CN-117572829-B - Multi-mode industrial process all-condition high real-time prediction control method and equipment

CN117572829BCN 117572829 BCN117572829 BCN 117572829BCN-117572829-B

Abstract

The invention discloses a multi-mode industrial process full-working condition high real-time prediction control method and equipment, when learning full-working condition display control law, the method firstly uses a working condition dataset to learn the explicit control law of the current working condition, namely, learn the parameters of a self-organizing fuzzy neural network; determining whether to increase fuzzy rules according to the data coverage rate of the learned working conditions, if the working conditions change in a small range, introducing an elastic weight consolidation item mechanism on the basis of original loss to ensure that a new working condition control strategy is learned and the control performance of the historical working conditions is kept, if the working conditions change in a large range, adopting a truncated radial basis neuron growth mechanism to learn the control strategy of the new working conditions by increasing the fuzzy rules, so that the explicit control law can adapt to the change in the large range of working conditions, and when in online control, obtaining a control sequence according to the current control state by using the full working condition explicit control law. The invention solves the problem of traditional multi-model online solving optimization and realizes the accurate control effect of the whole working condition of the industrial process.

Inventors

HUANG KEKE
YING XINYU
WU DEHAO
LIU YISHUN
SUN BEI
YANG CHUNHUA
GUI WEIHUA

Assignees

中南大学

Dates

Publication Date: 20260512
Application Date: 20231019

Claims (10)

