Mismatch diagnosis method for predictive control system model
By generating image data sets in the prediction control system and using image classification neural network for diagnosis, the lag and accuracy of model mismatch diagnosis in multivariable, strongly coupled, and nonlinear systems is solved, and efficient and accurate model mismatch diagnosis is achieved.
Patent Information
- Application Number
- CN202510650037.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-20
- Publication Date
- 2025-08-08
AI Technical Summary
In the prior art, in multivariable, strongly coupled, nonlinear prediction control systems, model mismatch diagnosis has the problem of strong lag and low real-time diagnostic accuracy.
By simulating the normal state and mismatch state of the prediction control system, one-dimensional time series of key variables are collected, image data sets are generated using the image transformation matrix, and state labels are marked, input to the image classification neural network for feature extraction and real-time diagnosis, and deep mining of image features is combined with deep learning models.
It significantly improves the accuracy and real-time nature of model mismatch diagnosis, overcomes the lack of nonlinear adaptability of traditional methods in complex industrial scenarios, reduces the risk of misjudgment, and provides more efficient and reliable diagnostic methods.
Smart Images
Figure CN120447525A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a prediction control system model mismatch diagnosis method, belonging to the technical field of industrial process manufacturing. Background Art
[0002] Control technology is crucial as modern industrial automation evolves towards efficiency, precision, and safety. In recent years, modern control theory has advanced rapidly, with advanced control (APC) technology, a leading area of advancement, garnering significant academic attention and becoming a research focus. Model predictive control (MPC), a key branch of advanced control, builds predictive models of the system and combines them with optimization algorithms to accurately predict and optimize future system behavior. This allows for proactive adjustments to the system to address changes, significantly improving control performance and stability. Leveraging these advantages, MPC technology has been widely adopted in numerous industrial sectors, including chemical, steel, metallurgy, and electric power. It has not only significantly improved the safety of production processes and reduced accident risks, but has also generated substantial economic benefits for businesses by optimizing production processes, improving product quality, and increasing production efficiency, thereby promoting their sustainable development.
[0003] Despite the remarkable achievements of MPC technology, in the actual process manufacturing industry, MPC systems often face numerous problems after a period of operation due to factors such as equipment aging, catalyst deactivation, and frequent fluctuations in production loads. Among these, the mismatch between the prediction model and the actual process (model mismatch MPM) is a maintenance pain point. Model mismatch can lead to mismatched controller parameters and altered disturbance characteristics, resulting in system performance degradation, impacting product quality, and even threatening production safety. To address this issue, traditional data-driven model diagnostic methods (such as multivariate covariance analysis and principal component analysis) utilize large amounts of sensor data. However, in the modern complex process control industry, faced with the high complexity of production processes, the multidimensionality of data, and the nonlinear characteristics, it is difficult to quickly and accurately diagnose system prediction model defects, thus impacting system maintenance and the normal operation of the enterprise. Summary of the Invention
[0004] The purpose of the present invention is to provide a method for diagnosing model mismatch in a predictive control system. By converting the one-dimensional time series of key variables in the predictive control system into a two-dimensional image, the problem of strong hysteresis and low accuracy of real-time diagnosis of the model in the existing technology for multivariable, strongly coupled, and nonlinear predictive control systems is solved.
[0005] In order to solve the above technical problems, the present invention is implemented by adopting the following technical solutions:
[0006] The present invention provides a method for diagnosing model mismatch in a predictive control system, comprising:
[0007] Simulate the predictive control system in the normal model state and the model mismatch state respectively;
[0008] Collect one-dimensional time series of key variables of simulation predictive control system;
[0009] The one-dimensional time series is visualized through a preset image transformation matrix to generate image datasets under the normal model state and the model mismatch state, and each image is annotated with the corresponding state label;
[0010] The labeled image dataset is input into the image classification neural network, and the neural network is trained by extracting image features;
[0011] The trained image classification neural network is used to perform real-time diagnosis on the predictive control system to be diagnosed.
[0012] Furthermore, the predictive control system transfer function is used to simulate the predictive control system in the normal model state and the model mismatch state respectively.
[0013] Furthermore, after collecting the one-dimensional time series of each key variable of the simulation predictive control system, it also includes:
[0014] The correlation coefficient is used to screen key variables whose correlation strength exceeds a preset threshold value to obtain the system multivariate correlation matrix;
[0015] According to the system multivariate correlation matrix, the key variables whose correlation strength exceeds the preset threshold are sequence reconstructed by principal component analysis method to obtain the reconstructed sequence of key variables;
[0016] The reconstructed sequence of key variables is divided by sliding windows to obtain the time windows of the divided key variables;
[0017] Within the time window, the segmented key variables are linearly scaled using a standard normalization method and then mapped to a polar coordinate system using a polar coordinate mapping formula.
[0018] Furthermore, the calculation formula of the correlation coefficient is expressed as:
[0019] ;
[0020] Where, represents the correlation coefficient, Representing variables middle samples, Representing variables middle samples, Representing variables The sample mean of Representing variables The sample mean of .
[0021] Furthermore, the system multivariate correlation matrix is expressed as:
[0022] ;
[0023] Where, represents the system multivariate correlation matrix, Indicates the first key variable and the second key variable The correlation coefficient of Indicates the first key variable and the second key variable The covariance of and Represent the first key variable and the second key variable The standard deviation of Indicates the first key variable and the second key variable in the system multivariate correlation matrix The correlation between the key variables and They represent the input dimension and output dimension of the predictive control system respectively.
[0024] Furthermore, the reconstruction sequence of the key variables is expressed as:
[0025] ;
[0026] Where, represents the reconstructed sequence of key variables, represents the first key variable, represents the mean vector of the first key variable, represents the standard deviation of the first key variable, Indicates the first key variable The eigenvectors of the principal components.
[0027] Furthermore, the time window is expressed as:
[0028] ;
[0029] Where, Indicates the time window, represents the sliding step length, Indicates the window size, Indicates the first key variable sample points.
[0030] Furthermore, the polar coordinate mapping formula is expressed as:
[0031] ;
[0032] Where, represents the angle component in polar coordinates, Represents the first key variable after normalization The samples The inverse function value on the cosine function, , represents the maximum value of the sample in the first key variable, represents the minimum value of the sample in the first key variable, represents the radius component in polar coordinates, Indicates the first key variable The time step where the sample is located, Represents the total number of samples in the first key variable.
[0033] Furthermore, the preset image conversion matrix is determined by permutation and combination of the GAF image conversion formula, wherein the GAF image conversion formula is expressed as:
[0034] ;
[0035] Where, Indicates the GAF images, Indicates the first key variable The angle of the first sample in Indicates the first key variable The The angle of the sample, represents the cosine function, represents the first key variable after normalization, express The transpose of Represents the identity matrix.
[0036] Furthermore, the preset image conversion matrix is expressed as:
[0037] ;
[0038] Where, Represents the preset multivariate fusion image transformation matrix, Represents the first GAF image transformation matrix, Represents the second GAF image transformation matrix, Indicates the A GAF image transformation matrix.
[0039] Compared with the prior art, the present invention has the following beneficial effects:
[0040] 1. The present invention simulates a predictive control system under normal model conditions and model mismatch conditions and collects one-dimensional time series of key variables. It uses a preset image conversion matrix to convert the sequence data into a visual image, annotates the state labels to form a data set, and finally inputs the image classification neural network for feature extraction and real-time diagnosis, which significantly improves the accuracy and real-time performance of model mismatch diagnosis. Among them, the visualization form of the two-dimensional image can intuitively display the dynamic characteristics of the system. Combined with the deep learning model's ability to deeply mine image features, it effectively overcomes the problem of traditional methods' lack of adaptability to nonlinear and non-stationary signals in complex industrial scenarios. At the same time, it avoids the risk of misjudgment caused by input-output coupling in traditional data-driven methods, thereby providing a more efficient and reliable model mismatch diagnosis method for industrial process control, and solving the problem that the existing technology has strong hysteresis and low model real-time diagnosis accuracy for multi-variable, strongly coupled, nonlinear predictive control systems.
[0041] 2. This invention uses a variable correlation calculation formula to screen out key variables whose correlation strength exceeds a preset threshold. Principal component analysis (PCA) is then used to reconstruct the sequence of these screened variables, achieving precise dimensionality reduction for high-dimensional data. This process effectively removes redundant information while preserving the core characteristics of the system's dynamics, significantly improving the data quality of subsequent diagnostic models and laying the foundation for accurate diagnosis of model mismatch.
[0042] 3. This invention uses a sliding window to segment the reconstructed time series of key variables and converts the one-dimensional time series into a two-dimensional image through standard normalization and polar coordinate mapping. Combined with a pre-defined image conversion matrix (such as the GAF conversion formula), this method enables multi-dimensional visualization of the data. This visualization method not only preserves the time-frequency characteristics of the time series but also reveals the dynamic patterns and periodic characteristics of the data through polar coordinate mapping. This makes abnormal patterns of model mismatch more intuitive and easier to identify in the two-dimensional image, significantly improving the accuracy and reliability of diagnosis.
[0043] 4. This method feeds a two-dimensional image dataset labeled with state labels into an image classification neural network. Through deep mining and integration of image features using a deep learning model, it achieves real-time diagnosis of the state of a predictive control system. This method leverages the advantages of deep learning in image recognition, automatically extracting complex features from images and accurately classifying normal and mismatched model states. Compared to traditional methods, this method significantly improves diagnostic efficiency and reduces manual intervention, providing strong technical support for real-time monitoring and optimization of industrial processes. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 1 is a flow chart of a method for diagnosing model mismatch in a predictive control system according to an embodiment of the present invention;
[0045] Figure 2 Schematic diagram of the structure of the Wood-Berry binary distillation column model provided in an embodiment of the present invention;
[0046] Figure 3 4 is a schematic diagram of the model mismatch diagnosis result provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0047] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.
[0048] Example 1
[0049] like Figure 1 As shown, this embodiment introduces a method for diagnosing model mismatch in a predictive control system, including:
[0050] Step 1: Simulate the predictive control system in the normal model state and the model mismatch state respectively.
[0051] This paper simulates the predictive control system under both normal and mismatched model conditions by adjusting its parameters. This generates one-dimensional time series data containing key variables under both normal and mismatched model conditions, contributing to a deeper understanding of the predictive control system's behavior, performance, and potential failure modes under different conditions. While the normal model condition represents the operation of the predictive control system under ideal or expected conditions, the mismatched model condition represents the operation of the predictive control system under abnormal conditions such as parameter deviations and external interference. Detailed analysis of the system under both conditions facilitates early detection of potential system failures and the development of appropriate response strategies.
[0052] Step 2: Collect the one-dimensional time series of each key variable of the simulation predictive control system.
[0053] The one-dimensional time series of each key variable of the predictive control system contains the key variables of the predictive control system during operation, reflecting the dynamic change information of important characteristics such as the system state, input, output and internal parameters. By collecting time series data, the present invention can fully understand the operation of the system at different times. By analyzing the time series data, the stability, trend changes and abnormal fluctuations of the system can be understood, which provides a possible way to timely discover system failures.
[0054] Step 3: Visualize the one-dimensional time series through a preset image transformation matrix to generate image datasets under the normal model state and the model mismatch state, and annotate each image with the corresponding state label.
[0055] The present invention converts one-dimensional time series data into image form through a preset image conversion matrix, making certain features in the data more prominent in the image. At the same time, each image is annotated with a corresponding state label, providing supervision information for the image classification task. The neural network can learn the differences between images in the normal state of the model and the model mismatch state, thereby realizing accurate classification and diagnosis of the state of the predictive control system.
[0056] Compared to one-dimensional time series data, image-based data is easier for humans to understand and analyze. By observing images, researchers can intuitively grasp the operating characteristics of predictive control systems under different states, which facilitates in-depth analysis and explanation of system failures. Furthermore, image datasets provide a rich source of training samples for image classification neural networks. By learning features and patterns in images, neural networks can automatically extract information useful for predictive control system state diagnosis, thereby improving the automation and accuracy of fault diagnosis.
[0057] Step 4: Input the labeled image dataset into the image classification neural network and train the neural network by extracting image features.
[0058] Image classification neural networks have excellent feature extraction and classification capabilities. They can automatically capture features closely related to the system status from the input image, accurately classify the image based on these features, and clearly determine whether the predictive control system is in a normal model state or a model mismatch state, without the need for manual pre-setting of complex feature extraction rules.
[0059] Step 5: Use the trained image classification neural network to perform real-time diagnosis on the predictive control system to be diagnosed.
[0060] During real-time diagnosis, the neural network efficiently processes and analyzes real-time image data. If a system mismatch is detected, it quickly issues an alarm, alerting personnel to immediate attention. Furthermore, the neural network automatically or with human assistance can implement appropriate troubleshooting measures based on pre-set fault handling strategies, effectively preventing further escalation of the system failure and ensuring stable operation. This real-time diagnostic capability enables predictive control systems to rapidly respond to abnormal situations and minimize losses caused by failures.
[0061] Example 2
[0062] Based on the same inventive concept as Example 1, this example introduces the implementation steps of a predictive control system model mismatch diagnosis method, including:
[0063] Step 1: Simulate the predictive control system in the normal model state and the model mismatch state respectively.
[0064] This embodiment is described with respect to the Wood-Berry binary distillation tower model, wherein the structure of the Wood-Berry binary distillation tower model is as follows: Figure 2 As shown, the Wood-Berry binary distillation column model in the normal model state and the model mismatch state is simulated respectively by using the predictive control system transfer function, wherein the predictive control system transfer function is expressed as:
[0065] ;
[0066] Where, Indicates the purity of the product at the top of the control tower, Indicates the concentration of residual components at the bottom of the control tower, Indicates adjusting the reflux ratio to control the top of the tower. Indicates adjusting the amount of heating steam to control the bottom of the tower. represents the feed disturbance, represents a complex variable in the Laplace transform.
[0067] Use the MPC toolbox in MATLAB to configure the performance mode state corresponding to Table 1, which is as follows:
[0068] Table 1 Predictive control performance mode description
[0069] Performance Mode describe System parameter settings Good model Good model none Model mismatch First channel gain 12.8 becomes 25.6 Unmeasured disturbance impact F(s) Random perturbation with variance 0.15
[0070] Step 2: Collect the one-dimensional time series of each key variable of the simulated predictive control system.
[0071] In this embodiment, the key variables collected for the Wood-Berry binary distillation column model include: system input sequence , the system output sequence , the system's predicted output sequence and the systematic error sequence .
[0072] The key variables of the Wood-Berry binary distillation column model collected in this embodiment also include a model quality index (MQI). The model quality index (MQI) is a key indicator for evaluating the accuracy of the binary distillation column model.
[0073] The model quality index (MQI) is expressed as:
[0074] ;
[0075] Where, Indicates the selected data length, represents the first values, express The transpose of The first values, express , Represents the output weight coefficient of the predictive control system controller.
[0076] In this embodiment, when the value of the model quality index (MQI) is smaller, it indicates that the binary distillation column model is closer to the mismatch state, and when the value of the model quality index (MQI) is closer to 1, it indicates that the binary distillation column model is closer to the normal state.
[0077] In this embodiment, after collecting the one-dimensional time series of each key variable of the simulation predictive control system, the following steps are also included:
[0078] Step 2.1: Use the correlation coefficient to screen key variables whose correlation strength exceeds the preset threshold to obtain the system multivariate correlation matrix.
[0079] In this embodiment, the input sequence of the system is calculated by the variable correlation formula , system output sequence , the system's predicted output sequence , the actual error sequence of the system , the system's prediction error sequence The correlation between the prediction model quality index (MQI) and the key variables whose correlation strength exceeds the preset threshold are measured to form the system multivariate correlation matrix.
[0080] In this embodiment, the calculation formula of the correlation coefficient is expressed as:
[0081] ;
[0082] Where, represents the correlation coefficient, Representing variables middle samples, Representing variables middle samples, Representing variables The sample mean of Representing variables The sample mean of .
[0083] In this embodiment, the system multivariate correlation matrix is expressed as:
[0084] ;
[0085] Where, represents the system multivariate correlation matrix, Representing variables and variables The correlation coefficient of Representing variables and variables The covariance of and Represent variables respectively and variables The standard deviation of Indicates the variables and The correlation between variables, and Represent the input dimension and output dimension of the system respectively.
[0086] Step 2.2: Based on the system multivariate correlation matrix, the principal component analysis method is used to reconstruct the sequence of key variables whose correlation strength exceeds the preset threshold to obtain the reconstructed sequence of key variables.
[0087] Step 2.2.1: Construct a standardized matrix based on key variables whose correlation strength exceeds a preset threshold.
[0088] Step 2.2.2: Construct the covariance matrix based on the standardized matrix.
[0089] Step 2.2.3: Perform eigenvalue decomposition on the covariance.
[0090] Step 2.2.4: Based on the eigenvector matrix obtained by eigenvalue decomposition, reconstruct the sequence of key variables with high correlation to obtain the reconstructed sequence of key variables, which is expressed as:
[0091] ;
[0092] Where, represents the reconstructed sequence of key variables, represents the first key variable, represents the mean vector of the first key variable, represents the standard deviation of the first key variable, Indicates the first key variable The eigenvectors of the principal components.
[0093] Step 2.3: Split the reconstructed sequence of key variables through a sliding window to obtain the time window of the split key variables.
[0094] In this embodiment, the time window is expressed as:
[0095] ;
[0096] Where, Indicates the time window, represents the sliding step length, Indicates the window size, Indicates the first key variable sample points.
[0097] Step 2.4: Within the time window, the segmented key variables are linearly scaled using a standard normalization method and then mapped to a polar coordinate system using a polar coordinate mapping formula.
[0098] In this embodiment, the polar coordinate mapping formula is expressed as:
[0099] ;
[0100] Where, represents the angle component in polar coordinates, Represents the first key variable after normalization The samples The inverse function value on the cosine function, , represents the maximum value of the sample in the first key variable, represents the minimum value of the sample in the first key variable, represents the radius component in polar coordinates, Indicates the first key variable The time step where the sample is located, Represents the total number of samples in the first key variable.
[0101] Step 3: Visualize the one-dimensional time series through a preset image transformation matrix to generate image datasets under the normal model state and the model mismatch state, and annotate each image with the corresponding state label.
[0102] In this embodiment, the preset image conversion matrix is determined by permutation and combination of the GAF image conversion formula, wherein the GAF image conversion formula is expressed as:
[0103] ;
[0104] Where, Indicates the GAF images, Indicates the first key variable The angle of the first sample in Indicates the first key variable The The angle of the sample, represents the cosine function, represents the first key variable after normalization, express The transpose of Represents the identity matrix.
[0105] In this embodiment, the preset image conversion matrix is expressed as:
[0106] ;
[0107] Where, Represents the preset multivariate fusion image transformation matrix, Represents the first GAF image transformation matrix, Represents the second GAF image transformation matrix, Indicates the A GAF image transformation matrix.
[0108] In this embodiment, images in a normal model state are labeled "1" and images in a mismatched model state are labeled "0" to obtain a labeled image dataset.
[0109] Step 4: Input the labeled image dataset into the image classification neural network and train the predictive control system by extracting image features.
[0110] The present invention builds an image classification neural network by introducing a ResNet neural network with residual connections, avoiding the degradation phenomenon of directly "skipping" certain layers of input information, allowing the network to effectively train deeper layers, thereby improving the accuracy of this embodiment.
[0111] In this embodiment, the structure of the ResNet neural network includes an input layer, an initial convolutional layer, a residual block, a pooling layer, a fully connected layer, and a Softmax layer, wherein the parameter settings of each layer are shown in Table 2:
[0112] Table 2 ResNet neural network parameters
[0113] Layer Name Detailed structure Output size Input layer 224×224×3 input image 224×224×3 Conv1 7×7 convolution, stride 2, 64 channels 112×112×64 MaxPool 3×3 max pooling with a stride of 2 56×56×64 ResBlock 1 2 3×3 convolutions, 64 channels 28×28×128 ResBlock 2 2 3×3 convolutions, 128 channels 14×14×256 ResBlock 3 2 3×3 convolutions, 256 channels 7×7×512 ResBlock 4 2 3×3 convolutions, 512 channels 1×1×512 AvgPool Global average pooling 1×1×512 FC layer Fully connected layer (2-class classification) 2 Softmax Calculating class probabilities 2
[0114] In this embodiment, the input layer is used to receive the labeled image dataset and input it into the initial convolutional layer.
[0115] In this embodiment, the initial convolution layer is used to extract image features of the labeled image data set and input them into the residual block, wherein the initial convolution layer adopts a 7×7 convolution kernel, a step size of 2, 64 output channels, and uses the ReLU function as the activation function.
[0116] In this embodiment, the residual block is used to input image features directly into the pooling layer through a residual connection, which effectively solves the degradation problem in the deep network and enables the number of network layers to be deepened without causing performance degradation. Each residual block includes two 3×3 convolutional layers, and the output of the first convolutional layer is added to the output of the second convolutional layer through a residual connection.
[0117] Among them, the output of the first convolutional layer is expressed as:
[0118] ;
[0119] In the formula, Represents the output of the first convolutional layer, represents a nonlinear activation function, represents the convolution operation, Represents input data, represents the first layer of convolution kernel, Represents the bias term of the first convolution layer.
[0120] Among them, the output of the second convolutional layer is expressed as:
[0121] ;
[0122] Where, Represents the output result of the second layer of convolution, Represents the output of the first layer, represents the weight of the second layer convolution kernel, represents the convolution operation, Represents the bias term of the second convolution layer
[0123] Among them, the internal convolution output of each residual block is expressed as:
[0124] ;
[0125] Where, represents the internal convolution output of the residual block, represents the residual mapping, represents the convolutional layer in the residual map parameters.
[0126] Among them, the final output of the residual block is expressed as:
[0127] ;
[0128] Where, Represents the final output of the residual block.
[0129] In this embodiment, the pooling layer is used to downsample the image features through a 2×2 convolution window and input them into the fully connected layer to reduce the amount of calculation.
[0130] In this embodiment, the fully connected layer is used to flatten the image features into a vector and perform classification through a Softmax layer.
[0131] Among them, the output of the fully connected layer is expressed as:
[0132] ;
[0133] Where, represents the output of the fully connected layer, Represents a full connection operation.
[0134] In this embodiment, the Softmax layer is used to output the classification probability of the image features using a softmax activation function.
[0135] Step 5: Use the trained image classification neural network to perform real-time diagnosis on the predictive control system to be diagnosed.
[0136] Example 3
[0137] Based on the same inventive concept as other embodiments, this embodiment introduces a computer-readable storage medium on which computer instructions are stored. When the computer instructions are executed by a processor, the steps of the method of the above-mentioned embodiment 1 or 2 are implemented.
[0138] Example 4
[0139] Based on the same inventive concept as other embodiments, this embodiment introduces a computer program product, including computer instructions. When the computer instructions are executed by a processor, the steps of the method in the above-mentioned embodiment 1 or 2 are implemented.
[0140] In summary, the present invention simulates the predictive control system under the normal state and the model mismatch state and collects the one-dimensional time series of key variables, uses the preset image conversion matrix to convert the sequence data into a visual image, and annotates the state label to form a data set, and finally inputs the image classification neural network for feature extraction and real-time diagnosis, thereby significantly improving the accuracy and real-time performance of model mismatch diagnosis. Among them, the visualization form of the two-dimensional image can intuitively display the dynamic characteristics of the system. Combined with the deep learning model's ability to deeply mine image features, it effectively overcomes the problem of insufficient adaptability of traditional methods to nonlinear and non-stationary signals in complex industrial scenarios, while avoiding the risk of misjudgment caused by input-output coupling of traditional data-driven methods, thereby providing a more efficient and reliable model mismatch diagnosis method for industrial process control, and solving the problem that the existing technology has strong hysteresis and low model real-time diagnosis accuracy for multi-variable, strongly coupled, nonlinear predictive control systems.
[0141] This method uses a variable correlation calculation formula to screen out key variables whose correlation strength exceeds a preset threshold. Principal component analysis (PCA) is then used to reconstruct the sequence of these screened variables, achieving precise dimensionality reduction for high-dimensional data. This process effectively removes redundant information while preserving the core characteristics of the system's dynamics. This significantly improves the data quality of subsequent diagnostic models and lays the foundation for accurate diagnosis of model mismatch.
[0142] This method uses a sliding window to segment the reconstructed time series of key variables and converts the one-dimensional time series into a two-dimensional image through standard normalization and polar coordinate mapping. Combined with a pre-defined image conversion matrix (such as the GAF conversion formula), this method enables multidimensional visualization of the data. This visualization method not only preserves the time-frequency characteristics of the time series but also reveals the dynamic patterns and periodic characteristics of the data through polar coordinate mapping. This makes abnormal patterns of model mismatch more intuitive and easier to identify in the two-dimensional image, significantly improving the accuracy and reliability of diagnosis.
[0143] This method feeds a two-dimensional image dataset labeled with state labels into an image classification neural network. Through deep mining and fusion of image features using a deep learning model, it achieves real-time diagnosis of the state of a predictive control system. This method leverages the advantages of deep learning in image recognition, automatically extracting complex features from images and accurately classifying normal and mismatched model states. Compared to traditional methods, this method significantly improves diagnostic efficiency and reduces manual intervention, providing strong technical support for real-time monitoring and optimization of industrial processes.
[0144] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0145] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0146] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.
[0147] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.
[0148] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which are all protected by the present invention.
Claims
1. A method for diagnosing model mismatch in a predictive control system, characterized in that: include: Simulate the predictive control system in the normal model state and the model mismatch state respectively; Collect one-dimensional time series of key variables of simulation predictive control system; The one-dimensional time series is visualized through a preset image transformation matrix to generate image datasets under the normal model state and the model mismatch state, and each image is annotated with the corresponding state label; The labeled image dataset is input into the image classification neural network, and the neural network is trained by extracting image features; The trained image classification neural network is used to perform real-time diagnosis on the predictive control system to be diagnosed.
2. The predictive control system model mismatch diagnosis method according to claim 1, characterized in that: The predictive control system in the normal model state and the model mismatch state is simulated respectively through the predictive control system transfer function.
3. The predictive control system model mismatch diagnosis method according to claim 1, characterized in that: After collecting the one-dimensional time series of each key variable of the simulated predictive control system, it also includes: The correlation coefficient is used to screen key variables whose correlation strength exceeds a preset threshold value to obtain the system multivariate correlation matrix; According to the system multivariate correlation matrix, the key variables whose correlation strength exceeds the preset threshold are sequence reconstructed by principal component analysis method to obtain the reconstructed sequence of key variables; The reconstructed sequence of key variables is divided by sliding windows to obtain the time windows of the divided key variables; Within the time window, the segmented key variables are linearly scaled using a standard normalization method and then mapped to a polar coordinate system using a polar coordinate mapping formula.
4. The predictive control system model mismatch diagnosis method according to claim 3, characterized in that: The calculation formula of the correlation coefficient is expressed as: ； Where, represents the correlation coefficient, Indicates the first key variable The samples, Represents the second key variable The samples, Indicates the first key variable The sample mean of Represents the second key variable The sample mean of .
5. The predictive control system model mismatch diagnosis method according to claim 3, characterized in that: The system multivariate correlation matrix is expressed as: ； Where, represents the system multivariate correlation matrix, Indicates the first key variable and the second key variable The correlation coefficient of Indicates the first key variable and the second key variable The covariance of and Represent the first key variable and the second key variable The standard deviation of Indicates the first key variable and the second key variable in the system multivariate correlation matrix The correlation between the key variables and They represent the input dimension and output dimension of the predictive control system respectively.
6. The predictive control system model mismatch diagnosis method according to claim 3, characterized in that: The reconstruction sequence of the key variables is expressed as: ； Where, represents the reconstructed sequence of key variables, represents the first key variable, represents the mean vector of the first key variable, represents the standard deviation of the first key variable, Indicates the first key variable The eigenvectors of the principal components.
7. The predictive control system model mismatch diagnosis method according to claim 3, characterized in that: The time window is expressed as: ； Where, Indicates the time window, represents the sliding step length, Indicates the window size, Indicates the first key variable sample points.
8. The predictive control system model mismatch diagnosis method according to claim 3, characterized in that: The polar coordinate mapping formula is expressed as: ； Where, represents the angle component in polar coordinates, Represents the first key variable after normalization The samples The inverse function value on the cosine function, , represents the maximum value of the sample in the first key variable, represents the minimum value of the sample in the first key variable, represents the radius component in polar coordinates, Indicates the first key variable The time step where the sample is located, Represents the total number of samples in the first key variable.
9. The predictive control system model mismatch diagnosis method according to claim 3, characterized in that: The preset image conversion matrix is determined by permutation and combination of the GAF image conversion formula, wherein the GAF image conversion formula is expressed as: ； Where, Indicates the GAF images, Indicates the first key variable The angle of the first sample in Indicates the first key variable The The angle of the sample, represents the cosine function, represents the first key variable after normalization, express The transpose of Represents the identity matrix.
10. The predictive control system model mismatch diagnosis method according to claim 9, characterized in that: The preset image conversion matrix is expressed as: ； Where, Represents the preset multivariate fusion image transformation matrix, Represents the first GAF image transformation matrix, Represents the second GAF image transformation matrix, Indicates the A GAF image transformation matrix.