A multimodal industrial process full-operating-mode high real-time predictive control method and device

Abstract

The application discloses a kind of multimode industrial process full working condition high real-time prediction control method and equipment, method in learning full working condition display control law: first using a working condition dataset to learn the explicit control law of current working condition, i.e. learning the parameters of self-organizing fuzzy neural network;Again, according to the data coverage of the learned working condition, determine whether to increase fuzzy rule;If it is small range working condition change, then introduce elastic weight consolidation term mechanism on the basis of original loss, ensure that the control strategy of new working condition is learned while the control performance of historical working condition is maintained;If it is large range working condition change, adopt truncated radial basis neuron growth mechanism, learn the control strategy of new working condition by increasing fuzzy rule, so that the explicit control law can adapt to large range working condition change;When online control, using full working condition explicit control law to obtain control sequence according to current control state.The application overcomes the traditional multi-model online solving optimization problem, realizes the precise control effect of industrial process full working condition.

Description

Technical Field

[0001] This invention belongs to the field of industrial control technology, specifically relating to a multimodal industrial process full-condition high real-time predictive control method and equipment. Background Technology

[0002] With the increasing scale of modern industrial systems, significant uncertainties exist in production environments and processes, leading to complex operational characteristics that often make accurate mathematical models difficult to establish. Traditional control methods may be limited in addressing these complex systems and fail to meet practical needs. Therefore, new computational methods and ideas are needed to address the challenges faced by industrial systems. Fuzzy theory provides a systematic and effective method to transform knowledge described in natural language into mathematical expressions, thereby overcoming the influence of uncertainties in complex systems through expert knowledge. However, traditional fuzzy control systems typically use fuzzy inference and fuzzy rules to handle the relationship between inputs and outputs, which has limited ability to model complex nonlinear systems. Fuzzy neural networks, a nonlinear modeling tool, combine the advantages of fuzzy logic and neural networks, enabling better handling of nonlinear systems. They can capture the complex relationships between inputs and outputs through multi-layer connections and nonlinear activation functions, thus modeling and controlling nonlinear systems more accurately.

[0003] Besides the uncertainties inherent in the production process, the diversification of raw materials and the complexity of production processes mean that some industrial systems are actually dynamic, multimodal processes, switching between different operating conditions. Therefore, achieving precise control of complex industrial systems under varying operating conditions is a crucial and challenging problem. Model Predictive Control (MPC) is an efficient industrial control method, currently recognized as an effective control algorithm for handling complex multivariate processes and considering multiple constraints. The basic idea of ​​MPC is to use a predictive model of the process as a foundation, solving a finite-time open-loop optimal control problem at each control interval to obtain the optimal control sequence, and then issuing the first control variable of the optimal control sequence to the actual process. Currently, MPC is widely used in complex industrial systems such as aerospace, autonomous driving, robotics, and non-ferrous metallurgy. When facing multi-condition industrial processes, MPC typically employs multi-model or multi-controller control strategies to address the fluctuations in the process's operating state. However, due to the simultaneous operation of multiple local predictive controllers, rolling optimization consumes significant computational resources and time. Especially for industrial processes with high sampling frequencies or rapid process changes, the above methods often fail to meet real-time requirements.

[0004] Explicit Model Predictive Control (EMPC) is a high-real-time control method. This method obtains an explicit control law through offline training, and during online execution, it only needs to substitute the current system state information into the explicit control law to calculate the optimal control sequence, eliminating the need for real-time rolling optimization and achieving fast control. EMPC has demonstrated its powerful capabilities in real-time control, but it still has some limitations. First, as the problem size increases, such as with an increased prediction time domain, a greater number of constraints, or an increased system input / output dimension, the size of the explicit control law grows exponentially. Second, EMPC was initially applied to linear time-invariant processes. Although some research has now applied EMPC to nonlinear processes, this inevitably leads to increased offline computational complexity and a larger explicit control law. Compared to piecewise affine functions, neural networks possess more powerful parallel computing capabilities and nonlinear fitting abilities, providing a new approach to solving large-scale multi-parameter quadratic programming problems. Explicit control laws based on neural networks can provide smaller control law size and faster computation speed while ensuring that control performance meets requirements. However, when facing multi-condition industrial processes with fluctuating operating states, pre-trained explicit control laws may experience mismatch, severely affecting control performance. Summary of the Invention

[0005] To address the problem of real-time control of multimodal processes under all operating conditions, this invention provides a high real-time predictive control method and device for multimodal industrial processes under all operating conditions. It proposes and adopts a self-organizing fuzzy neural network explicit control law learning framework oriented to all operating conditions. By self-adjusting network parameters and structure, it adapts to all operating conditions. The learned control law can eliminate the impact of control strategy mismatch during switching phases and improve operating speed.

[0006] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0007] A multimodal industrial process full-condition high real-time predictive control method includes:

[0008] Offline learning:

[0009] We learn the corresponding output state prediction model using the dataset for each operating condition of the industrial system, and learn the full-condition explicit control law based on a self-organizing fuzzy neural network using the dataset for all operating conditions.

[0010] In the process of learning the explicit control law for all operating conditions, the first step is to use the dataset of the first operating condition to learn the explicit control law for the current operating condition, which means learning the parameters of the self-organizing fuzzy neural network. Then, the datasets of the other operating conditions are used to adjust the currently learned explicit control law in turn: if the data coverage of the currently learned operating condition does not meet the preset conditions, the current network parameters are directly updated and learned; if the data coverage of the currently learned operating condition meets the preset conditions, fuzzy rules are added first, that is, neurons representing fuzzy rules are added in the self-organizing fuzzy neural network, and then the parameters of the self-organizing fuzzy neural network are updated and learned.

[0011] Online control:

[0012] The current output state, together with the historical output state sequence, control sequence, and reference trajectory of future output state, are used to construct the control state data for the current moment. This data is then input into the learned full-condition explicit control law to obtain the optimal control sequence for the current moment.

