISO compact-format model-free controller are part or all of the parameters of the SISO compact-format model-free controller and comprise any one or combination of a punishment factor lambda and a step factor rho; determining the number of input layer nodes, the number of hidden layer units and the number of output layer nodes of a cyclic neural network, wherein the number of the output layer nodes is not less than the number of parameters to be set of the SISO compact-format model-free controller; initializing all weight coefficients to be trained and learned in a hidden layer and an output layer of the recurrent neural network; determining a learning rate parameter value of the recurrent neural network back propagation; initializing partial derivatives in the gradient information set; recording the current time as k time; will output the expected value y*(k) Taking the difference between the system error and the output actual value y (k) as a system error e (k) at the time k, recording any one or any combination of the calculated system error and a function set thereof, a system output expected value and a system output actual value as a system error set, wherein the system error set is taken as a first part of an original input set; respectively taking partial derivatives of parameters to be set of each SISO tight-format model-free controller at corresponding time in the previous time control input u (k-1), the previous two time control input u (k-2) and the previous three time control input u (k-3) in the gradient information set as a second part of the original input set; the first portion of the original input set and the second portion of the original input set together comprise an original input set.
And performing similarity calculation on the original input set and the hidden layer value of the recurrent neural network at the previous moment, recording the similarity score as a similarity score, inputting the similarity score into a normalization exponential function, namely a softmax function, calculating to obtain each input weight coefficient in the recurrent neural network input set, and performing multiplication operation on the weight set formed by the weight coefficients and the original input set to obtain a new input set of the recurrent neural network at the current moment. Based on the new input set, the recurrent neural network carries out forward calculation, firstly, the product of the input at the current moment and the weight coefficient to be trained and the product of the hidden layer state value at the previous moment and the weight coefficient to be trained are calculated, the results of the two products are summed and input to an activation function, and the hidden layer value at the current moment is obtained. Similarly, the hidden layer value at the current moment is multiplied by a weight coefficient, the value of a final output layer is obtained through an activation function, and the output value is used as the value of the parameter to be set of the SISO compact-format model-free controller. Calculating to obtain control input u (k) of the SISO compact format model-free controller aiming at a controlled object at the time k by adopting a control algorithm of the SISO compact format model-free controller based on the system error and the value of a parameter to be set of the SISO compact format model-free controller; and calculating partial derivatives of the parameters to be set of each SISO compact format model-free controller at the time k by using the control input u (k).
The method aims at minimizing the value of a system error function, adopts a gradient descent method based on a chain rule to perform the back propagation calculation of the cyclic neural network, updates all weight coefficients to be trained and learned of the cyclic neural network, and uses the weight coefficients as the weight coefficients for performing the forward calculation of the cyclic neural network at the later moment; in the back propagation calculation process, when all weight coefficients to be trained and learned of the recurrent neural network are updated, the partial derivatives of parameters to be set of each SISO (SISO) compact-format model-free controller at the time k need to be respectively used by using a control input u (k); taking the current moment as an example, for the weight coefficient to be trained and learned of the output layer of the recurrent neural network, firstly, calculating the gradient of the error function about the weight coefficient of the output layer at the current moment, and updating the weight coefficient of the output layer at the current moment by using a gradient descent method; for all weight coefficients to be trained and learned in the hidden layer of the recurrent neural network, firstly, calculating the gradient value of an error function relative to the state of the hidden layer by using a chain rule, then obtaining a gradient algebraic expression of the error function relative to the ownership coefficient in the hidden layer by using a full derivative formula, and updating the weight coefficients by using a gradient descent method; the updating process of the weight coefficient to be trained in the input layer is the same as the weight coefficient of the hidden layer.
