A method for diagnosing the root cause of performance degradation of DRN-SVM predictive control based on rolling feature time domain

Abstract

The application provides a DRN-SVM prediction control performance degradation source diagnosis method based on a rolling feature time domain, belongs to the machine learning and target classification fields, and comprises the following steps: step 1: historical data in a performance degradation mode of a prediction controller is obtained through a Wood-Berry rectifying column simulation model and is used as a training set; step 2: dynamic feature extraction is realized through a deep residual network based on rolling promotion of the rolling feature time domain; step 3: a DRN-SVM classification model is constructed; step 4: the DRN-SVM classification model is trained, and a prediction control performance degradation diagnosis model is obtained; and step 5: factors causing the performance degradation of the prediction controller are determined. The application can capture dynamic time sequence features of complex industrial process data and sufficiently mine data features of the industrial process data, and thus the accuracy of the prediction control performance degradation source diagnosis is improved.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and object classification, specifically a method for diagnosing the root causes of performance degradation in DRN-SVM predictive control based on rolling feature time domain. Background Technology

[0002] Model predictive control (MPC) is the most widely used advanced control technology in industry. However, predictive controllers experience performance degradation over time. Evaluating and diagnosing the performance of predictive controllers is crucial for ensuring the continuous and efficient operation of predictive control systems.

[0003] Many factors contribute to the degradation of predictive controller performance, including process model mismatch, changes in disturbance characteristics, controller constraint saturation, and improper controller parameter tuning. Regarding the root cause diagnosis of predictive controller performance degradation, some studies focus on performance evaluation, while others treat it as a multi-classification problem. For root cause diagnosis from a performance evaluation perspective, since model mismatch is a major factor causing predictive controller performance degradation, much research focuses on model performance evaluation. For example, Qin et al. proposed detecting autocorrelation by monitoring Kalman filter information to identify whether the mismatch is process model mismatch or disturbance model mismatch. Li et al. proposed a correlation analysis method (CAID) between input and disturbance to address the model mismatch assessment problem, combining it with the Model Quality Index (MQI) for multivariate systems. Ling et al. proposed a method to evaluate model performance using the ratio of disturbance to model variance. For predictive controller performance degradation caused by controller parameters, relevant literature proposes a method to evaluate controller parameter performance by estimating the Markov parameters of the disturbance model. For predictive controller performance degradation diagnosis from a multi-classification perspective, the main approach is to start with the differences in predictive controller response data features caused by different root causes. Methods such as nonlinear mapping and PCA are used to extract data features, and then methods such as SVM are used to classify the extracted data features. For example, Tian Xuemin et al. defined eigenvector angles to describe the similarity between current data and introduced a classifier based on eigenvector angles to diagnose the root causes of predictive controller performance degradation.

[0004] The classification accuracy of multi-classification problems is directly related to the feature extraction of data; fully mining the features of the data can improve the classification accuracy to a certain extent. Currently, research on feature extraction based on the SVM classification algorithm has achieved certain results. To improve the complexity and latency of spectrum detection in the GEO-LEO satellite constellation network, Bao Jianrong et al. processed the sensing signal using PCA and then used SVM to classify the PCA results, thus constructing a high-performance PCA-SVM classification model to effectively detect low-complexity and high-performance spectrum signals. Zhou Taotao et al., addressing the insufficient technology for detecting and diagnosing mechanical defects in disconnect switches, used KPCA to extract the time-domain features of the motor output power time curve and then used SVM to classify the extracted time-domain features, constructing a KPCA-SVM classification model, which improved the classification accuracy of mechanical state parameters where the motor operating voltage was partially known and partially unknown. Wang Junhong et al., to achieve accurate muscle fatigue detection and classification, used CNN to extract surface electromyography features and then used SVM to classify the features, constructing a CNN-SVM model, which improved the accuracy of muscle fatigue detection and classification. Based on the basic SVM algorithm, the improved algorithms proposed by the aforementioned scholars have all enhanced the accuracy of SVM classification. However, due to the nonlinear, temporal, and dynamic characteristics of complex industrial process data, traditional feature extraction methods such as PCA, KPCA, and shallow CNNs are difficult to characterize the features of complex industrial process data and perform poorly in predicting controller performance degradation classification.