1. The method for predicting and controlling the full working condition of the multi-mode industrial process in high real time is characterized by comprising the following steps: Offline learning: Respectively learning a corresponding output state prediction model by using a data set of each working condition of the industrial system, and learning a full-working-condition explicit control law based on a self-organizing fuzzy neural network by using data sets of all working conditions; When learning the full-working condition display control law, firstly, learning the explicit control law of the current working condition by using the data set of the first working condition, namely, learning the parameters of the self-organizing fuzzy neural network, and then sequentially adjusting the explicit control law which is learned by the current working condition by using the data sets of the other working conditions, namely, directly updating and learning the current network parameters if the data coverage rate of the current learned working condition does not meet the preset condition, and firstly, adding the fuzzy rule, namely, adding the neurons representing the fuzzy rule in the self-organizing fuzzy neural network, and updating and learning the parameters of the self-organizing fuzzy neural network if the data coverage rate of the current learned working condition meets the preset condition; on-line control: The output state at the current moment, the historical output state sequence, the control sequence and the reference track of the future output state are constructed together to serve as control state data at the current moment, and the control state data are input into the learned full-working-condition explicit control law to obtain an optimal control sequence at the current moment; The industrial system is controlled using the first control variable of the optimal control sequence at the current time.
2. The multi-modal industrial process full-condition high real-time predictive control method of claim 1, wherein the self-organizing fuzzy neural network comprises an input layer, a fuzzy layer, a normalization layer and an output layer; The input layer comprises Neurons, respectively representing fuzzy neural networks The mathematical expression of the input layer is as follows: (1) Wherein, the Representing the input of the ith neuron at time t, The output of the ith neuron at time t; The input of the fuzzy neural network at the moment t is shown and is used for inputting control state data of the industrial system at the moment t; the output of the input layer at the time t is represented; The fuzzy layer is provided with P groups of neurons, each group of neurons represents a fuzzy rule, wherein each fuzzy rule adopts a radial basis function as a membership function of a fuzzy input variable, and the mathematical expression is as follows: (2) Wherein, the Representing the membership function of the ith input variable corresponding to the jth fuzzy rule, And After obtaining the membership degree of each input variable, calculating the membership degree of each fuzzy rule, wherein the mathematical expression is as follows: (3) Wherein, the Representation of Time of day (time) The output of the group of radial basis neurons, Representation of Time of day (time) The group is radial to the center of the basal neurons, Representation of Time of day (time) The width of the group radial basis neurons; the normalization layer has P neurons, and the number of the neurons is the same as that of the fuzzy rules of the fuzzy layer, and the neurons are used for normalizing the output of the fuzzy layer to obtain normalized output ; The output layer is a linear layer, and the mathematical expression of the output is as follows: ; Wherein, the The output of the self-organizing fuzzy neural network at the time t is represented and is used for outputting a control variable of the industrial system at the time t, Representing the weight matrix of the output layer at time t, Is the number of neurons in the output layer.
3. The multi-modal industrial process full-condition high real-time predictive control method according to claim 2, wherein the learning of the explicit control law of the current condition using the data set of the first condition specifically converts the optimization problem of the model predictive control strategy into a loss function to learn the explicit control law; Firstly, the optimization problem of the model predictive control strategy is that the objective function only considers the tracking performance of the output state of the industrial system, and the constraint condition only considers the upper and lower limiting constraints of the control variable, as follows: (6) Wherein the method comprises the steps of Representing a neural network predictive model, Representing the predicted output state of the prediction model, A reference trajectory is represented and a reference trajectory is represented, Representing the predicted control variable(s), And Representing the upper and lower limits of the control variable, Representing the predicted time-domain range, Representing the control time domain range, 、 Representing the order of the predictive model with respect to the input control variable u and the output state y; the objective function is then converted into a loss function as: (7) Wherein, the Representing the batch size of the batch training of the ad hoc fuzzy neural network, The sample index for each batch is represented, And The weight factors representing the output states and control variables, , , And The upper and lower limit penalty term representing the control variable is specifically as follows: (8) Finally, the explicit control law of the first working condition is learned through the model predictive control loss function shown in the formula (7), namely, the parameter learning of the self-organizing fuzzy neural network is completed through the gradient of the loss function shown in the following back propagation formula, and the parameter learning is as follows: ; Wherein, the Weights representing the output layers of the ad hoc fuzzy neural network, Representing the width of the blur layer neurons of the ad hoc blur neural network, Representing the centers of the blur layer neurons of the self-organizing blur neural network, Representing the learning rate.
4. The method for high real-time predictive control of all operating modes in a multi-modal industrial process as defined in claim 3 wherein the data coverage rate is calculated by first using a truncated radial basis function as an activation function of neurons in a fuzzy layer of a self-organizing fuzzy neural network Time of day (time) Output of fuzzy rule Expressed as: (12) Wherein, the Representing the truncated radial basis function, Representation of Time of day (time) The first fuzzy rule The output of the individual neurons is referred to as, Representation of Time of day (time) The first fuzzy rule The activation state of the individual neurons and, , Represent the first The activation state of the individual fuzzy rules is determined, The upper and lower cutoff limits of the radial basis functions are set as follows: (13) Wherein, the Is a positive integer; then according to the activation state of all fuzzy rules Judging the first batch of data Status signal of whether individual data is in the range of fuzzy rule And calculating the data coverage rate of the current self-organizing fuzzy neural network according to the status signals of all the data Expressed as: (14) (15)。
5. The method for high real-time predictive control of full operating mode in a multi-modal industrial process according to claim 4, wherein the preset conditions are: (16) Wherein, the Representing a threshold.
6. The method for high real-time predictive control of full operating mode in a multi-modal industrial process according to claim 4, wherein the adding of the fuzzy rule is to add neurons representing the fuzzy rule in the ad hoc fuzzy neural network, and the specific newly added neuron group is designed as follows: (17) Wherein, the Representing the data quantity which is not covered by the self-organizing fuzzy neural network in the current working condition to be learned, Representing data not covered by the ad hoc fuzzy neural network, And Representation of The moment newly added fuzzy rule corresponds to the center and width of the truncated radial basis neurons, Representing the weight of the corresponding output layer of the new fuzzy rule, And And randomly initializing by the self-organizing fuzzy neural network.
7. The method for high real-time predictive control of full operating mode in multi-modal industrial process according to claim 6, wherein the parameters of the self-organizing fuzzy neural network are updated and learned after adding the fuzzy rule, and the loss function is specifically: (18) Wherein, the Representing the current working condition to be learned Model predictive control loss function of (2); representing current network parameters of the ad hoc fuzzy neural network, Represent the first The parameters of the individual networks are set to be, Indicating the completed working condition Obtained after explicit control law learning The parameters of the individual networks are set to be, Indicating a set of completed conditions Model predictive control loss function of (c) For a pair of Is used for the first partial derivative of (c), Indicating a set of completed conditions Is of importance of (a); 、、 representing a set of three network parameters and for the jth fuzzy rule, Represent the first The parameters of the individual networks are set to be, Indicating the completed working condition Obtained after explicit control law learning The parameters of the individual networks are set to be, Indicating a set of completed conditions Model predictive control loss function of (c) For a pair of Second partial derivative of (2); Representing the activation state of the jth fuzzy rule; Then, the parameter updating of the self-organizing fuzzy neural network is completed through the gradient of the loss function shown in the following back propagation formula, as follows: (10)。
8. The method for high real-time predictive control of full operating mode in a multi-modal industrial process as set forth in claim 7, wherein the width of the radial basis neurons is additionally limited during the update learning of the parameters of the ad hoc fuzzy neural network The ranges of (2) are as follows: (19) Wherein, the Represent the first The first fuzzy rule The width of the individual neurons and the number of the individual neurons, And Respectively representing upper and lower clipping.
9. The method for high real-time predictive control of full operating mode in multi-modal industrial process according to claim 3, wherein if the data coverage rate of the current learned operating mode does not meet the preset condition, the current network parameters are directly updated and learned, and the loss function adopted by the updated and learned is: (9) Wherein, the Representing the current working condition to be learned Model predictive control loss function of (2); representing current network parameters of the ad hoc fuzzy neural network, Representation of The first of (3) The parameters of the individual networks are set to be, Representing a set of all conditions for which explicit control learning has been completed, Indicating a set of completed conditions Obtained after explicit control law learning The parameters of the individual networks are set to be, Indicating a set of completed conditions Model predictive control loss function of (c) For a pair of Is used for the first partial derivative of (c), Indicating a set of completed conditions Is of importance of (a); Then, the parameter updating of the self-organizing fuzzy neural network is completed through the gradient of the loss function shown in the following back propagation formula, as follows: (10)。
10. An electronic device comprising a memory and a processor, the memory having stored therein a computer program, wherein execution of the computer program by the processor causes the processor to implement the method of any of claims 1-9.