Updating partial derivatives in the gradient information set, i.e.: recording the partial derivative of the first two-time control input u (k-2) for each parameter to be set as the partial derivative of the first three-time control input u (k-3) for each parameter to be set; recording the partial derivative of the control input u (k-1) at the previous moment for each parameter to be set as the partial derivative of the control input u (k-2) at the previous two moments for each parameter to be set; and recording the partial derivative of the current moment control input u (k) for each parameter to be set as the partial derivative of the previous moment control input u (k-1) for each parameter to be set.
And (c) after the control input u (k) acts on the controlled object, obtaining an output actual value of the controlled object at the next moment, repeating the process, and performing a parameter self-tuning process of the SISO compact format model-free controller at the next moment based on the recurrent neural network.
The following is a specific embodiment of the present invention.
The controlled object three-container water tank is a single-input single-output SISO system, is a typical industrial object with complex characteristics of nonlinearity, large inertia and the like, the figure 2 is a schematic diagram of the three-container water tank and consists of 3 water tanks, wherein the actual output value y of the system is the liquid level height (cm) of the water tank 3, and the control input u is the valve opening (%) of a flow regulating valve flowing into the water tank 1. The initial working conditions of the three-container water tank are as follows: u (0) ═ 40%, y (0) ═ 50 cm. At 20 seconds, the system outputs a desired value y to meet the requirement of industrial field working condition adjustment*(20) Adjusting the length of the sample to be 60cm from 50 cm; subsequently, at 60 seconds, the system outputs the desired value y*(60) From 60cm again, adjust back to 50 cm. And three groups of tests are carried out for comparison and verification aiming at the typical actual working conditions of the industrial field. The hardware platform for operating the control method of the invention adopts an industrial control computer.
First set of experiments (RUN 1): the number of input layer nodes of the recurrent neural network is preset to be 9, the number of hidden layer nodes is preset to be 20, the number of output layer nodes is preset to be 2, wherein 2 output layer nodes respectively output punishment factors lambda and step factors rho; then, a control algorithm of a SISO compact format model-free controller is adopted to control the SISO system of the three water tanks; the RUN1 curve in fig. 3 is a graph of control effect of output, the RUN1 curve in fig. 4 is a curve of control input, the RUN1 curve in fig. 5 is a curve of penalty factor λ variation of control input, and the RUN1 curve in fig. 6 is a curve of step factor ρ variation of control input; FIG. 7 shows the importance weight distribution of RUN1 at the time 21 and the time 22; from the RUN1 curve of fig. 3, it can be found that the system output actual value can quickly track the change of the system output expected value, and meanwhile, the overshoot of the system output actual value is small, so that ideal control performance is realized; from the RUN1 curves shown in fig. 5 and fig. 6, it can be found that the penalty factor λ and the step factor ρ can be self-tuned online in time according to the change of the system error, so that the system can track and output the expected value more quickly, accurately and stably; as can be seen from fig. 7, the algorithm proposed in the present patent can calculate the importance distribution of the input information in real time, and reflects the effectiveness of the Attention mechanism algorithm. The method of the invention can realize good control effect by self-setting the penalty factor lambda and the step factor rho at the same time, and can effectively overcome the difficult problem that the penalty factor lambda and the step factor rho need to be time-consuming and labor-consuming to set.
Second set of experiments (RUN 2): the penalty factor lambda is fixed, and the value of the penalty factor lambda is the average value of the change curve of the penalty factor lambda in the first group of tests (RUN 1); the step factor rho is fixed and takes the value as the average value of the step factor rho change curve in the first group of experiments (RUN 1); then, a control algorithm of a SISO compact format model-free controller is adopted to control the SISO system of the three water tanks; the RUN2 curve in fig. 3 is a graph of control effect of the output, the RUN2 curve in fig. 4 is a graph of control input, the RUN2 curve in fig. 5 is a penalty factor λ of the control input, and the RUN2 curve in fig. 6 is a step factor ρ of the control input; from the RUN2 curve of fig. 3, it can be seen that the system output actual value can slowly track the change of the system output expected value, and the overshoot of the system output actual value is small; the second set of trials (RUN2) were inferior in terms of the rapidity index of control performance compared to the control performance of the first set of trials (RUN 1).
