Model predictive control performance diagnosis method based on spatio-temporal feature extraction

Abstract

The invention discloses a model predictive control performance diagnosis method based on spatio-temporal feature extraction. The method comprises the steps of obtaining a pre-constructed data set and performing preprocessing; inputting the preprocessed data set into a pre-trained model prediction control system spatial-temporal feature extraction model to obtain a classification result output by the model prediction control system spatial-temporal feature extraction model; according to a classification result of the spatial-temporal feature extraction model of the model predictive control system, outputting a category label causing performance degradation of model predictive control, and realizing performance diagnosis of the model predictive control system; wherein the spatial-temporal feature extraction model of the model predictive control system comprises a structural feature extraction model based on a graph neural network, a time sequence feature extraction model based on a long-short memory network and a full connection layer which are connected in series. According to the method, the spatial-temporal characteristics of the model prediction control system are jointly extracted, real-time performance diagnosis of the control system is achieved, model mismatch and other degradation factors are effectively distinguished, and then model performance diagnosis is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial process manufacturing, and particularly to a model predictive control performance diagnosis method based on spatio-temporal feature extraction. Background Art

[0002] In recent years, advanced process control (APC) has become a research hotspot in the control field. Among them, model predictive control (MPC) has received extensive attention due to its advantages in dealing with multi-variable, non-linear, and multi-constraint problems. MPC has been successfully applied to industries such as petroleum, chemical, steel, metallurgy, and power, and is one of the most influential control technologies in the future. The "14th Five-Year Plan" emphasizes promoting the high-end, intelligent, and green development of the manufacturing industry. MPC has significant advantages in this strategy and can integrate economic, safety, and environmental goals into the optimal control framework of the manufacturing process to achieve the integrated management of product quality and production safety.

[0003] However, during the entire life cycle of the model predictive control (MPC) system, multiple disturbance factors such as the drift of industrial equipment operation parameters, the time-varying characteristics of key process parameters, and the change of production load will cause the model-object mismatch phenomenon, resulting in the attenuation of the prediction accuracy of the control system and the deterioration of the regulation quality, and further causing an increase in energy consumption, product quality deviation, and potential safety hazards in equipment operation. Among them, model mismatch is the most common problem and the core of monitoring and diagnosis. Traditional model performance diagnosis methods, such as the minimum variance criterion (MVC) and the linear quadratic Gaussian criterion (LQG), are widely used in practice. However, for multi-variable constrained systems with complex structures, their diagnosis has technical defects such as large diagnosis lag and failure to consider the structural characteristics of the control system. Summary of the Invention

[0004] The purpose of the present invention is to provide a model predictive control performance diagnosis method based on spatio-temporal feature extraction. By extracting the structural features and temporal features of the model predictive control system, the spatio-temporal features of the model predictive control system can be extracted, and then classified through the spatio-temporal features to obtain the reasons for the performance degradation of the model predictive control system, so as to realize the performance diagnosis of the model predictive control system. The present invention is realized through the following technical solutions.

[0005] In the first aspect, the present invention provides a model predictive control performance diagnosis method based on spatio-temporal feature extraction, including the following:

[0006] Obtain a pre-constructed data set and perform preprocessing;

[0007] Input the preprocessed data set into a pre-trained spatio-temporal feature extraction model of the model predictive control system to obtain the classification result output by the spatio-temporal feature extraction model of the model predictive control system;

[0008] According to the classification results of the spatio-temporal feature extraction model of the model predictive control system, the categories causing performance degradation of the model predictive control are output to complete the performance diagnosis of the model predictive control;

[0009] Among them, the spatio-temporal feature extraction model of the model predictive control system includes a structure feature extraction model based on a graph neural network, a temporal feature extraction model based on a long short-term memory network, and a fully connected layer connected in series.

[0010] In practical applications, due to the existing model predictive control performance diagnosis methods for multi-variable constraint systems with complex structures, there are technical defects such as large diagnostic hysteresis and lack of consideration of the control system structure characteristics. The present invention constructs an online diagnosis method for model predictive control performance based on spatio-temporal feature extraction, and effectively distinguishes model mismatch from other degradation factors by jointly extracting the structure features and temporal features of the model predictive control system in real time, so as to realize the performance diagnosis of the model predictive control running on the model predictive control system.