Description

Multi-mode industrial process all-condition high real-time prediction control method and equipment Technical Field The invention belongs to the technical field of industrial control, and particularly relates to a method and equipment for high-real-time prediction control of all working conditions of a multi-mode industrial process. Background With the large-scale of modern industrial systems, there is a great deal of uncertainty in the production environment, process, etc., and industrial systems exhibit complex operating characteristics, often making it difficult to build accurate mathematical models. Conventional control methods may be limited in the face of these complex systems, and it is difficult to meet the actual demands. Therefore, new computing methods and ideas need to be introduced to address challenges facing industrial systems. The fuzzy theory gives a set of systematic and effective methods for converting the knowledge described by natural language into mathematical expression form, so that the influence of uncertainty of complex systems can be overcome by expert knowledge. However, conventional fuzzy control systems typically use fuzzy reasoning and fuzzy rules to handle the relationships between inputs and outputs, with limited modeling capabilities for complex nonlinear systems. The fuzzy neural network is a nonlinear modeling tool, combines the advantages of fuzzy logic and the neural network, and can better process a nonlinear system. It can capture complex relationships between inputs and outputs through multi-layer connections and nonlinear activation functions, thereby more accurately modeling and controlling nonlinear systems. In addition to the uncertainty of the production process, some industrial systems are actually a dynamic multi-modal process, the state of which switches between different operating conditions, due to the influence of factors such as diversification of production raw materials, complexity of the production process, etc. Thus, how to achieve precise control of complex industrial systems under varying conditions is an important and challenging problem. Model predictive control (Model Predictive Control, MPC) is an efficient industrial control method, a control algorithm currently recognized to be able to efficiently handle multi-variable complex processes and take into account a variety of constraints. The basic idea of model predictive control is that based on a predictive model of a process, an optimal control sequence is obtained by solving a finite time domain open-loop optimal control problem at each control interval, and then a first control quantity of the optimal control sequence is issued to an actual process. At present, model predictive control has been widely applied to complex industrial systems such as aerospace, autopilot, robotics and nonferrous metallurgy. In the face of multi-operating industrial processes, model predictive control typically employs a multi-model or multi-controller control strategy to cope with fluctuations in the operating state of the industrial process. However, since there are a plurality of local predictive controllers operating simultaneously, the scroll optimization requires a large amount of computing resources and computing time. Especially for industrial processes with high sampling frequency or fast process variation, the above method often has difficulty in meeting the real-time requirement. The explicit model predictive control (Explicit Model Predictive Control, EMPC) is a high real-time control method, the method obtains an explicit control law through offline training, and the optimal control sequence can be obtained by substituting system state information at the current moment into the explicit control law for calculation during online operation, so that real-time rolling optimization solution is omitted, and rapid control is realized. The explicit model prediction control method verifies the strong capability of the explicit model prediction control method on the aspects of control instantaneity and the like, but the explicit model prediction control method still has some defects. First, as the problem size increases, such as increasing prediction horizon, increasing number of constraints, increasing system input-output dimensions, etc., the size of the explicit control laws increases exponentially. On the other hand, explicit model predictive control is initially applied to a linear time-invariant process, and although some studies have been made to apply explicit model predictive control to a nonlinear process, this inevitably results in an increase in the complexity of offline computation and an increase in the scale of explicit control laws. Compared with the piecewise affine function, the neural network has stronger parallel computing capacity and nonlinear fitting capacity, and provides a new thought for solving the large-scale multiparameter quadratic programming problem. Explicit control laws based on