Third set of experiments (RUN 3): the values of the penalty factor lambda and the step factor rho are both fixed to be common values of 0.5; then, a control algorithm of a SISO compact format model-free controller is adopted to control the SISO system of the three water tanks; the RUN3 curve in fig. 3 is a graph of control effect of output, and the RUN3 curve in fig. 4 is a graph of control input; from the RUN3 curve of fig. 3, it can be seen that the system output actual value can track the change of the system output expected value at the fastest speed, but the overshoot of the system output actual value is large; the third set of tests (RUN3) was inferior in stability index of control performance compared to the control performance of the first set of tests (RUN 1).
The results of the three groups of tests show that the parameter self-tuning method of the SISO compact-format model-free controller based on the recurrent neural network of the Attention mechanism adopted in the first group of tests (RUN1) has the optimal control performance comprehensive index, and meanwhile, the Attention mechanism can help the recurrent neural network to screen the important part of input information in real time and has the optimization effect on the calculation of the recurrent neural network.
It should be noted that in the above-described embodiment, the desired value y will be output*(k) The difference from the output actual value y (k) is used as the system error e (k) at time k, i.e. e (k) y*(k) -y (k), only one method of calculating a function for said error; the system at the moment k +1 can also output the expected value y*The difference between (k +1) and the time k output y (k) is taken as the system error e (k), i.e. e (k) y (k)*(k +1) -y (k); the error calculation function may also employ other methods of calculating the output desired value and the output actual value, such as, for example,
for the controlled object of the above embodiment, good control effects can be achieved by using the different system error calculation functions.
It should be more particularly noted that, in the above-described embodiment, when the implicit layer weight coefficients and the output layer weight coefficients of the recurrent neural network are updated with the goal of minimizing the value of the systematic error function, the systematic error function employs the square of the systematic error e2(k) Only one of said systematic error functions; for example, the systematic error function can also be ae2(k)+bΔu2(k) Wherein Δ u (k) -u (k-1), a and b are constants greater than or equal to 0; it is clear that the systematic error function only takes into account e when b equals 02(t) contribution, indicating that the objective of minimization is to minimize the systematic errorNamely, the pursuit precision is high; and when b is greater than 0, the systematic error function takes e into account2Contribution of (t) and Δ u2The contribution of (t) indicates that the goal of minimization is to pursue small system error and small control input variation, i.e., both high precision and stable steering. For the controlled object of the above embodiment, good control effect can be achieved by adopting the different system error functions; considering only e with the systematic error function2(k) Control effects in contribution to the system error function while considering e2Contribution of (t) and Δ u2The contribution of (t) is that the control precision is slightly reduced and the operation stability is improved.
Finally, it should be particularly noted that the parameters to be set by the SISO compact-format model-less controller include one or all of a penalty factor λ and a step factor ρ; in the above specific embodiment, the penalty factor λ and the step factor ρ realize simultaneous self-tuning during experimental verification; in practical application, any combination of parameters to be set can be selected according to specific conditions, for example, the step factor rho is fixed, and the penalty factor realizes self-setting; in addition, the parameters to be set by the SISO compact-format modeless controller include, but are not limited to, a penalty factor λ and a step factor ρ, and for example, according to the specific situation, the parameters may also include parameters such as an estimation value of pseudo-partial derivative Φ (k) of the SISO system.
The above-described embodiments are intended to illustrate the present invention, but not to limit the present invention, and any modifications, equivalents, improvements, etc. made within the spirit of the present invention and the scope of the claims fall within the scope of the present invention.