[0011] Optionally, the structure feature extraction model based on a graph neural network is used to extract the structure features of the model predictive control system and input the extracted structure features into the temporal feature extraction model based on a long short-term memory network;

[0012] The temporal feature extraction model based on a long short-term memory network is used to extract the temporal features of the model predictive control system, combine the temporal features and the structure features to obtain the spatio-temporal features of the model predictive control system, and input the spatio-temporal features into the fully connected layer;

[0013] The fully connected layer is used to classify according to the spatio-temporal features and output the categories causing performance degradation of the model predictive control.

[0014] The spatio-temporal feature extraction method of the present invention solves the problems of diagnostic hysteresis and lack of structure features by extracting the structure and temporal features of the model predictive control system in real time. By using a graph neural network and a long short-term memory network to form comprehensive spatio-temporal features, the performance degradation factors are accurately identified, and the diagnostic ability and real-time performance of complex systems are significantly improved, providing support for improving the reliability and efficiency of the model predictive control system.

[0015] Optionally, the pre-constructed data set includes a data set constructed according to the historical operation data of the model predictive control under different performance modes;

[0016] Each data in the historical operation data includes the input variable value, output variable value of the model predictive control system at the same moment, and the performance mode in which the model predictive control currently running on the model predictive control system is located;

[0017] The different performance modes of the model predictive control include good model, model gain mismatch, model time constant mismatch, model time delay mismatch, and disturbance characteristic change.

[0018] The dataset constructed in the present invention collects historical operation data of the model predictive control under different performance modes, including input and output variable values and performance modes, such as good model, gain mismatch, etc. This dataset provides rich information for spatio-temporal feature extraction and classification, significantly improves the recognition ability of different performance modes, and improves the accuracy and reliability of the performance diagnosis of the model predictive control system.

[0019] Optionally, the graph neural network-based structural feature extraction model includes a graph structure input layer, a multi-order graph convolution layer, and a graph feature aggregation layer connected in series;

[0020] The graph structure input layer is used to receive the input variable value and the output variable value, map the input variable value and the output variable value to the initial feature vector of the graph node, and construct an adjacency matrix according to the connection relationship of the variables of the model predictive control system;

[0021] The multi-order graph convolution layer includes two graph convolution modules connected in series. The first graph convolution module is used to receive the initial feature vector of the graph node and the adjacency matrix and perform graph convolution operations to obtain the updated graph node and adjacency matrix; the second graph convolution module is used to receive the updated graph node and adjacency matrix and perform graph convolution operations again to obtain the structural features extracted at the current moment;

[0022] The graph feature aggregation layer is used to receive the structural features extracted at each moment by the multi-order graph convolution layer and splice them into a structural feature sequence in chronological order as the input of the long short-term memory network-based temporal feature extraction model.

[0023] Optionally, the graph neural network-based structural feature extraction model extracts features through the following formula:

[0024] ,

[0025] ,

[0026] In the formula, is the message passing function, is the embedding representation of the -th node, is the embedding representation of the -th node, is the embedding representation of the edge between the node and the node. The message passing function and to obtain the node embedding representation In layer, the aggregated messages , and the update function will combine the aggregated messages and the features of the node itself to update and obtain the features of the layer , which is the structural feature of the model predictive control system extracted by the

[0027] Optionally, the correlation relationships between the variables as edges in the graph structure include:

[0028] If there is a correlation between an input variable and another input variable in the model predictive control system, there is an edge between them, otherwise there is no edge; if there is a correlation between an input variable and an output variable in the model predictive control system, there is an edge between them, otherwise there is no edge; if there is a correlation between an output variable and another output variable in the model predictive control system, there is an edge between them, otherwise there is no edge.

[0029] Optionally, constructing the time series feature extraction model based on the long short-term memory network includes the following steps:

[0030] Input the data with structural features extracted by the structural feature extraction model based on the graph neural network into the long short-term memory network, and the sequential data of each variable within the time window is used to extract time series features through an independent long short-term memory network. The specific steps are as follows:

[0031] Step 1, input construction: For each variable , the historical sequential data within the time window length is , and the expression is as follows:

[0032] ,

[0033] where is the observation value of the variable at time , ranges from t - T + 1 to t, and t is the current time; the output of each time is used as the input of the long short-term memory network, and the input of the long short-term memory network has the following expression:

[0034] ,

[0035] In the formula, is the structural feature output by the graph neural network at time

[0036] Step 2, long short-term memory network time series modeling: For the structural feature sequence of each variable , the long short-term memory network unit processes step by step at each time. The processing process includes: inputting the structural feature at the current time and the hidden state at the previous time into the input gate and forget gate of the long short-term memory network unit for calculation to obtain the candidate cell state . According to the candidate cell state , the cell state is updated to obtain the current cell state . Input the hidden state at the previous time and the current cell state into the output gate for calculation to obtain the hidden state at the current time.