[0005] In recent years, with the development of deep learning technology and its successful applications in fields such as speech recognition, image recognition, natural language processing, and object detection, the advantages of deep learning in feature extraction and learning of complex industrial process data have gradually become apparent. However, due to the nonlinear, temporal, and dynamic characteristics of complex industrial process data, SVM, being essentially a shallow learning structure, struggles to represent the complex nonlinear features of industrial process data with its single-layer nonlinear mapping. Furthermore, SVM training cannot reflect the temporal characteristics of the samples, thus the model cannot reflect the dynamic changes in the data. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a DRN-SVM-based method for diagnosing the root causes of predictive control system performance degradation. First, a deep residual network is used to extract deep-level dynamic temporal features from a continuous time-domain dataset of a certain width, fully exploring the data change characteristics reflecting the performance degradation of the predictive control system. Then, an SVM is used to establish a classification and diagnostic model based on the extracted features, enabling real-time diagnosis of the root causes of performance degradation in actual predictive control systems. This invention achieves both capturing the dynamic temporal features of complex industrial process data and fully exploring the data characteristics of industrial process data, thereby improving the accuracy of predictive controller performance degradation root cause diagnosis.

[0007] A method for diagnosing the root causes of performance degradation in DRN-SVM predictive control based on rolling feature time domain includes the following steps:

[0008] Step 1: Obtain historical data under the performance degradation mode of the predictive controller using the Wood-Berry distillation column simulation model as the training set;

[0009] Step 2: Based on the rolling feature time domain, continuously roll forward with a certain step size, and then realize the dynamic feature extraction of the entire training set through a deep residual network;

[0010] Step 3: The dynamic features of the predictive control system under the performance degradation mode extracted by the deep residual network are used as the input of the SVM to construct the DRN-SVM classification model.

[0011] Step 4: Train the DRN-SVM classification model based on the training set to predict the control performance degradation diagnosis model;

[0012] Step 5: Input the predictive controller performance degradation test set obtained based on the Wood-Berry model into the DRN-SVM control performance degradation diagnostic model to determine the factors that cause predictive controller performance degradation.

[0013] Preferably, the controller performance degradation mode states in step 1 include four types: process model mismatch state, disturbance characteristic change state, controller constraint saturation state, and controller parameter improper adjustment state.

[0014] Furthermore, step 2 employs a deep residual network for feature extraction, utilizing multiple sliding weight windows to traverse the process data within the rolling feature time domain, facilitating the discovery of deep-level features in complex industrial process data.

[0015] Specifically, feature extraction using deep residual networks includes the following steps:

[0016] Step 2.1: Construct a two-dimensional matrix X from the historical dataset consisting of the controlled variable y and the manipulated variable u;

[0017] Step 2.2: Input the constructed two-dimensional data matrix X into the deep residual network;

[0018] Step 2.3: Before performing convolution feature extraction, the two-dimensional data matrix X needs to be padded at the matrix edges to prevent missing features of the matrix edge data;

[0019] Step 2.4: Extract features from the two-dimensional data matrix using the weight window in the deep residual network;

[0020] Step 2.5: Perform max pooling on the deep features extracted from multiple weight windows to select the deep features with the largest weight values;

[0021] Step 2.6: Transpose the deep features selected by the max pooling operation, and use the result as the input for the subsequent SVM.

[0022] Furthermore, a Batch Normalization (BN) strategy is introduced in the deep residual network in step 2.4 to address the issues that the whitening process alters the distribution of network layers, thereby changing the expressive power of the data within the network layers, and that the parameter information learned by the underlying network is lost during the whitening process.

[0023] Furthermore, in step 4, the parameters W and (γβ) of the deep residual network are trained, where W is the weight set of the deep residual network and (γβ) is the set of weights and biases of the deep residual network introduced by the BN strategy, so that the features of the process data extracted by the network are distributed as far away from the optimal hyperplane as possible.

[0024] Preferably, the objective function of the DRN-SVM classification model is:

[0025]

[0026] in, This indicates that the objective function of DRN-SVM is to be minimized, ω ij Let X represent the normal vector of the optimal classification hyperplane for the performance degradation data of the predictive controller in classes i and j. Let C represent the penalty parameter, used to adjust the distribution of features extracted by the deep residual network on both sides of the classification hyperplane, so that the features are as far away from the classification hyperplane as possible. Let a represent the index of the data feature vector extracted by the deep residual network. N is the number of data features of the process data under the performance mode state of the predictive controller extracted by the deep residual network. i and X j This represents the datasets showing performance degradation in classes i and j of the predictive controller; BN 1，a (g(φ(Xi ;W));(γ,β)) represents the data features of the i-th type of performance degradation data of the predictive controller extracted by the deep residual network, BN 1，a (g(φ(X j ;W));(γ,β)) represents the data features of the j-th type of performance degradation data of the predictive controller extracted by the deep residual network; W (l) W represents the weights of the convolutional layers in a deep residual network. (k) λ1 and λ2 represent the weights of the input dimension of the deep residual network; they correspond to the two regularization terms of the network, representing the network's penalty coefficients, and λ1 represents the network's convolution weights W. (l) The penalty coefficient, λ2 represents the weight W adjusted for the network input dimension. (k) The penalty coefficient; |||| represents the magnitude of the vector, |||| 2 Represents the square of the modulus. This represents the square of the Frobenius norm.

[0027] Furthermore, the rule for applying the penalty parameter C is as follows:

[0028]

[0029] Where ↑ represents increasing the penalty parameter C, ↓ represents decreasing the penalty parameter C, and D represents the square of the distance between the features of the process data in the feature space of the predictive controller under the i-th and j-th performance modes extracted by the deep residual network. The difference of squares is expressed as follows:

[0030]

[0031] If D > 0, it means that the process data features extracted by the deep residual network in the feature space are not as far away from the optimal classification hyperplane as possible, and a larger penalty parameter C should be applied; if D < 0, it means that the extracted process data features are as far away from the optimal classification hyperplane as possible, and a smaller penalty parameter C should be applied.

[0032] Furthermore, the performance of the improved DRN-SVM performance degradation diagnostic model was evaluated using accuracy.

[0033] Specifically, the accuracy metric formula is as follows:

[0034]

[0035] Here, Accuracy represents the accuracy rate, and TP and FP represent the number of samples that were correctly identified as true and the number of samples that were incorrectly identified as true, respectively.

[0036] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention addresses the problem of classifying and diagnosing the root causes of performance degradation in predictive control systems. It proposes a DRN-SVM method based on rolling feature time domain for diagnosing the root causes of predictive control system performance degradation. First, a deep residual network is used to extract deep-level dynamic time-series features from continuous time-domain data of a certain width, fully exploring the data change characteristics reflecting the performance degradation of the predictive control system. Then, an SVM is used to establish a classification and diagnostic model based on the extracted features, which is used for real-time diagnosis of the root causes of performance degradation in actual predictive control systems. This invention achieves both capturing the dynamic time-series features of complex industrial process data and fully exploring the data features of industrial process data, thereby improving the accuracy of diagnosing the root causes of predictive controller performance degradation.

[0037] 1. Due to the nonlinear, dynamic, and temporal characteristics of complex industrial process data, this invention proposes the concept of a "rolling feature time domain." Assuming the width of the rolling feature time domain is L and the rolling step size is h, feature extraction at time k considers not only the process variable characteristics at the current time but also the dynamic changes within the subsequent k+L-1 time range. This reflects the performance characteristics of the control system within the width L. Since DRN possesses the characteristic of deep-level complex nonlinear feature extraction, the extracted features can accurately capture the complex dynamic temporal performance within the width L. At the next rolling time k+1, the dynamic feature extraction within the subsequent L-1 time domain range is still considered, continuously rolling forward to achieve dynamic feature extraction across the entire dataset.

[0038] 2. This invention constructs an SVM classification model based on four types of predictive controller performance degradation depth features extracted by a deep residual network, and then constructs a DRN-SVM performance degradation diagnostic model. Given the introduction of a deep residual network into the SVM classification model, the objective function of the traditional SVM is no longer applicable. Based on the principle of the DRN-SVM performance degradation diagnostic model, the objective function of this model is constructed. The data feature extraction effect and data representation ability of the deep residual network are superior to the implicit nonlinear transformation in the support vector machine, while the classification effect of the support vector machine using the maximum data margin is superior to the classification effect of the softmax layer in the deep residual network. The DRN-SVM performance degradation diagnostic model combines the advantages of both, therefore, the classification accuracy of this model is superior to that of the deep residual network and the support vector machine. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of data feature extraction based on DRN-SVM;

[0040] Figure 2 This is a framework diagram of the residual module in a Deep Residual Network (DRN);

[0041] Figure 3 This is a framework diagram of a Deep Residual Network (DRN).

[0042] Figure 4 A framework diagram of a predictive control performance degradation diagnosis method based on DRN-SVM;

[0043] Figure 5 This is a schematic diagram illustrating the distance of depth features in the feature space.

[0044] Figure 6 This is a diagnostic process for predictive control performance degradation based on DRN-SVM. Detailed Implementation

[0045] The following detailed description, in conjunction with the accompanying drawings and specific implementation methods, provides a further detailed explanation of the present invention's method for diagnosing the root causes of performance degradation in DRN-SVM predictive control based on rolling characteristic time domain.

[0046] To address the problem that the original SVM algorithm does not perform well in classifying complex industrial process data, this application provides an improved DRN-SVM algorithm that captures the dynamic temporal characteristics of complex industrial process data while fully mining the data features of industrial process data, thereby improving the accuracy of predictive controller performance degradation root cause diagnosis.

[0047] This invention obtains process datasets under four performance modes (process model mismatch, disturbance characteristic change, controller constraint saturation, and improper controller parameter adjustment) of a predictive control system based on a Wood-Berry predictive control simulation model. These process datasets under the four performance modes are used as the training set. Drawing inspiration from the "rolling optimization" approach in predictive control algorithms, the concept of a "rolling feature time domain" is proposed to characterize the temporal and dynamic nature of the predictive control system's operating data. A deep residual network (DRN) is used to extract data features within a rolling feature time domain of width L. Based on this rolling feature time domain, the model continuously rolls forward with a step size h, thereby achieving dynamic feature extraction across the entire dataset. A DRN-SVM classification model is constructed based on the process data features extracted from the four performance modes of the predictive control system using the deep residual network. An improved DRN-SVM predictive control performance degradation diagnostic model is trained. The process data of predictive controller performance degradation obtained from the Wood-Berry predictive control simulation model (different from the training set data) is used as the test set and fed into the DRN-SVM performance degradation diagnostic model to determine the factors leading to predictive controller performance degradation.

[0048] Specifically, the performance degradation modes of the predictive controller are derived from its historical dataset {y} 1i y 2i , ...,yn i ;u 1iu 2i ,…,u mi} (i = 1, 2, ..., N) characterizes the predictive controller's performance modes c as follows:

[0049] c = f(y) 1i y 2i , ..., y n i;uu 1i u 2i , ..., u mi )

[0050] Where f represents the correlation function in the DRN network, y and u represent the controlled variable and manipulated variable, respectively, i represents the i-th sampling time, n represents the number of controlled variables in the predictive controller, m represents the number of manipulated variables, N represents the length of the evaluation data, and y1 i y2 represents the value of the first controlled variable in the predictive control system at the i-th sampling time. i Let y represent the value of the second controlled variable in the predictive control system at the i-th sampling time. ni u1 represents the value of the nth controlled variable in the predictive control system at the i-th sampling time. i u2 represents the value of the first manipulated variable in the predictive control system at the i-th sampling time. i u represents the value of the second manipulated variable in the predictive control system at the i-th sampling time. mi This represents the value of the m-th manipulated variable in the predictive control system at the i-th sampling time.

[0051] Step 2: Based on the rolling feature temporal domain, continuously roll forward with a certain step size, and then achieve dynamic feature extraction of the entire training set range through a deep residual network;

[0052] The operational data of predictive control systems exhibit both temporal and dynamic characteristics. Drawing inspiration from the "rolling optimization" approach in predictive control algorithms, the concept of a "rolling characteristic time domain" is proposed to characterize the temporal and dynamic nature of the predictive control system's operational data. (See attached image) Figure 1 Assuming the rolling feature has a time domain width of L and a rolling step size of h, feature extraction at time k considers not only the process variable characteristics at the current time but also the dynamic changes within the subsequent k+L-1 time range, thus reflecting the performance characteristics of the control system within the width L. A deep residual network is used to extract data features within the rolling feature's time domain of width L. Deep residual networks possess the characteristic of extracting deep, complex, and nonlinear features, thus accurately capturing the complex dynamic temporal performance within the width L. As the rolling feature's time domain continuously advances with a step size h, dynamic feature extraction across the entire dataset can be achieved.

[0053] Secondly, deep residual networks utilize multiple sliding weight windows to traverse process data within a rolling feature time domain of width L, thus enabling them to uncover deeper features of complex industrial process data.

[0054] In summary, a deep residual network is preferred for feature extraction, and the execution steps are as follows:

[0055] First, the historical dataset consisting of the controlled variable y and the manipulated variable u is constructed into a two-dimensional matrix X;

[0056] Then, the constructed two-dimensional data matrix X is fed as input into the deep residual network;

[0057] Next, before performing convolution feature extraction, the two-dimensional data matrix X needs to be padded at the matrix edges to prevent missing features of the matrix edge data.

[0058] Next, weights and bias learning parameters γβ are introduced through BN, with the aim of restoring the expressive power of the data itself;

[0059] Below, the weight window in the deep residual network is used to process the two-dimensional data matrix BN after BN processing. γ，β (X) performs feature extraction;

[0060] Then, the deep features extracted from multiple weighted windows are subjected to max pooling to select the deep features with the largest weight values.

[0061] Finally, the deep features selected through max pooling are transposed, and the result is used as the input to the support vector machine.

[0062] The proposed feature extraction method using deep residual networks for complex industrial process data outperforms the original single-layer implicit nonlinear mapping feature extraction method. Furthermore, the latter does not take into account the dynamic temporal nature of industrial data, resulting in discrete data features compared to the proposed method.

[0063] This application, in another aspect, provides a method for diagnosing the root causes of predictive control performance degradation based on DRN-SVM. This method integrates deep residual networks and support vector machines, proposing an improved DRN-SVM performance degradation root cause diagnosis model. It fully utilizes the superior representation ability of deep residual networks for complex industrial data and the advantage of support vector machines in classifying data by maximizing the data interval. The improved method includes:

[0064] The SVM-based method for diagnosing the root causes of predictive control performance degradation mainly consists of three parts: data input, feature extraction via implicit nonlinear mapping, and the establishment of an SVM classification model. Implicit nonlinear mapping maps nonlinear data to a high-dimensional feature space to achieve linear separability. However, nonlinear mapping only extracts shallow features from industrial process data and cannot represent complex industrial process data.

[0065] We utilize deep residual networks to mine deep features from industrial process data, and then construct a support vector machine (SVM) classification model based on the extracted deep features.

[0066] The two-dimensional data matrix to be classified needs to be padded at its edges to avoid missing edge data features. The padded data matrix then enters the residual module of the deep residual network for weight calculation, i.e., feature extraction. Because the deep residual network has a skip-layer connection structure, the data features obtained after traversing the weight window are concatenated with the original input data features after adjusting the dimensions using weights W, and then used as the output of the entire residual module to enter the next residual module. (See attached image) Figure 2 This refers to the residual module in a deep residual network. The batch normalization (BN) strategy in the residual block can solve the ICS (Internal Covariate Shift) problem that often occurs in deep networks. Simultaneously, the unique skip-layer connection structure in the residual module makes it easier for the network to learn data features, avoiding gradient vanishing or exploding phenomena during network training and reducing the number of training parameters. (See attached image) Figure 3 A deep residual network consists of multiple residual modules. The output H(x) of a residual module is composed of two parts: the input x and the residual F(x), from which the following calculation relationship can be obtained.

[0067] H(x)=F(x)+Wx

[0068] Where W represents the weights of the input x dimension.

[0069] There are four main factors affecting the performance degradation of predictive controllers: process model mismatch, changes in disturbance characteristics, controller constraint saturation, and improper controller parameter adjustment. A two-dimensional data matrix X is constructed based on process data under these four performance degradation modes of the predictive controller.

[0070]

[0071]

[0072] Where i represents the i-th type of performance degradation mode of the predictive control system. This represents process data under the first performance degradation mode of the predictive control system. This represents process data under the second performance degradation mode of the predictive control system. This represents process data under the third performance degradation mode of the predictive control system. y represents the process data under the fourth performance degradation mode of the predictive control system. n，T u represents the value of the nth controlled variable in the predictive control system at the Tth sampling time. m，T Let represent the value of the m-th manipulated variable in the predictive control system at the T-th sampling time, where n represents the number of controlled variables in the predictive controller, m represents the number of manipulated variables, T represents the sampling time, and the subscripts p and q represent indices in the dataset. p represents the row index of the process data matrix under the i-th type of predictive controller performance degradation mode, and q represents the column index of the process data matrix under the i-th type of predictive controller performance degradation mode, 1≤p≤(n+m), 1≤q≤T.

[0073] The Internal Covariate Shift (ICS) problem, a frequent occurrence in deep networks, is typically addressed using whitening. However, whitening alters the distribution of network layers, thus changing the expressive power of the data within those layers, and also leads to the loss of learned parameter information from lower layers. Therefore, Batch Normalization (BN) is introduced into deep residual networks. The BN algorithm steps are as follows:

[0074] Define the mean of X as μ X Its expression is as follows:

[0075]

[0076] x p，q This represents the elements of a two-dimensional data matrix X constructed from process data under four performance degradation modes of the predictive controller. The construction form of the process data under each performance degradation mode of the predictive controller is as follows: Based on the concatenation form of the two-dimensional data matrix X, its dimension is m+n*4T, therefore the maximum value of q in X is 4T.

[0077] Define the variance of X as Its expression is as follows:

[0078]

[0079] Based on the mean and variance of X, its standardized form as follows,

[0080]

[0081] The purpose of the standardization parameter ε is to ensure the stability of the numerical value and avoid invalid expressions in the standardization expression.

[0082] A linear transformation of the dataset X after Batch Normalization (BN) yields the following expression:

[0083]

[0084] Here, γ and β are the weights introduced by BN, which are bias learning parameters aimed at restoring the expressive power of the data itself. This represents the form of the dataset X after a linear transformation following BN; Represents the standardized form of dataset X; BN γ，β (X) represents the linear transformation after BN. The abbreviation of .

[0085] As attached Figure 4 Deep residual networks (DRN-SVMs) replace the implicit nonlinear transformations in traditional SVM algorithms for feature extraction from complex industrial process data. A DRN-SVM performance degradation diagnostic model is constructed based on the deep features extracted by the network, and the objective optimization function of the DRN-SVM model is defined. By training the hyperparameters in the deep residual network, the optimal classification hyperplane in the feature space is found, and thus the classification decision function is obtained. (Based on the appendix...) Figure 4 The DRN-SVM method framework in the DRN can be used to obtain the output of the first residual module in the Deep Residual Network (DRN), which is H1(X), and its expression is as follows:

[0086]

[0087]

[0088] Similarly, the output of the second residual module is H2(H1(X)), and its expression is as follows:

[0089]

[0090] The output of the nth residual module in a Deep Residual Network (DRN) is H. n (H n-1 (…H1(X)…)), its expression is as follows:

[0091]

[0092] Where, φ (l) (·)(1≤l≤2n) represents the convolution in the residual module of a deep residual network; W (l)(1≤l≤2n) represents the weights of the convolutional layers in the residual module; g (l) (1≤l≤2n) represents the activation function in the residual module; (γ) τ ,β τ (1≤τ≤2n) represents the weights introduced by the Batch Normalization (BN) policy in a deep residual network, and the bias learning parameters; W (k) (1≤k≤n) represents adjusting the input X, H1(X), ..., H of the deep residual network. n The weights of the (...(H1(X))...) dimension.

[0093] The dataset X, consisting of process data under four performance degradation modes of the predictive controller, is processed through n residual modules to extract data features. The extracted data features are then passed through a max pooling layer and a fully connected layer to obtain the final output H of the deep residual network. n (x), its expression is as follows.

[0094]

[0095] In this study, the number of data features extracted by the deep residual network for the process data under the four performance modes of the predictive controller is N. The data features extracted from the process dataset X by the deep residual network can be simplified to BN. 1，a (g(φ(X;W));(γ,β)), (1≤a≤4N), where a represents the index of the data feature vector extracted by the deep residual network.

[0096] W = {W (1) W (2) , ..., W (2n) W (1) W (2) , ..., W (n) Let} be the weight set of the deep residual network, (γ, β) = {(γ1, β1), (γ2, β2), ..., (γ...} 2n ,β 2n The weights of the deep residual network are introduced into the BN strategy, and the bias learning parameter set is used.

[0097] Since the improved SVM classification model is built upon features extracted by a deep residual network, the objective function for traditional SVM-based root cause diagnosis of performance degradation is no longer applicable. The optimization objective of the DRN-SVM model is to train the parameter set W, (γ, β) of the deep residual network so that the features H of the process data extracted by the network... n (X) is distributed as far away from the optimal hyperplane as possible (ω). ij ) * x+(b ij ) *=0 on both sides. To address the above situation, an improved objective optimization function for a predictive control performance degradation root cause diagnosis method based on DRN-SVM is proposed, and its expression is as follows:

[0098]

[0099] in, This indicates that the objective function of DRN-SVM is to be minimized, ω ij X represents the normal vector of the hyperplane to be used to determine the optimal classification hyperplane for the performance degradation data of the i-th and j-th classes of the predictive controller, and C represents the penalty parameter; i and X j W represents the dataset of performance degradation for the i-th and j-th classes of the predictive controller; (l) W represents the weights of the convolutional layers in a deep residual network. (k) This represents adjusting the inputs X, H1(X), ..., H of the deep residual network. n (...(H1(X))...) represents the weights of the dimension; |||| represents the magnitude of the vector. Represents the square of the modulus. Denotes the square of the Frobenius norm;

[0100] BN 1，a (g(φ(X i ;W));(γ,β)) represents the data features of the i-th type of performance degradation data of the predictive controller extracted by the deep residual network, BN 1，a (g(φ(X j ;W));(γ,β)) represents the data features of the j-th type of performance degradation data of the predictive controller extracted by the deep residual network.

[0101] Define parameter D as follows: Figure 5 It represents the square of the distance between the features of the process data in the i-th and j-th performance mode states of the predictive controller extracted by the deep residual network in the feature space and the sum of the squares of the distances between the features ... in the feature space. The difference of squares is expressed as follows:

[0102]

[0103] According to the objective function of the improved DRN-SVM performance degradation model, the second term of the objective function aims to maximize the distance between the deep features extracted by the deep residual network in the feature space. λ1 and λ2 correspond to the two regularization terms of the network, representing the network's penalty coefficients, and λ1 represents the network's convolution weight W. (l) The penalty coefficient, λ2 represents the weight W adjusted for the network input dimension. (k)The penalty coefficient is calculated as follows: The parameter C adjusts the distribution of features extracted by the deep residual network on both sides of the classification hyperplane, ensuring the features are as far away from the hyperplane as possible. If D > 0, it indicates that the process data features extracted by the deep residual network are not as far away from the optimal classification hyperplane as possible, and a larger penalty parameter C should be applied. If D < 0, it indicates that the extracted process data features are as far away from the optimal classification hyperplane as possible, and a smaller penalty parameter C should be applied. The rules for applying the penalty term C are as follows:

[0104]

[0105] ↑ indicates increasing the penalty parameter C, ↓ indicates decreasing the penalty parameter C, based on the deep optimization algorithm for the weight set W = {W (1) W (2) , ..., W (2n) W (1) W (2) , ..., W (n)} and the weights introduced by the BN policy, the bias learning parameter set (γβ)={(γ1,β1), (γ2,β2), ..., (γ 2n ,β 2n )} is optimized to extract the process data features of the i-th and j-th performance degradation of the predictive controller, i.e., BN 1，a (g(φ(X i ;W));(γ,β)),BN 1，a (g(φ(X j The goal is to maximize the distance between the two classes of data in the feature space, thereby finding the optimal classification hyperplane (ω) between them. ij ) * x+(b ij ) * =0, thus achieving the optimal classification effect. W * and (γ) * ,β * ) represent the set of weights of the deep residual network obtained by optimization training and the set of weights and bias learning parameters introduced by the BN policy, respectively.

[0106] As attached Figure 6 This is a flowchart of the diagnostic process for predictive control performance degradation diagnosis algorithm based on DRN-SVM.

[0107] This invention focuses on the root cause diagnosis of performance degradation in predictive control systems. The core of this invention is the concept of a "rolling feature time domain," which characterizes the temporal and dynamic nature of predictive control system operating data. A deep residual network (DRN) is used to capture the dynamic temporal features of the data within this feature time domain, and multiple sliding weight windows based on the DRN can fully extract the data's features. As the rolling feature time domain progresses forward, feature extraction of the entire dataset is achieved. Secondly, leveraging the maximum data interval classification characteristic of SVM, the features extracted by the DRN are used as input to construct a DRN-SVM classification model, fully combining the advantages of both algorithms. The aim is to fully extract the features of the predictive control system operating data, thereby improving the accuracy of root cause diagnosis of predictive control system performance degradation.

[0108] The detailed steps of this invention are as follows:

[0109] The process datasets for four performance modes of the predictive control system were obtained based on the Wood-Berry distillation simulation model. The simulation time for each performance mode was 500 min and the sampling time was 1 min.

[0110] Given the amount of process data under each performance mode, the predictive control system uses four different values ​​for each of the four performance modes for simulation, thereby expanding the dataset and using it as the training set.

[0111] The training set data was preprocessed and labeled using an Excel spreadsheet.

[0112] The preprocessed training dataset is fed into the improved DRN-SVM performance degradation diagnostic model for training.

[0113] The experimental environment uses the Windows 10 operating system and the PyTorch framework, implemented using Python 3.8.

[0114] The performance of the improved DRN-SVM performance degradation diagnostic model was evaluated using accuracy.

[0115] The accuracy metrics are as follows:

[0116]

[0117] TP and FP represent the number of samples correctly identified as true and the number of samples incorrectly identified as true, respectively.

[0118] Matters not covered in this invention are common knowledge.

[0119] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for diagnosing the root causes of performance degradation in DRN-SVM predictive control based on rolling feature time domain, characterized in that, Includes the following steps: Step 1: Obtain historical data under the performance degradation mode of the predictive controller using the Wood-Berry distillation column simulation model as the training set; Step 2: Based on the rolling feature time domain, continuously roll forward with a certain step size, and then realize the dynamic feature extraction of the entire training set through a deep residual network; Step 3: The dynamic features of the predictive control system under the performance degradation mode extracted by the deep residual network are used as the input of the SVM to construct the DRN-SVM classification model. Step 4: Train the DRN-SVM classification model based on the training set to predict the control performance degradation diagnosis model; Step 5: Input the predictive controller performance degradation test set obtained based on the Wood-Berry model into the DRN-SVM control performance degradation diagnostic model to determine the factors that cause predictive controller performance degradation. Step 4 trains the parameter set W,(γ,β) of the deep residual network. W is the weight set of the deep residual network, and (γ,β) is the set of weights and biases learned by the BN strategy in the deep residual network, so that the features of the process data extracted by the network are distributed as far away from the optimal hyperplane as possible. The objective function of the DRN-SVM classification model is: in, This indicates that the objective function of DRN-SVM is to be minimized, ω ij Let X represent the normal vector of the optimal classification hyperplane for the performance degradation data of the predictive controller in classes i and j. Let C represent the penalty parameter, used to adjust the distribution of features extracted by the deep residual network on both sides of the classification hyperplane, so that the features are as far away from the classification hyperplane as possible. Let a represent the index of the data feature vector extracted by the deep residual network. N is the number of data features of the process data under the performance mode state of the predictive controller extracted by the deep residual network. i and X j This represents the datasets showing performance degradation in classes i and j of the predictive controller; BN 1,a (g(φ(X i ;W));(γ,β)) represents the data features of the i-th type of performance degradation data of the predictive controller extracted by the deep residual network, BN 1,a (g(φ(X j ;W));(γ,β)) represents the data features of the j-th type of performance degradation data of the predictive controller extracted by the deep residual network; W (l) W represents the weights of the convolutional layers in a deep residual network. (k) λ1 and λ2 represent the weights of the input dimension of the deep residual network; λ1 and λ2 correspond to the two regularization terms of the network, where λ1 represents the network convolution weights W. (l) The penalty coefficient, λ², represents the weight W adjusted for the network input dimension. (k) The penalty coefficient; ||| represents the magnitude of the vector, ||| 2 Represents the square of the modulus. This represents the square of the Frobenius norm.

2. The method for diagnosing the root causes of performance degradation in DRN-SVM predictive control based on rolling feature time domain as described in claim 1, characterized in that: In step 1, the controller performance degradation mode includes four states: process model mismatch, disturbance characteristic change, controller constraint saturation, and improper controller parameter adjustment.

3. The method for diagnosing the root causes of DRN-SVM predictive control performance degradation based on rolling feature time domain as described in claim 2, characterized in that: Step 2 uses a deep residual network for feature extraction, and utilizes multiple sliding weight windows to traverse the process data within the rolling feature time domain, which facilitates the discovery of deep features in complex industrial process data.

4. The method for diagnosing the root causes of performance degradation in DRN-SVM predictive control based on rolling feature time domain as described in claim 3, characterized in that, Feature extraction using deep residual networks specifically includes the following steps: Step 2.1: Construct a two-dimensional matrix X from the historical dataset consisting of the controlled variable y and the manipulated variable u; Step 2.2: Input the constructed two-dimensional data matrix X into the deep residual network; Step 2.3: Before performing convolution feature extraction, the two-dimensional data matrix X needs to be padded at the matrix edges to prevent missing features of the matrix edge data; Step 2.4: Extract features from the two-dimensional data matrix using the weight window in the deep residual network; Step 2.5: Perform max pooling on the deep features extracted from multiple weight windows to select the deep features with the largest weight values; Step 2.6: Transpose the deep features selected by the max pooling operation, and use the result as the input for the subsequent SVM.

5. The method for diagnosing the root causes of DRN-SVM predictive control performance degradation based on rolling feature time domain as described in claim 4, characterized in that: In step 2.4, a Batch Normalization (BN) strategy is introduced into the deep residual network to address the issues that the whitening process alters the distribution of network layers, thereby changing the expressive power of the data within the network layers, and that the parameter information learned by the underlying network is lost during the whitening process.

6. The method for diagnosing the root causes of performance degradation in DRN-SVM predictive control based on rolling feature time domain as described in claim 1, characterized in that, The rules for applying the penalty parameter C are as follows: Where ↑ represents increasing the penalty parameter C, ↓ represents decreasing the penalty parameter C, and D represents the square of the distance between the features of the process data in the feature space of the predictive controller under the i-th and j-th performance modes extracted by the deep residual network. The difference of squares is expressed as follows: If D > 0, it means that the process data features extracted by the deep residual network in the feature space are not as far away from the optimal classification hyperplane as possible, and a larger penalty parameter C should be applied; if D < 0, it means that the extracted process data features are as far away from the optimal classification hyperplane as possible, and a smaller penalty parameter C should be applied.

7. The method for diagnosing the root causes of DRN-SVM predictive control performance degradation based on rolling feature time domain as described in claim 6, characterized in that: The performance of the improved DRN-SVM performance degradation diagnostic model was evaluated using accuracy.

8. The method for diagnosing the root causes of performance degradation in DRN-SVM predictive control based on rolling feature time domain as described in claim 7, characterized in that, The accuracy metric formula is as follows: Here, Accuracy represents the accuracy rate, and TP and FP represent the number of samples that were correctly identified as true and the number of samples that were incorrectly identified as true, respectively.

Images