Claims (8)
- A parameter self-tuning method of a SISO (SISO) compact format model-free controller based on an Attention mechanism cyclic neural network is characterized by comprising the following steps of:step (1): parameters of the SISO compact format model-free controller comprise a penalty factor lambda and a step factor rho; determining parameters to be set of a SISO (system-in-process) compact-format model-free controller, wherein the parameters to be set of the SISO compact-format model-free controller are part or all of the parameters of the SISO compact-format model-free controller and comprise any one or combination of a punishment factor lambda and a step factor rho; determining the number of input layer nodes, the number of hidden layer units and the number of output layer nodes of a cyclic neural network, wherein the number of the output layer nodes is not less than the number of parameters to be set of the SISO compact-format model-free controller; initializing all weight coefficients to be trained and learned in a hidden layer and an output layer of the recurrent neural network; determining a learning rate parameter value of the recurrent neural network back propagation; initializing partial derivatives in the gradient information set;step (2): recording the current moment as k moment, and calculating by adopting a system error calculation function to obtain a system error of the k moment based on a system output expected value and a system output actual value, and recording as e (k);and (3): recording any one or any combination of the system error and the function set thereof, the system output expected value and the system output actual value obtained by the calculation in the step (2) as a system error set, wherein the system error set is taken as a first part of an original input set; respectively taking partial derivatives of parameters to be set of each SISO tight-format model-free controller at corresponding time in the previous time control input u (k-1), the previous two time control input u (k-2) and the previous three time control input u (k-3) in the gradient information set as a second part of the original input set; the first part of the original input set and the second part of the original input set jointly form an original input set;and (4): performing similarity calculation on the original input set in the step (3) and the hidden layer value of the recurrent neural network at the last moment, marking as a similarity score, inputting the similarity score into a normalization exponential function, namely a softmax function, and calculating to obtain a weight coefficient of each input in the recurrent neural network input set; multiplying a weight set formed by the weight coefficients with the original input set in the step (3) to obtain a new input set of the recurrent neural network at the current moment;and (5): based on the new input set in the step (4), the recurrent neural network carries out forward calculation, firstly, the product of the input at the current moment and the weight coefficient to be trained and the product of the hidden layer state value at the previous moment and the weight coefficient to be trained are calculated, the results of the two products are summed and input to an activation function, and the hidden layer value at the current moment is obtained; multiplying the hidden layer value at the current moment by a weight coefficient and obtaining the value of a final output layer through an activation function, wherein the value of the final output layer is used as the value of a parameter to be set of the SISO compact-format model-free controller;and (6): calculating to obtain a control input u (k) of the SISO tight format model-free controller at the time k for the controlled object by adopting a control algorithm of the SISO tight format model-free controller based on the system error e (k) obtained in the step (2) and the value of the parameter to be set of the SISO tight format model-free controller obtained in the step (5);and (7): based on the control input u (k) obtained in the step (6), calculating partial derivatives of the control input u (k) at the time k for the parameters to be set of each SISO compact-format model-free controller, wherein the specific calculation formula is as follows:when the parameters to be set of the SISO compact-format model-free controller comprise a penalty factor lambda, the partial derivative of the control input u (k) at the moment k for the penalty factor lambda is as follows:
when the parameter to be set of the SISO compact-format model-free controller contains a step factor rho, the partial derivative of the control input u (k) at the k moment with respect to the step factor rho is as follows:
wherein phi (k) is a pseudo gradient estimation value at the k moment;and (8): the method comprises the steps of taking the minimization of a value of a system error function as a target, adopting a gradient descent method based on a chain rule to carry out backward propagation calculation on a cyclic neural network, updating all weight coefficients to be trained and learned of the cyclic neural network, and taking the updated weight coefficients as the weight coefficients for carrying out forward calculation on the cyclic neural network at the later moment; the gradient descent method formula is as follows:
wherein w is a weight coefficient to be trained and learned; j (w) is a systematic error function with respect to w; alpha is the learning rate of the recurrent neural network training and is a real number between 0 and 1; in the back propagation calculation process, when all weight coefficients to be trained and learned of the recurrent neural network are updated, the partial derivatives of the parameters to be set of each SISO tight format model-free controller at the time k are respectively used according to the control input u (k) obtained in the step (7);and (9): updating partial derivatives in the gradient information set, namely: recording partial derivatives of the control input u (k-2) at the first two moments in the corresponding moments of the parameters to be set of each SISO tight format model-free controller as partial derivatives of the control input u (k-3) at the first three moments in the corresponding moments of each SISO tight format model-free controller; recording partial derivatives of the control input u (k-1) at the previous moment at the corresponding moment of the parameters to be set of each SISO tight format model-free controller respectively as partial derivatives of the control input u (k-2) at the previous two moments at the corresponding moment of each SISO tight format model-free controller respectively; recording partial derivatives of the parameters to be set of each SISO tight-format model-free controller at the time k obtained in the step (7) at the current time control input u (k) as partial derivatives of the parameters to be set of each SISO tight-format model-free controller at the corresponding time at the previous time control input u (k-1), namely: when the parameters to be set of the SISO tight format model-free controller contain penalty factor lambda, the partial derivative at the k momentIs recorded as partial derivative of previous time
When the parameters to be set of the SISO compact-format model-free controller contain step length factors rho, the partial derivative at the k moment
Is recorded as partial derivative of previous time
Step (10): and (e) after the control input u (k) acts on the controlled object, obtaining a system output actual value of the controlled object at the later moment, returning to the step (2), and repeating the steps (2) to (10). - 2. The SISO compact format model-less controller parameter self-tuning method of claim 1, wherein the argument of the system error computation function in step (2) comprises a system output expected value and a system output actual value.
- 3. The SISO tight format model-less controller of claim 2, wherein the system error computation function employs e (k) -y*(k) -y (k), wherein y*(k) The system output expected value is set for the time k, and y (k) is the system output actual value obtained by sampling at the time k; or using e (k) ═ y*(k +1) -y (k), wherein y*(k +1) is the system output expected value at the moment of k + 1; or using e (k) ═ y (k) — y*(k) (ii) a Or using e (k) ═ y (k) — y*(k+1)。
- 4. The SISO compact-format model-less controller parameter self-tuning method of claim 1, wherein the system error and its function set in step (3) comprises the system error at time k, e (k), and all the previous k timesAccumulation of systematic errors at time of dayAny one or any combination of first order backward differences e (k) -e (k-1) of the k-time systematic error e (k), second order backward differences e (k) -2e (k-1) + e (k-2) of the k-time systematic error e (k), and high order backward differences of the k-time systematic error e (k).
- 5. The SISO compact format model-less controller parameter self-tuning method of claim 1, wherein the argument of the system error function in step (8) comprises any one or any combination of system error, system output desired value and system output actual value.
- 6. The SISO compact-format model-less controller of claim 5, wherein the system error function is ae2(k)+bΔu2(k) Where e (k) is the systematic error, Δ u (k) -u (k-1), and a and b are constants greater than or equal to 0.
- 7. The SISO compact format model-less controller of claim 1, based on the parameter self-tuning method of the Attention mechanism recurrent neural network, characterized in that: the controlled object comprises a reactor, a rectifying tower, a machine, equipment, a device, a production line, a workshop and a factory.