[0013] The industrial system is controlled using the first control variable of the optimal control sequence at the current moment.

[0014] Furthermore, the self-organizing fuzzy neural network includes an input layer, a fuzzy layer, a normalization layer, and an output layer;

[0015] The input layer includes Each neuron represents a fuzzy neural network. Dimensional input variables; the mathematical expression for the input layer is shown below:

[0016] (1)

[0017] in, This represents the input of the i-th neuron at time t. This represents the output of the i-th neuron at time t; This represents the input to the fuzzy neural network at time t, used to input the control state data of the industrial system at time t; This represents the output of the input layer at time t;

[0018] The fuzzy layer consists of P groups of neurons, each group representing a fuzzy rule. Each fuzzy rule uses a radial basis function as the membership function of the fuzzification input variable, as shown in the following mathematical expression:

[0019] (2)

[0020] in, Let represent the membership function of the i-th input variable corresponding to the j-th fuzzy rule. and Here, we define the center and width of the radial basis function neurons. After obtaining the membership degree of each input variable, we calculate the membership degree of each fuzzy rule, with the following mathematical expression:

[0021] (3)

[0022] in, express Time of the first The output of the radial basis neurons, express Time of the first The center of the radial basis neurons, express Time of the first The width of the radial basis neurons in the group;

[0023] The normalization layer has P neurons, the same number as the number of fuzzy rules in the fuzzy layer, and is used to normalize the output of the fuzzy layer to obtain a normalized output. ;

[0024] The output layer is a linear layer, and its mathematical expression is as follows:

[0025]

[0026] in, This represents the output of the self-organizing fuzzy neural network at time t, used to output the control variables of the industrial system at time t. This represents the weight matrix of the output layer at time t. This represents the number of neurons in the output layer.

[0027] Furthermore, the method of using the dataset of the first working condition to learn the explicit control law of the current working condition specifically transforms the optimization problem of the model predictive control strategy into a loss function to learn the explicit control law.

[0028] First, the optimization problem of the model predictive control strategy has an objective function that only considers the tracking performance of the industrial system's output state, and the constraints only consider the upper and lower limits of the control variables, as shown below:

[0029] (6)

[0030] in This represents a neural network prediction model. This represents the predicted output state of the prediction model. Indicates the reference trajectory. Indicates the predictive control variable. and Indicates the upper and lower limits of the control variable. Indicates the time domain range of the prediction. Indicates the control time domain range, , This indicates the order of the prediction model with respect to the input control variable u and the output state y;

[0031] Then, the objective function described above is transformed into a loss function as follows:

[0032] (7)

[0033] in, This indicates the batch size for batch training of a self-organizing fuzzy neural network. This represents the sample index for each batch. and The weighting factors represent the output state and control variables. , , and The upper and lower limit penalty terms for the control variable are shown in the following form:

[0034] (8)

[0035] Finally, the explicit control law for the first operating condition is learned through the model prediction control loss function shown in formula (7), that is, the parameter learning of the self-organizing fuzzy neural network is completed through the gradient of the loss function shown in the following backpropagation formula, as shown below:

[0036]

[0037] in, This represents the weights of the output layer of a self-organizing fuzzy neural network. This represents the width of the neurons in the fuzzy layer of a self-organizing fuzzy neural network. This represents the center of a fuzzy layer neuron in a self-organizing fuzzy neural network. This represents the learning rate.

[0038] Furthermore, the data coverage is calculated as follows: First, a truncated radial basis function is used as the activation function for the fuzzy layer neurons of the self-organizing fuzzy neural network. Time of the first The output of the fuzzy rule Represented as:

[0039] (12)

[0040] in, This indicates the truncated radial basis functions. express Time of the first In the fuzzy rule, the th The output of each neuron express Time of the first In the fuzzy rule, the th The activation state of each neuron , Indicates the first The activation state of a fuzzy rule. Let the upper and lower cutoff limits of the radial basis functions be defined as follows:

[0041] (13)

[0042] in, It is a positive integer;

[0043] Then, based on the activation state of all fuzzy rules... Determine the first element in a batch of data. A status signal indicating whether a data point falls within the range of fuzzy rules. And calculate the current data coverage of the self-organizing fuzzy neural network based on the state signals of all data. , is represented as:

[0044] (14)

[0045] (15).

[0046] Furthermore, the preset conditions are:

[0047] (16)

[0048] in, This represents the threshold.

[0049] Furthermore, the addition of fuzzy rules, i.e., the addition of neurons representing fuzzy rules in the self-organizing fuzzy neural network, is specifically designed as follows:

[0050] (17)

[0051] in, This represents the amount of data in the current learning scenario that is not covered by the self-organizing fuzzy neural network. This represents data not covered by the self-organizing fuzzy neural network. and express The newly added fuzzy rule corresponds to the truncated center and width of the radial basis neurons. This indicates the weights of the output layer corresponding to the new fuzzy rule. and Random initialization can be performed using a self-organizing fuzzy neural network.

[0052] Furthermore, the parameters of the self-organizing fuzzy neural network are updated and learned after adding fuzzy rules. The specific loss function used is as follows:

[0053] (18)

[0054] in, Indicates the current learning status. The model predicts the control loss function; This represents the current network parameters of the self-organizing fuzzy neural network. Indicates the first Network parameters, Indicates that the work has been completed. The first explicit control law obtained after learning Network parameters, Indicates the set of completed working conditions Model predictive control loss function right The second-order partial derivative, Indicates the set of completed working conditions The degree of importance; , , Let represent the set of three network parameters; for the j-th fuzzy rule, Indicates the first Network parameters, Indicates that the work has been completed. The first explicit control law obtained after learning Network parameters, Indicates the set of completed working conditions Model predictive control loss function right The second-order partial derivative;

[0055] Then, the parameters of the self-organizing fuzzy neural network are updated using the gradient of the loss function shown in the following backpropagation formula, as follows:

[0056] (10).

[0057] Furthermore, when updating the parameters of a self-organizing fuzzy neural network, the width of the radial basis neurons is additionally restricted. The range is as follows:

[0058] (19)

[0059] in, Indicates the first The fuzzy rule of the first The width of a neuron, and These represent the upper and lower limits, respectively.

[0060] Furthermore, if the data coverage of the currently learned working condition does not meet the preset conditions, the current network parameters are directly updated and learned, and the loss function used for the update and learning is:

[0061] (9)

[0062] in, Indicates the current learning status. The model predicts the control loss function; This represents the current network parameters of the self-organizing fuzzy neural network. express The first in Network parameters, This represents the set of all operating conditions that have completed explicit control learning. Indicates the set of completed working conditions The first explicit control law obtained after learning Network parameters, Indicates the set of completed working conditions Model predictive control loss function right The second-order partial derivative, Indicates the set of completed working conditions The degree of importance;

[0063] Then, the parameters of the self-organizing fuzzy neural network are updated using the gradient of the loss function shown in the following backpropagation formula, as follows:

[0064] (10).

[0065] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the method described in any of the preceding claims.

[0066] Beneficial effects

[0067] The multimodal industrial process full-condition high real-time predictive control method and equipment proposed in this invention adopts an explicit model predictive control strategy. The learning framework of the explicit control law consists of two modules: a fuzzy neural network based explicit control law (FNNECL) and a neural network based predictive model (NNPM). By adjusting the structure and parameters of FNNECL, accurate control across all operating conditions is ensured. It can not only adapt to all operating conditions with a single network, but also eliminate the impact of control strategy mismatch during switching phases. Furthermore, since only the forward propagation calculation of the self-organizing fuzzy neural network is required, without the need for optimization, computation time and resources are greatly saved, and the real-time performance of control is significantly improved. Attached Figure Description

[0068] Figure 1 This application's embodiment is a full-condition self-organizing fuzzy neural network explicit control law learning framework;

[0069] Figure 2 These are the network parameters in the embodiments of this application. and For operating condition A, the model predictive control loss function The impact results;

[0070] Figure 3 The loss function in the embodiments of this application Results of parameter changes when dealing with large-scale changes in operating conditions;

[0071] Figure 4 This is a comparison of the four control effects described in the embodiments of this application: ETASI4PC, PWA, MMPC, and FNNECL.

[0072] Figure 5 It is the hardware-in-the-loop simulation platform described in the embodiments of this application;

[0073] Figure 6 This is a comparison diagram of the control effects described in the embodiments of this application, wherein the sub-diagram shows the control effect during the switching of operating conditions. Detailed Implementation

[0074] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.

[0075] This invention provides a multimodal industrial process full-condition high real-time predictive control method and device. It adopts an explicit model predictive control strategy, and the learning framework of the explicit control law consists of two modules: a fuzzy neural network based explicit control law (FNNECL) and a neural network based predictive model (NNPM). By adjusting the structure and parameters of FNNECL, accurate control under all operating conditions is ensured.

[0076] like Figure 1 As shown, specifically, to ensure that FNNECL can adapt to different operating conditions, a truncated radial basis activation function is first introduced, and a data coverage-based operating condition change metric is proposed to accurately identify the degree of mismatch between the current and historical operating conditions. Then, for small-scale operating condition changes, considering the inherent robustness of neural network models, an Elastic Weight Consolidation (EWC) mechanism is introduced to ensure that FNNECL can learn the control strategy for new operating conditions while maintaining control performance for historical operating conditions. Finally, for large-scale operating condition changes, considering the limited expressive power of fixed-structure FNNECL for new operating conditions, a radial basis neuron growth mechanism based on data coverage is proposed. This mechanism selectively adds new neurons by calculating the data coverage of fuzzy rules to learn the control strategy for new operating conditions, enabling the explicit control law to adapt to large-scale operating condition changes. It is worth noting that the method of this invention does not require prior knowledge of the current operating conditions. It quickly obtains the accurate control sequence based on a single FNNECL model, overcoming the traditional problem of online optimization using multiple models, thus greatly expanding the scope of application of this invention and achieving precise control effects across all operating conditions.

[0077] The multimodal industrial process full-condition high real-time predictive control method provided by this invention mainly includes two parts: offline learning and online control.

[0078] 1. Offline learning: Learn the corresponding output state prediction model using the dataset for each working condition of the industrial system, and learn the explicit control law based on the self-organizing fuzzy neural network for all working conditions using the dataset for all working conditions.

[0079] In the process of learning the explicit control law for all operating conditions, the first step is to use the dataset of the first operating condition to learn the explicit control law for the current operating condition, which means learning the parameters of the self-organizing fuzzy neural network. Then, the datasets of the other operating conditions are used to adjust the currently learned explicit control law in turn: if the data coverage of the currently learned operating condition does not meet the preset conditions, the current network parameters are directly updated and learned; if the data coverage of the currently learned operating condition meets the preset conditions, fuzzy rules are added first, that is, neurons representing fuzzy rules are added in the self-organizing fuzzy neural network, and then the parameters of the self-organizing fuzzy neural network are updated and learned.

[0080] To illustrate this more vividly, we will use the working conditions as an example. , , For example, working conditions and working conditions Their operating characteristics are similar, and their working conditions are similar. The conditions differ significantly from the other two conditions. The following will be presented in three parts: (1) constructing a self-organizing fuzzy neural network, (2) training the first condition A, and (3) adapting to the condition through network parameter self-adjustment. and working conditions (4) Adapt to changes in working conditions through self-adjustment of network structure. , and working conditions Large-scale changes in operating conditions.

[0081] 1. Constructing a self-organizing fuzzy neural network

[0082] First, the basic structure of a self-organizing fuzzy neural network (SNN) is introduced. A fuzzy neural network is a combination of a fuzzy system and a neural network, where the parameters of the fuzzy system can be adjusted using a neural network-based learning algorithm. The main advantage of fuzzy neural networks is that they use a language model rather than a complex mathematical model to model the problem. The language model is essentially a fuzzy rule base, consisting of a set of IF-THEN fuzzy rules. These rules are very intuitive and easy for human users to understand, thus solving the black-box nature of the neural network paradigm. The number of fuzzy rules in a fuzzy neural network is a crucial parameter affecting its performance. A self-organizing fuzzy neural network can automatically determine the number of fuzzy rules by learning from training data. Especially when dealing with multi-condition tasks, the number of fuzzy rules changes with the number of conditions. Therefore, this invention uses a self-organizing fuzzy neural network to describe the explicit control law for all conditions. The self-organizing fuzzy neural network has four layers: an input layer, a fuzzy layer, a normalization layer, and an output layer.

[0083] Input layer: This layer has a total of Each neuron represents a fuzzy neural network. Dimensional input variables. The mathematical expression for the input layer is shown below:

[0084] (1)

[0085] in This represents the output of the i-th neuron at time t. This represents the input of the i-th neuron at time t. This represents the input of the fuzzy neural network at time t. This represents the output of the input layer at time t.

[0086] Fuzzy Layer: This layer contains P groups of neurons, each group representing a fuzzy rule. The membership function used to fuzzify the input variables is the radial basis function, and its mathematical expression is shown below:

[0087] (2)

[0088] in, Let represent the membership function of the i-th input corresponding to the j-th fuzzy rule. and Here, represents the center and width of the radial basis function neurons. After obtaining the membership degree of each input variable, the membership degree of each fuzzy rule is calculated, and its mathematical expression is as follows:

[0089] (3)

[0090] in, express Time of the first The output of the radial basis neurons, express Time of the first The center of the radial basis neurons, express Time of the first The width of the radial basis neurons in the group.

[0091] Normalization Layer: This layer has P neurons, the same number as the number of fuzzy rules in the fuzzy layer, and is used for defuzzifying rule membership degrees. Specifically, this layer normalizes the output of the fuzzy layer, as shown in the following mathematical expression:

[0092] (4)

[0093] in, This represents the output after normalizing the membership degree of the l-th fuzzy rule at time t. .

[0094] Output layer: This is a linear layer, and the output mathematical expression is as follows:

[0095] (5)

[0096] in, This represents the output of the fuzzy neural network at time t. Let represent the weight matrix of the output layer at time t.

[0097] 2. Training for the first working condition A

[0098] For ease of analysis, consider a typical model predictive control optimization problem, where the objective function only considers the tracking performance of the system output, and the constraints only consider the upper and lower limits of the manipulated variables, as shown below:

[0099] (6)

[0100] in, Indicates the predicted output. Indicates the reference trajectory. This represents the predicted manipulated variable, i.e., the output of the self-organizing fuzzy neural network. and Indicates the upper and lower limits of the manipulated variable. This indicates a neural network prediction model, which is a trained model whose parameters are frozen and not updated during the training of the fuzzy neural network. Indicates the time domain range of the prediction. Indicates the control time domain range, , This indicates the order of the prediction model with respect to the input control variable u and the output state y.

[0101] The predictive control problem shown in Equation (6) is transformed into a loss function, as shown below:

[0102] (7)

[0103] in, This indicates the batch size for batch training of a self-organizing fuzzy neural network. This represents the sample index for each batch. and The weighting factors represent the output state and control variables. , , and The upper and lower limit penalty terms for the control variable are shown in the following form:

[0104] (8)

[0105] The model predicts the control loss function based on the operating conditions as shown in formula (7). The explicit control law, that is, the parameter learning of the self-organizing fuzzy neural network is completed through the gradient of the loss function shown in the following backpropagation formula, as follows:

[0106]

[0107] in, This represents the weights of the output layer of a self-organizing fuzzy neural network. This represents the width of the neurons in the fuzzy layer of a self-organizing fuzzy neural network. This represents the center of a fuzzy layer neuron in a self-organizing fuzzy neural network. This represents the learning rate.

[0108] 3. Network parameter self-adjustment for small-scale operating condition changes

[0109] When the training dataset is switched to working condition B, the prediction model is switched to working condition B. The prediction model. At this point, if the model prediction control loss function is still used to learn the operating conditions... The explicit control law can cause a catastrophic forgetting problem, that is, the network parameters of the self-organizing fuzzy neural network are adjusted to make the operating conditions... The control performance reaches its optimal state while ignoring the operating conditions. This caused working conditions The control performance deteriorates. During online control, explicit control laws still cannot adapt to multiple operating conditions. Therefore, EWC loss is added to the model predictive control loss function, mathematically described as follows:

[0110] (9)

[0111] in, Indicates the current learning status. The model predicts the control loss function; This represents the current network parameters of the self-organizing fuzzy neural network. express The first in Network parameters, This represents the set of all operating conditions that have completed explicit control learning. Indicates the set of completed working conditions The first explicit control law obtained after learning Network parameters, Indicates the set of completed working conditions Model predictive control loss function right The second-order partial derivative, Indicates the set of completed working conditions The importance of.

[0112] When the explicit control law completes the operating condition After learning the control strategy, different network parameters affect the operating conditions. The model predicts the control loss function with different convergence patterns. For example... Figure 2 As shown, operating conditions Model predictive control loss function For parameters The second-order partial derivative is relatively large, while for the parameter The second-order partial derivative is relatively small. When the explicit control law begins operation... When learning the control strategy, if the parameters are adjusted... To improve working conditions Control performance, operating conditions Model predictive control loss function This will inevitably lead to a corresponding increase, which means that the operating conditions will be affected. The control performance decreases. However, if the parameters are adjusted... Operating conditions Model predictive control loss function It will basically not be affected, which means the operating conditions The control performance remains unchanged. Therefore, the EWC loss in the learning condition... During the control strategy process, parameters Changes Apply a larger penalty weight to the parameter. Changes Apply a smaller penalty weight. This ensures that the operating conditions are met. Improving operating conditions while maintaining control performance This improves control performance, enabling explicit control laws to adapt to small-scale changes in operating conditions.

[0113] Finally, Figure 1 The full-condition fuzzy neural network explicit control law learning framework shown below completes the parameter update of the self-organizing fuzzy neural network through the gradient of the loss function shown in formula (10) via backpropagation, as follows:

[0114] (10)

[0115] in, This represents the weights of the output layer of a self-organizing fuzzy neural network. This represents the width of the neurons in the fuzzy layer of a self-organizing fuzzy neural network. This represents the center of a fuzzy layer neuron in a self-organizing fuzzy neural network. This represents the learning rate.

[0116] 4. Self-adjusting network structure under varying operating conditions

[0117] like Figure 3 As shown, the blue area represents the conditions that enable the self-organizing fuzzy neural network to meet the operating conditions. The parameter space for control requirements; the yellow area corresponds to the operating condition. The green area corresponds to the operating condition. The red trajectory represents the loss function. The network parameter update status under the influence of [the system]. In the first stage, the network parameters are adjusted to simultaneously meet the operating conditions. and The location of the control requirements can be adapted to the working conditions. and Small-scale changes in operating conditions. The second stage, due to changes in operating conditions... The control strategy differs significantly from the other two operating conditions, and the loss function... In order to learn the working conditions The control strategy can maintain the operating conditions and The control performance was not affected, causing the network parameters to eventually adjust to an intermediate position, resulting in a sharp decline in the control performance of the explicit control law across all operating conditions. Therefore, simply adjusting the network parameters would cause the self-organizing fuzzy neural network to fail to meet the control performance requirements for all three operating conditions. This section proposes a radial basis function (RBF) neuron growth mechanism based on data coverage to adjust the network structure, learning the operating conditions by adding new RBF neurons. The control strategy, and maintain the operating condition and working conditions The control performance remains unchanged.

[0118] First, a novel truncated radial basis function is employed as the activation function for neurons in the fuzzy layer of a self-organizing fuzzy neural network. Here, we first introduce the one-dimensional truncated radial basis function, as shown in the curve... Figure 2 As shown, the mathematical expression is as follows:

[0119] (11)

[0120] in, This indicates the center of the truncated radial basis functions. This indicates the width of the truncated radial basis functions. When hour, Normal output; otherwise, The output is 0. Extending the one-dimensional truncated radial basis function to a multi-dimensional form:

[0121] (12)

[0122] in, express Time of the first The output of a fuzzy rule Indicates the first The activation state of a fuzzy rule. express Time of the first The center of a fuzzy rule, express Time of the first The width of a fuzzy rule. If Established, Indicates the first If a fuzzy rule is activated, the corresponding value is output; otherwise, the rule is not activated. Indicates the first The fuzzy rule was not activated, so the output was 0. and The settings are as follows:

[0123] (13)

[0124] in, It is a positive integer.

[0125] The special structure of the multidimensional truncated radial basis activation function means that only data in the first position are activated. The range of fuzzy rules The radial basis function neuron will only output a corresponding value if the data falls within the range of the fuzzy rule; otherwise, it will output 0. Therefore, a data coverage metric can be defined based on whether the data falls within the range of the fuzzy rule, as shown in the following mathematical expression:

[0126] (14)

[0127] (15)

[0128] in, This represents the current data coverage of the self-organizing fuzzy neural network. This indicates the batch size for training the neural network. This indicates the first element in a batch of data. A status signal indicating whether a data point falls within the range of fuzzy rules.

[0129] Next, based on data coverage To determine whether it is necessary to add radial basis neurons.

[0130] (16)

[0131] in, This represents the threshold. If the condition shown in formula (16) is met, it indicates that a neuron needs to be added. The new neuron is designed as follows:

[0132] (17)

[0133] in, This represents the amount of data not covered by the self-organizing fuzzy neural network. This represents data not covered by the self-organizing fuzzy neural network. express The center of the new fuzzy rule at every moment, express The width of the new fuzzy rule at any given moment. This represents the weight of the output layer corresponding to the new fuzzy rule. and Random initialization can be performed using a self-organizing fuzzy neural network.

[0134] Learning operating conditions by adding new fuzzy rules. While developing control strategies, it is also necessary to consider the adaptability of self-organizing fuzzy neural networks to different operating conditions. and working conditions Is the control performance affected by operating conditions? The impact of this. Considering that the activation state of the EWC loss should be consistent with that of the neurons in the fuzzy layer, the activation state is added to the EWC loss. The mathematical description of the loss function is as follows:

[0135] (18)

[0136] in, , , This represents a set of three network parameters. Indicates the first The activation state of a fuzzy rule. As shown in formulas (20) and (21), due to the special structure of the multidimensional truncated radial basis activation function, if the data Not by A fuzzy rule coverage, data during parameter self-adjustment. The calculated loss function for the th The gradient of the neuron parameters in the fuzzy rule is 0. This means that the neurons in the fuzzy layer only update their parameters based on the data they cover. and Due to operating conditions Stable point, operating point range and operating conditions and The difference is significant, especially if the width of the radial basis neurons is large. Suitable, working conditions The data will not affect the coverage conditions. and The parameters of the neurons in the data are affected, thus enabling the self-organizing fuzzy neural network to adapt to different operating conditions. and The control performance remains unchanged. Therefore, the width of the radial basis neurons is further limited during parameter self-tuning. The range is as follows:

[0137] (19)

[0138] in, Indicates the first The fuzzy rule of the first The width of a neuron, and These represent the upper and lower limits, respectively.

[0139] (20)

[0140] (twenty one)

[0141] Due to data Not by A fuzzy rule coverage, Therefore, equations (20) and (21) hold true.

[0142] II. Online Control: The current output state, along with the historical output state sequence, control sequence, and reference trajectory of the future output state, are used to construct the control state data for the current moment. This data is then input into the learned full-condition explicit control law to obtain the optimal control sequence for the current moment. Finally, the first control variable of the optimal control sequence for the current moment is used to control the industrial system.

[0143] After the above self-adjustment of network parameters and network structure, the explicit control law based on the self-organizing fuzzy neural network learns the full-condition control strategy for the industrial process, as described mathematically below:

[0144] (twenty two)

[0145] in, This represents the explicit control law for all operating conditions based on a self-organizing fuzzy neural network. Indicates the current control status. , This indicates the current output state. Representing history The output state sequence at each time step. Representing history The control sequence at each time point Indicates the future The output state reference trajectory at each moment.

[0146] During online execution, it is only necessary to construct the control state of the multimodal process at the current moment. By inputting the learned full-condition explicit control law, the optimal control sequence can be obtained. Then, the first manipulation variable of the control sequence is input into the actual system.

[0147] The method proposed in this invention only requires forward propagation computation of a self-organizing fuzzy neural network, without the need for optimization, thus significantly saving computation time and resources and greatly improving control real-time performance. Furthermore, compared to traditional multi-model predictive control methods, this method eliminates the switching mechanism, ensuring rapid matching of different modal control strategies. Therefore, this method can meet the full-condition control requirements of multi-modal processes with a single network.

[0148] To verify the effectiveness of the method proposed in this invention, the following numerical simulation experiments and hardware-in-the-loop simulation platform experiments for the calcination process were designed for verification:

[0149] I. Numerical Simulation Experiment

[0150] Consider the numerical simulation system shown below as the controlled object:

[0151] (twenty three)

[0152] in, yes Manipulated variables of the time system yes Unmeasured intermediate variables of the time system yes The controlled variable of the system at any given time, This indicates that uncertainties and disturbances that change over time cause changes in the system's operating conditions. The mean is Gaussian noise with a standard deviation of 0.01. The settings are as follows:

[0153] (twenty four)

[0154] in, The system is operating under working conditions. , The system is operating under working conditions. and The system is operating under working conditions. Due to operating conditions and working conditions The stable point operating ranges differ little, and the operating conditions are similar. and working conditions The changes are minor variations in operating conditions. (Due to the operating conditions...) and working conditions The stable point operating range varies considerably, and the operating conditions are also different. and working conditions The changes are considered to be large-scale changes in operating conditions.

[0155] In the offline learning phase, a sufficiently large random sampling dataset was used as the training dataset, followed by the learning of explicit control laws for all operating conditions based on a self-organizing fuzzy neural network. To illustrate the advantages of the proposed FNNECL method, ETASI4PC, PWA, and MMPC were used as comparative methods. The parameters of the self-organizing fuzzy neural network were set as follows: initial number of fuzzy rules... The center of the radial basis neurons is in Perform uniformly distributed random initialization within the range, with the width set to... Batch size for batch training Set to 32, parameter Set to 3, threshold OL to 2, learning rate The number of iterations is set to 300. The control parameters are set as follows: prediction time domain range. Setting it to 2 controls the time domain range. Set to 2, weighting factor and Upper and lower limits of manipulated variables and The width of radial basis neurons is limited to Within the range.

[0156] During the online control phase, the system is in operating condition A during the 300th cycle, in operating condition B during the 300th to 600th cycles, and in operating condition C during the 600th to 1000th cycles.

[0157] The system output setpoint is set to a reference trajectory in the form of a sine function, and the performance metric is selected as MSE. Figure 4It is evident that the ETASI4PC and PWA methods can only adapt to small-scale operating condition changes; their control performance deteriorates sharply when operating conditions change significantly. In contrast, MMPC and FNNECL can adapt to a wide range of operating condition changes, achieving tracking control during the switching between the three operating conditions. Table 1 shows that for multimodal processes, MMPC's convergence is slower due to the need for an error-triggered mechanism to update the prediction model during large-scale operating condition changes, resulting in slightly inferior overall control performance compared to FNNECL. Meanwhile, FNNECL, during online control, only requires inputting the current state into the fuzzy neural network to obtain the optimal control sequence, eliminating the need for online optimization problems or row-by-row table lookups, thus requiring the shortest runtime. Comparative experiments demonstrate that for multimodal processes, FNNECL has a faster operating speed and, with a single network, can adapt to both large and small-scale operating condition changes simultaneously.

[0158]

[0159] II. Experiment using a semi-physical simulation platform for the roasting process

[0160] To verify the performance of the proposed full-condition explicit model predictive control based on self-organizing fuzzy neural network in industrial processes with fluctuating operating conditions, a system was built as follows: Figure 5 The hardware-in-the-loop simulation platform shown is capable of effectively simulating industrial environments and validating engineering deployment schemes. It has been used for verifying control algorithms such as multi-model predictive control and explicit model predictive control. The platform uses a roasting furnace as the simulation object; roasting is the first step in zinc smelting, converting zinc concentrate into zinc roasted sand. The entire roasting furnace system consists of the furnace body, feeding system, blasting system, exhaust system, and zinc tapping system. In industry, the feed belt speed is typically used as the manipulated variable in the roasting process, while the standard temperature is used as the controlled variable. Considering communication bandwidth, the sampling period is set to 10 seconds, and the control period to 30 seconds. To fully excite the roasting process, the feed belt speed is set to... Random signals within a certain range are input into the HILS platform to generate process operation data.

[0161]

[0162] This study considers the variations in operating conditions during the calcination process. The system parameters for the three operating conditions are shown in Table 2. Conditions A and B represent small-range variations, while conditions A and B, along with condition C, represent large-range variations. In the simulation experiment, the system operates under condition A for the first 400 cycles, under condition B from cycle 400 to 800, and under condition C from cycle 800 to 1400. The setpoint for the calcination process temperature is set to the following step change:

[0163] (32)

[0164] Considering that the change in rotation speed is generally used as the manipulated variable in the roasting process control, the original dataset is differentially processed and normalized to serve as the training dataset. The parameter settings for the self-organizing fuzzy neural network are as follows: initial number of fuzzy rules. The center of the radial basis neurons is in Perform uniformly distributed random initialization within the range, with the width set to... Batch size for batch training Set to 32, parameter Set to 3, Set to 1.5, threshold Set the learning rate to 4. The iteration step count is set to 100. The control parameters are set as follows: prediction time domain range. Setting it to 4 controls the time domain range. Set to 2, weighting factor and Upper and lower limits of manipulated variables and The width of radial basis neurons is limited to Within the specified range. The number of fuzzy rules obtained through training is 326. The performance metric chosen is MSE. The comparison method selected is error-triggered multi-model predictive control, while matrix transformation is used to simplify the optimization problem to improve the solution speed. The control effects of the two methods are as follows: Figure 6 As shown in Table 3, the quantitative results are as follows.

[0165]

[0166] Experimental results show that FNNECL can stably track the setpoint in the furnace during the fluctuating roasting process. The temperature abrupt change in the 800th cycle was caused by a large-scale change in operating conditions. FNNECL has a shorter settling time, smaller overshoot and fluctuation, especially when operating conditions change, because FNNECL does not need to switch prediction models at this time; the truncated radial basis neurons in the fuzzy layer select appropriate fuzzy rules based on the data to obtain the control sequence. Experimental verification shows that the FNNECL proposed in this invention only needs 0.006s to solve for the optimal control quantity on average, while MMPC requires 0.513s on average, which is 85.5 times faster. For edge devices that need to handle multiple tasks, this greatly reduces the computing power requirement and significantly improves the real-time performance of control decisions. In summary, compared with traditional MMPC, FNNECL can not only adapt to all operating conditions with a single network and eliminate the impact of control strategy mismatch during the switching phase, but also requires only a small amount of computing resources to meet real-time requirements. Therefore, FNNECL has more significant advantages.

[0167] The explicit predictive control method based on self-organizing fuzzy neural networks proposed in this invention can be applied to the precise control of multimodal industrial processes. It not only adapts to all operating conditions with a single network but also eliminates the impact of control strategy mismatch during switching phases, while reducing computational requirements and significantly improving the real-time performance of control decisions. Specifically, compared to the traditional MMPC algorithm, which requires an average of 0.513 seconds per solution, the proposed FNNECL algorithm only requires an average of 0.006 seconds per solution to find the optimal control value, making the former 85.5 times faster. Therefore, this invention is suitable for scenarios with frequent changes in operating states, high sampling frequencies, high real-time requirements, and limited computational resources.

[0168] The above embodiments are preferred embodiments of this application. Those skilled in the art can make various changes or improvements based on them. Without departing from the overall concept of this application, these changes or improvements should fall within the scope of protection claimed in this application.

Claims

1. A high real-time predictive control method for multimodal industrial processes under all operating conditions, characterized in that, include: Offline learning: We learn the corresponding output state prediction model using the dataset for each operating condition of the industrial system, and learn the full-condition explicit control law based on a self-organizing fuzzy neural network using the dataset for all operating conditions. In the process of learning the explicit control law for all operating conditions, the first step is to use the dataset of the first operating condition to learn the explicit control law for the current operating condition, which means learning the parameters of the self-organizing fuzzy neural network. Then, the datasets of the other operating conditions are used to adjust the currently learned explicit control law in turn: if the data coverage of the currently learned operating condition does not meet the preset conditions, the current network parameters are directly updated and learned; if the data coverage of the currently learned operating condition meets the preset conditions, fuzzy rules are added first, that is, neurons representing fuzzy rules are added in the self-organizing fuzzy neural network, and then the parameters of the self-organizing fuzzy neural network are updated and learned. Online control: The current output state, together with the historical output state sequence, control sequence, and reference trajectory of future output state, are used to construct the control state data for the current moment. This data is then input into the learned full-condition explicit control law to obtain the optimal control sequence for the current moment. The industrial system is controlled using the first control variable of the optimal control sequence at the current moment.

2. The multimodal industrial process full-condition high real-time predictive control method according to claim 1, characterized in that, The self-organizing fuzzy neural network includes an input layer, a fuzzy layer, a normalization layer, and an output layer; The input layer includes Each neuron represents a fuzzy neural network. Dimensional input variables; the mathematical expression for the input layer is shown below: （1） in, This represents the input of the i-th neuron at time t. This represents the output of the i-th neuron at time t; This represents the input to the fuzzy neural network at time t, used to input the control state data of the industrial system at time t; This represents the output of the input layer at time t; The fuzzy layer consists of P groups of neurons, each group representing a fuzzy rule. Each fuzzy rule uses a radial basis function as the membership function of the fuzzification input variable, as shown in the following mathematical expression: （2） in, Let represent the membership function of the i-th input variable corresponding to the j-th fuzzy rule. and Here, we define the center and width of the radial basis function neurons. After obtaining the membership degree of each input variable, we calculate the membership degree of each fuzzy rule, with the following mathematical expression: （3） in, express Time of the first The output of the radial basis neurons, express Time of the first The center of the radial basis neurons, express Time of the first The width of the radial basis neurons in the group; The normalization layer has P neurons, the same number as the number of fuzzy rules in the fuzzy layer, and is used to normalize the output of the fuzzy layer to obtain a normalized output. ; The output layer is a linear layer, and its mathematical expression is as follows: ； in, This represents the output of the self-organizing fuzzy neural network at time t, used to output the control variables of the industrial system at time t. This represents the weight matrix of the output layer at time t. This represents the number of neurons in the output layer.

3. The multimodal industrial process full-condition high real-time predictive control method according to claim 2, characterized in that, The method of using the dataset of the first working condition to learn the explicit control law of the current working condition specifically transforms the optimization problem of the model predictive control strategy into a loss function to learn the explicit control law. First, the optimization problem of the model predictive control strategy has an objective function that only considers the tracking performance of the industrial system's output state, and the constraints only consider the upper and lower limits of the control variables, as shown below: （6） in This represents a neural network prediction model. This represents the predicted output state of the prediction model. Indicates the reference trajectory. Indicates the predictive control variable. and Indicates the upper and lower limits of the control variable. Indicates the time domain range of the prediction. Indicates the control time domain range, , This indicates the order of the prediction model with respect to the input control variable u and the output state y; Then, the objective function described above is transformed into a loss function as follows: （7） in, This indicates the batch size for batch training of a self-organizing fuzzy neural network. This represents the sample index for each batch. and The weighting factors represent the output state and control variables. , , and The upper and lower limit penalty terms for the control variable are shown in the following form: （8） Finally, the explicit control law for the first operating condition is learned through the model prediction control loss function shown in formula (7), that is, the parameter learning of the self-organizing fuzzy neural network is completed through the gradient of the loss function shown in the following backpropagation formula, as shown below: ； in, This represents the weights of the output layer of a self-organizing fuzzy neural network. This represents the width of the neurons in the fuzzy layer of a self-organizing fuzzy neural network. This represents the center of a fuzzy layer neuron in a self-organizing fuzzy neural network. This represents the learning rate.

4. The multimodal industrial process full-condition high real-time predictive control method according to claim 3, characterized in that, The data coverage is calculated as follows: First, a truncated radial basis function is used as the activation function for the neurons in the fuzzy layer of the self-organizing fuzzy neural network. Time of the first The output of the fuzzy rule Represented as: （12） in, This indicates the truncated radial basis functions. express Time of the first In the fuzzy rule, the th The output of each neuron express Time of the first In the fuzzy rule, the th The activation state of each neuron , Indicates the first The activation state of a fuzzy rule. Let the upper and lower cutoff limits of the radial basis functions be defined as follows: （13） in, It is a positive integer; Then, based on the activation state of all fuzzy rules... Determine the first element in a batch of data. A status signal indicating whether a data point falls within the range of fuzzy rules. And calculate the current data coverage of the self-organizing fuzzy neural network based on the state signals of all data. , represented as: （14） （15）。 5. The multimodal industrial process full-condition high real-time predictive control method according to claim 4, characterized in that, The preset conditions are: （16） in, This represents the threshold.

6. The multimodal industrial process full-condition high real-time predictive control method according to claim 4, characterized in that, The addition of fuzzy rules refers to adding neurons representing fuzzy rules in the self-organizing fuzzy neural network. Specifically, the newly added neuron group is designed as follows: （17） in, This represents the amount of data in the current learning scenario that is not covered by the self-organizing fuzzy neural network. This represents data not covered by the self-organizing fuzzy neural network. and express The newly added fuzzy rule corresponds to the truncated center and width of the radial basis neurons. This indicates the weights of the output layer corresponding to the new fuzzy rule. and Random initialization can be performed using a self-organizing fuzzy neural network.

7. The multimodal industrial process full-condition high real-time predictive control method according to claim 6, characterized in that, After adding fuzzy rules, the parameters of the self-organizing fuzzy neural network are updated and learned. The specific loss function used is as follows: （18） in, Indicates the current learning status. The model predicts the control loss function; This represents the current network parameters of the self-organizing fuzzy neural network. Indicates the first Network parameters, Indicates that the work has been completed. The first explicit control law obtained after learning Network parameters, Indicates the set of completed working conditions Model prediction control loss function right The second-order partial derivative, Indicates the set of completed working conditions The degree of importance; , , Let represent the set of three network parameters; for the j-th fuzzy rule, Indicates the first Network parameters, Indicates that the work has been completed. The first explicit control law obtained after learning Network parameters, Indicates the set of completed working conditions Model prediction control loss function right The second-order partial derivative; This indicates the activation state of the j-th fuzzy rule; Then, the parameters of the self-organizing fuzzy neural network are updated using the gradient of the loss function shown in the following backpropagation formula, as follows: （10）。 8. The multimodal industrial process full-condition high real-time predictive control method according to claim 7, characterized in that, When updating the parameters of a self-organizing fuzzy neural network, the width of the radial basis neurons is additionally restricted. The range is as follows: （19） in, Indicates the first The fuzzy rule of the first The width of a neuron, and These represent the upper and lower limits, respectively.

9. The multimodal industrial process full-condition high real-time predictive control method according to claim 3, characterized in that, If the data coverage of the currently learned working condition does not meet the preset conditions, the current network parameters are directly updated and learned. The loss function used for the update and learning is: （9） in, Indicates the current learning status. The model predicts the control loss function; This represents the current network parameters of the self-organizing fuzzy neural network. express The first in Network parameters, This represents the set of all operating conditions that have completed explicit control learning. Indicates the set of completed working conditions The first explicit control law obtained after learning Network parameters, Indicates the set of completed working conditions Model predictive control loss function right The second-order partial derivative, Indicates the set of completed working conditions The degree of importance; Then, the parameters of the self-organizing fuzzy neural network are updated using the gradient of the loss function shown in the following backpropagation formula, as follows: （10）。 10. An electronic device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, the processor causes the processor to implement the method as described in any one of claims 1 to 9.

Images