[0037] Step 3, time series feature extraction: After time step iterations, take the final hidden state of the long short-term memory network as the spatio-temporal coupling feature of the variable . Concatenate the spatio-temporal coupling features of all variables into a global time series feature vector through the following formula:

[0038] ,

[0039] This feature vector encodes both the structural dependence relationship and the dynamic time series evolution law between variables, providing a spatio-temporal coupling feature representation for model predictive control.

[0040] On the second aspect, the present invention provides a training method for model predictive control performance diagnosis based on spatio-temporal feature extraction, including:

[0041] Obtain data samples and annotation information from different performance mode categories of model predictive control;

[0042] Divide the data samples into training set samples and test set samples;

[0043] Preprocess the data samples to determine three types of annotation information: model mismatch, unmeasurable disturbance influence, and good model;

[0044] Input the preprocessed training set samples and sample annotation information into a pre-constructed spatio-temporal feature extraction model of a model predictive control system for training;

[0045] Detect the classification accuracy of the trained model prediction control system spatio-temporal feature extraction model through the test set samples. If the classification accuracy reaches the preset accuracy requirement, stop the training; otherwise, continue to train the model prediction control system spatio-temporal feature extraction model through the training set samples.

[0046] Beneficial effects

[0047] (1) The present invention describes the model prediction control system in a graph structure and uses the graph convolution algorithm to extract the structural features of the control system, which can effectively fit the fault propagation characteristics in the actual production process and improve the accuracy of the overall system performance evaluation. The present invention also constructs a time series feature extraction model based on the long short-term memory network to extract the time series features of the model prediction control system, and uses the structural features extracted by the graph convolution network as the input to further extract the spatio-temporal coupling features of the control system.

[0048] (2) The present invention uses the extracted structural features as the input of the long short-term memory network to realize the extraction of time series features. This process combines the real-time data during the operation of the control system through the joint extraction and classification of time features and space features, and uses artificial intelligence technology to realize the real-time performance diagnosis of the model prediction control system. Description of the drawings

[0049] Figure 1 The following shows a schematic structural diagram of a data gateway system based on a message queue in an embodiment of the present invention;

[0050] Figure 2 The following shows a schematic structural diagram of the Wood-Berry binary distillation column model of the present invention;

[0051] Figure 3 The following shows a diagnostic result in an embodiment of the present invention. Detailed implementation manners

[0052] The following is further described in conjunction with the drawings and specific embodiments. In the description of the present invention, it should be understood that terms such as "first", "second", etc. are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first", "second", etc. may explicitly or implicitly include one or more of such features.

[0053] Embodiment 1

[0054] This embodiment introduces a model prediction control performance diagnosis method based on spatio-temporal feature extraction, which specifically includes the following contents:

[0055] Obtain a pre-constructed data set and perform preprocessing;

[0056] Input the preprocessed dataset into the pre-trained model predictive control spatio-temporal feature extraction model to obtain the classification result output by the model predictive control system spatio-temporal feature extraction model;

[0057] According to the classification result of the model predictive control system spatio-temporal feature extraction model, output the category label that causes the performance degradation of the model predictive control, and complete the performance diagnosis of the model predictive control; the category label includes model mismatch and unmeasured disturbance influence;

[0058] Among them, the model predictive control system spatio-temporal feature extraction model includes a structure feature extraction model based on a graph neural network, a time series feature extraction model based on a long short-term memory network, and a fully connected layer connected in series.

[0059] In practical applications, due to the existing model predictive control performance diagnosis methods for multi-variable constrained systems with complex structures, there are technical defects such as large diagnosis lag and unconsideration of the control system structure features. The present invention constructs an online diagnosis method for model predictive control performance based on spatio-temporal feature extraction, which jointly extracts the structure features and time series features of the model predictive control system in real time, thereby extracting the due spatio-temporal features, and further effectively distinguishing model mismatch from other degradation factors, realizing the performance diagnosis of the model predictive control running on the model predictive control system.

[0060] Embodiment 2

[0061] On the basis of Embodiment 1, this embodiment introduces the specific implementation process of a model predictive control performance diagnosis method based on spatio-temporal feature extraction, as Figure 1 shown, which specifically includes the following contents:

[0062] I. Acquisition and processing of the dataset

[0063] The present invention constructs a dataset according to the historical operation data of the model predictive control under different performance modes;

[0064] Each data in the historical operation data includes the input variable value, output variable value of the model predictive control system at the same moment, and the performance mode in which the model predictive control currently running on the model predictive control system is located;

[0065] The different performance modes of the model predictive control performance include: good model, model gain mismatch, model time constant mismatch, model time delay mismatch, and disturbance characteristic change;

[0066] Process the data set in the form of a sliding window, define the sliding window size and step size parameters, label the data within the entire window according to the proportion of data labels within the window, and divide the data set into a training set and a validation set. In this embodiment, the sliding window size is set to 20, the step size is 2, and the data set is divided into a training set and a validation set in a ratio of 8:2.

[0067] II. Construction and Training of the Model

[0068] 2.1 Construction of the Model

[0069] 2.11 Construct a structural feature extraction model based on a graph neural network

[0070] The structural feature extraction model based on the graph neural network is used to extract the structural features of the model predictive control system and input the extracted structural features into the time series feature extraction model based on the long short-term memory network.

[0071] The structural feature extraction model based on the graph neural network includes a graph structure input layer, a multi-order graph convolutional layer, and a graph feature aggregation layer connected in series;

[0072] The graph structure input layer is used to receive the input variable value and the output variable value, map the input variable value and the output variable value to the initial feature vector of the graph node, and construct an adjacency matrix according to the connection relationship of the model predictive control system variables;

[0073] The multi-order graph convolutional layer includes two graph convolutional modules connected in series. The first graph convolutional module is used to receive the initial feature vector of the graph node and the adjacency matrix and perform graph convolutional operations to obtain the updated graph node and adjacency matrix; the second graph convolutional module is used to receive the updated graph node and adjacency matrix and perform graph convolutional operations again to obtain the structural features extracted at the current moment;

[0074] Since model predictive control is applied to the model predictive control system, the input variables and output variables of the model predictive control system are used as nodes in the graph structure, the variable values of the input variables and output variables are used as the features of the nodes, and the relationship between the variables is used as the edge in the graph structure; there are three corresponding relationships between the relationship between the variables and the existence of an edge: if one input variable is related to another input variable, there is an edge between them, otherwise there is no edge; if one input variable is related to an output variable, there is an edge between them, otherwise there is no edge; if one output variable is related to another output variable, there is an edge between them, otherwise there is no edge.

[0075] Step 1, obtain the number of input variables and output variables of the model predictive control system, the correlation relationship between the variables, and the variable values of the input variables and output variables;

[0076] Step 2: The input variables and output variables of the control system are predicted by the model to form nodes in the graph structure , and the nodes are expressed as follows:

[0077] ,

[0078] where f represents the number of layers of the graph neural network model, is an input variable or an output variable of the model predictive control system, i is the serial number of the input variable or output variable of the model predictive control system, n is the total number of input variables and output variables of the model predictive control system, and the variable values of the input variables and output variables are used as the features of the node ;

[0079] Construct a graph adjacency matrix representing the relationships between graph nodes based on the prior knowledge of the model predictive control system structure , which is used to obtain the graph data of the model predictive control system. Among them, the importance of the relationship between variable and variable is represented by the element in the th row and the th column of the graph adjacency matrix . The structure features of the model predictive control system are extracted by using a graph convolutional algorithm based on space through the following formula:

[0080] ,

[0081] ,

[0082] where is the message passing function, is the embedding representation of the th layer node, is the embedding representation of the th node, is the embedding representation of the edge between the node and the node. The message passing function aggregates and 's embedding representations to obtain the aggregated message of the node in the th layer. The update function combines the aggregated message and the features of the node itself to update and obtain the features of the th layer, is The structural features of the model predictive control system extracted from the layer.

[0083] 2.12 Construct a time series feature extraction model based on the long short-term memory network

[0084] The time series feature extraction model based on the long short-term memory network is used to extract the time series features of the model predictive control system, and combine the time series features and the structural features to obtain the spatio-temporal features of the model predictive control system and input them into the fully connected layer; the fully connected layer is used to classify according to the spatio-temporal features and output the labels causing performance degradation of the model predictive control system.

[0085] Input the data with structural features extracted by the structural feature extraction model based on the graph neural network into the long short-term memory network, and the sequence data of each variable within the time window is used to extract time series features through an independent long short-term memory network. The specific steps are as follows:

[0086] Step 1, input construction: For each variable , the historical sequence data within the time window length is , and the expression is as follows:

[0087] ,

[0088] In the formula, is the observation value of variable at time , ranges from t - T + 1 to t, where t is the current time; the output of each time is used as the input of the long short-term memory network , and the input of the long short-term memory network is expressed as:

[0089] ,

[0090] In the formula, is the structural features output by the graph neural network at time.

[0091] Step 2, long short-term memory network time series modeling: For the structural feature sequence of each variable , the long short-term memory network unit processes the input step by step at each time, and the processing process includes: inputting the structural features at the current time and the hidden state at the previous time into the input gate and forget gate of the long short-term memory network unit for calculation to obtain the candidate cell state , and according to the candidate cell state Update the cell state to obtain the current cell state , and input the previous hidden state and the current cell state into the output gate for calculation to obtain the current hidden state .

[0092] Step 3, Temporal feature extraction: After iterations of time steps, take the final hidden state of the long short-term memory network as the spatio-temporal coupling feature of the variable . Concatenate the spatio-temporal coupling features of all variables into a global temporal feature vector through the following formula :

[0093] ,

[0094] This feature vector encodes both the structural dependence relationship and the dynamic temporal evolution law between variables, providing a spatio-temporal coupling feature representation for model predictive control.

[0095] 2.2 Training of the model

[0096] Input the divided training set into the spatio-temporal feature extraction model of the model predictive control system, and optimize the parameters of the spatio-temporal feature extraction model of the model predictive control system by minimizing the cross-entropy loss function between the classification category and the true category through the gradient descent algorithm to obtain the trained weights. Load the trained model weights, and then obtain the trained spatio-temporal feature extraction model of the model predictive control system.

[0097] Input the processed real-time operation data into the trained model to identify the performance mode of model predictive control and diagnose the performance mode of model predictive control running on the model predictive control system.

[0098] III. Experiments

[0099] In this embodiment, the Wood-Berry binary distillation column model is selected for experiments. As Figure 2 shown is the structural schematic diagram of the Wood-Berry binary distillation column model. Configure the corresponding performance mode status through the MPC toolbox in MATLAB. The performance modes implemented in this example are as follows in the table:

[0100] The model predictive control performance modes include good model, model gain mismatch, model time constant mismatch, model time delay mismatch, and disturbance characteristic change. Model gain mismatch, model time constant mismatch, and model time delay mismatch belong to model mismatch. Therefore, there are a total of 3 label categories: good model, model mismatch, and unmeasured disturbance influence.

[0101] In this experiment, the model gain mismatch performance mode simulates the gain mismatch situation by changing the first-channel gain of the actual model from 12.8 to 25.6; the model time constant mismatch performance mode simulates the time constant mismatch situation by changing the first-channel time constant of the actual model from 16.7 to 20; the model time delay mismatch performance mode simulates the time delay mismatch situation by changing the first-channel time delay of the actual model from 1 to 3.

[0102] The unmeasured disturbance influence performance mode sets three different disturbance signals: sine disturbance, step disturbance and random disturbance. Among them, the sine disturbance cases with different oscillation amplitudes are experimented: amplitude of 0.5 and amplitude of 1, the random disturbance is random noise with a variance equal to 0.15, and the step disturbance is a step signal with a step amplitude of 1.

[0103] According to the structure of the model predictive control system, a structure feature extraction model based on a graph neural network is constructed. The transfer function of the control system model in this example is as follows:

[0104] In the formula, is the complex frequency variable, representing the complex frequency in the Laplace domain, which is used to describe the dynamic characteristics of the system. and respectively represent the top product concentration and the bottom liquid concentration of the Wood-Berry binary distillation column model, represent the top reflux flow rate and the bottom reboiler steam flow rate of the distillation column model respectively, and represent the model feed flow disturbance. The bottom liquid concentration refers to the concentration of the liquid substance at the bottom. This system has two inputs and two outputs, so a graph structure containing 4 nodes is constructed. According to the coupling relationship between the control variables, the edges in the graph structure can be obtained: to , to , to and to There are edges between them, and thus the adjacency matrix A can be constructed: , and the input dimension of the graph convolutional network is equal to the number of nodes 4, and the output hidden layer dimension is 16.

[0105] A time series feature extraction model based on a long short-term memory network is constructed. The input is the structure features of the model predictive control system extracted. In this example, the output hidden layer dimension of the graph convolutional network is 16, and the number of nodes in the control system is 4, so the input dimension of the long short-term memory network is 16 * 4 = 64, and the output hidden layer number of the long short-term memory network is 64. On this basis, to achieve the performance mode classification function, a fully connected layer is connected after the output hidden layer of the long short-term memory network. The input layer number of the fully connected layer is 64, and the output layer number is 3.

[0106] The hardware operating environment of this embodiment is shown in the following table:

[0107] Load the trained model predictive control performance classification model based on spatio-temporal feature extraction, and load the processed real-time operation data to obtain the performance diagnosis result of the current predictive control system, such as Figure 3 This is the confusion matrix of the diagnosis result of this example. In the figure, the vertical coordinate is the true category of the model predictive control system, and the horizontal coordinate is the category identified by the method proposed in the present invention. The intersection of the horizontal coordinate and the vertical coordinate is the recognition accuracy of the method proposed in the present invention. The recognition accuracies for model mismatch and unmeasurable disturbance effects are 70% and 99% respectively. The result of the performance diagnosis is highly consistent with the true category, which proves the accuracy of the diagnosis result.

[0108] In this embodiment, by setting different performance modes to collect the data of the model predictive control system under each performance mode, the performance diagnosis of the model predictive control system is realized by using artificial intelligence technology. It has a clear structure, a specific implementation method and a certain portability, providing an efficient and practically applicable new method for realizing the real-time performance diagnosis of the model predictive control system, and laying a foundation for the wide application of model predictive control.

[0109] Embodiment 3

[0110] This embodiment provides a training method for model predictive control performance diagnosis based on spatio-temporal feature extraction, including:

[0111] Obtain data samples and annotation information from different performance mode categories of model predictive control;

[0112] Divide the data samples into training set samples and test set samples;

[0113] Preprocess the data samples to determine three types of annotation information: model mismatch, unmeasurable disturbance effect, and good model;

[0114] Input the preprocessed training set samples and sample annotation information into the pre-constructed spatio-temporal feature extraction model of the model predictive control system for training;

[0115] Detect the classification accuracy of the trained spatio-temporal feature extraction model of the model predictive control system through the test set samples. If the classification accuracy reaches the preset accuracy requirement, stop training; otherwise, continue to train the spatio-temporal feature extraction model of the model predictive control system through the training set samples.

[0116] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative rather than restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit of the present invention and the scope protected by the claims. All of these fall within the protection scope of the present invention.

Claims

1. A model predictive control performance diagnosis method based on spatiotemporal feature extraction, characterized in that: include: Get a pre-built dataset and pre-process it; The preprocessed data set is input into a pre-trained spatiotemporal feature extraction model of a model predictive control system to obtain a classification result output by the spatiotemporal feature extraction model of the model predictive control system; Extract the classification results of the model based on the spatiotemporal characteristics of the model predictive control system, output the categories that cause the performance degradation of the model predictive control, and complete the model predictive control performance diagnosis; Among them, the spatiotemporal feature extraction model of the model predictive control system includes a serially connected structural feature extraction model based on a graph neural network, a temporal feature extraction model based on a long short-term memory network, and a fully connected layer.

2. The model predictive control performance diagnosis method based on spatiotemporal feature extraction according to claim 1 is characterized in that: The structural feature extraction model based on graph neural network is used to extract the structural features of the model predictive control system and input the extracted structural features into the temporal feature extraction model based on long short-term memory network; The time series feature extraction model based on the long short-term memory network is used to extract the time series features of the model predictive control system, and the time series features are combined with the structural features to obtain the spatiotemporal features of the model predictive control system, and the spatiotemporal features are input into the fully connected layer; The fully connected layer is used to perform classification according to the spatiotemporal features and output the category that causes performance degradation of the model predictive control.

3. The model predictive control performance diagnosis method based on spatiotemporal feature extraction according to claim 1 is characterized in that: The pre-built data set includes building a data set based on historical operating data under different performance modes of model predictive control; Each data in the historical operation data includes the input variable value, output variable value of the model predictive control system at the same moment and the performance mode of the model predictive control currently running on the model predictive control system; The different performance modes of the model predictive control include good model, model gain mismatch, model time constant mismatch, model lag mismatch and disturbance characteristic change.

4. The model predictive control performance diagnosis method based on spatiotemporal feature extraction according to claim 1 is characterized in that: The structural feature extraction model based on graph neural network includes a serially connected graph structure input layer, a multi-order graph convolution layer and a graph feature aggregation layer; The graph structure input layer is used to receive the input variable values ​​and the output variable values, map the input variable values ​​and the output variable values ​​into initial feature vectors of graph nodes, and construct an adjacency matrix according to the connection relationship of the model predictive control system variables; The multi-order graph convolution layer includes two layers of serially connected graph convolution modules, wherein the first layer of graph convolution modules is used to receive the initial feature vector and adjacency matrix of the graph node and perform graph convolution operation to obtain the updated graph node and adjacency matrix; the second layer of graph convolution module is used to receive the updated graph node and adjacency matrix, and perform graph convolution operation again to obtain the structural features extracted at the current moment; The graph feature aggregation layer is used to receive the structural features extracted by the multi-order graph convolution layer at each moment and splice them into a structural feature sequence in chronological order as the input of the temporal feature extraction model based on the long short-term memory network.

5. The model predictive control performance diagnosis method based on spatiotemporal feature extraction according to claim 4 is characterized in that: The structural feature extraction model based on graph neural network extracts features through the following formula: ， ， In the formula, is the message passing function, For the layer The embedding representation of the node, For the The embedding representation of the node, for Node and Embedding representation of edges between nodes, message passing function polymerization and The embedding representation of the node exist Aggregate messages for layers , update function Will combine aggregate messages and the characteristics of the node itself To update Layer characteristics , that is The structural features of the model predictive control system extracted by the layer.

6. The model predictive control performance diagnosis method based on spatiotemporal feature extraction according to claim 4 is characterized in that: The correlation between the variables as edges in the graph structure includes: If an input variable in the model predictive control system is correlated with another input variable, there is an edge between the two, otherwise there is no edge; if an input variable in the model predictive control system is correlated with an output variable, there is an edge between the two, otherwise there is no edge; if an output variable in the model predictive control system is correlated with another output variable, there is an edge between the two, otherwise there is no edge.

7. The model predictive control performance diagnosis method based on spatiotemporal feature extraction according to claim 1 is characterized in that: Constructing the temporal feature extraction model based on the long short-term memory network includes the following steps: The data with structural features extracted by the structural feature extraction model based on graph neural network is input into the long short-term memory network. The sequence data of each variable in the time window is extracted with the time series features through an independent long short-term memory network. The specific steps are as follows: Step 1: For each variable , in the time window length The historical sequence data in is , the expression is as follows: ， In the formula, For variables At the moment The observed value of The value range is from t-T+1 to t, where t is the current time. The output of is used as the input of the long short-term memory network , the input of the long short-term memory network The expression is: ， In the formula, for Structural features of the output of the moment graph neural network; Step 2: For each variable Structural feature sequence , the long short-term memory network unit is processed step by step according to the time, and the processing process includes: Structural features and the previous hidden state Input to the input gate and forget gate of the long short-term memory network unit to calculate the candidate cell state , according to the candidate cell state Update the cell status to get the current cell status , hide the previous state and the current cell state Input to the output gate to calculate the current hidden state ; Step 3, after After time steps, take the final hidden state of the long short-term memory network As a variable The spatiotemporal coupling characteristics of all variables are spliced ​​into a global time series feature vector by the following formula: : 。 8. The model predictive control performance diagnosis method based on spatiotemporal feature extraction according to claim 1 is characterized in that: The training method of the model predictive control performance diagnosis includes: Obtain data samples and annotation information from different performance mode categories of model predictive control; Divide the data samples into training set samples and test set samples; Preprocess the data samples to determine three types of annotation information: model mismatch, unmeasurable disturbance impact, and model goodness; Input the preprocessed training set samples and sample annotation information into a pre-built model predictive control system spatiotemporal feature extraction model for training; The classification accuracy of the trained model predictive control system spatiotemporal feature extraction model is detected through the test set samples. If the classification accuracy reaches the preset accuracy requirement, the training is stopped. Otherwise, the training of the model predictive control system spatiotemporal feature extraction model continues through the training set samples.

Images