- 8. The SISO compact format model-less controller of claim 1, based on the parameter self-tuning method of the Attention mechanism recurrent neural network, characterized in that: the hardware platform for operating the SISO compact-format model-free controller comprises any one or any combination of a digital signal processing controller, an embedded system controller, a programmable logic controller, an industrial control computer, a single chip microcomputer controller, a microprocessor controller, a field programmable gate array controller, a distributed control system, a field bus control system, an industrial Internet of things control system and an industrial Internet control system.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010560142.9A CN111781821B (en) | 2020-06-18 | 2020-06-18 | Parameter self-tuning method of SISO (SISO) compact-format model-free controller based on Attention mechanism cyclic neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010560142.9A CN111781821B (en) | 2020-06-18 | 2020-06-18 | Parameter self-tuning method of SISO (SISO) compact-format model-free controller based on Attention mechanism cyclic neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111781821A CN111781821A (en) | 2020-10-16 |
| CN111781821B true CN111781821B (en) | 2021-09-24 |
Family
ID=72757310
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010560142.9A Active CN111781821B (en) | 2020-06-18 | 2020-06-18 | Parameter self-tuning method of SISO (SISO) compact-format model-free controller based on Attention mechanism cyclic neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111781821B (en) |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6360131B1 (en) * | 1998-10-16 | 2002-03-19 | Cheng | Model-free adaptive control for flexible production systems |
| CN106354016B (en) * | 2016-11-04 | 2019-03-08 | 重庆大学 | A kind of control method of metal tube compacting height |
| CN108073072B (en) * | 2017-11-06 | 2020-06-09 | 浙江大学 | Parameter self-tuning method of SISO (Single input Single output) compact-format model-free controller based on partial derivative information |
| CN109782586B (en) * | 2019-02-01 | 2021-01-05 | 浙江大学 | Parameter self-tuning MISO different-factor compact-format model-free control method |
-
2020
- 2020-06-18 CN CN202010560142.9A patent/CN111781821B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN111781821A (en) | 2020-10-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105807607B (en) | A kind of genetic algorithm optimization predictive fuzzy PID coking furnace temprature control methods | |
| CA2718911C (en) | Method of multi-dimensional nonlinear control | |
| Xiong et al. | Neural network model-based on-line re-optimisation control of fed-batch processes using a modified iterative dynamic programming algorithm | |
| CN109254530A (en) | MFA control method based on grinding process basis circuit | |
| CN111142494B (en) | Intelligent control method and system for amine liquid regeneration device | |
| JP7034046B2 (en) | Control device and control method | |
| CN111522229A (en) | Parameter self-tuning MIMO different-factor offset format model-free control method | |
| CN109143853B (en) | Self-adaptive control method for liquid level of fractionating tower in petroleum refining process | |
| CN104898426A (en) | Room temperature loop control method based on gradient descent method and generalized prediction control | |
| CN114384931B (en) | Multi-target optimal control method and equipment for unmanned aerial vehicle based on strategy gradient | |
| CN109782586B (en) | Parameter self-tuning MISO different-factor compact-format model-free control method | |
| CN101285816A (en) | Copper matte air refining procedure parameter soft sensing instrument and its soft sensing method | |
| Alitasb et al. | Multiple-input multiple-output Radial Basis Function Neural Network modeling and model predictive control of a biomass boiler | |
| CN111752152B (en) | LSTM neural network-based parameter self-tuning method for SISO (SISO) compact format model-free controller | |
| CN111781821B (en) | Parameter self-tuning method of SISO (SISO) compact-format model-free controller based on Attention mechanism cyclic neural network | |
| CN108181809A (en) | Tight methods of self-tuning of the form Non-Model Controller based on systematic error of MISO | |
| CN108319146B (en) | Discrete particle swarm training-based method for radial basis function neural network | |
| CN112015083B (en) | Parameter self-tuning method of SISO (SISO) compact-format model-free controller based on ensemble learning | |
| CN112015081B (en) | Parameter self-tuning method of SISO (SISO) compact-format model-free controller based on PSO-LSTM (particle swarm optimization-least Square transform) cooperative algorithm | |
| CN109814388B (en) | Parameter self-tuning MISO different-factor offset format model-free control method | |
| Hung et al. | Design of multi model predictive control for nonlinear process plant | |
| CN114690628A (en) | Pseudo Jacobian matrix parameter self-tuning MIMO (multiple input multiple output) compact format model-free control method | |
| Zhao et al. | An identification approach to nonlinear state space model for industrial multivariable model predictive control | |
| Krishnan et al. | Robust model predictive control of an industrial solid phase polymerizer | |
| CN116149262B (en) | Tracking control method and system of servo system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |