Industrial process fault diagnosis method, device, equipment and medium

Abstract

The invention relates to the technical field of process fault diagnosis. The invention discloses an industrial process fault diagnosis method and device, equipment and a medium, and the method comprises the steps: obtaining historical operation data, and dividing the historical operation data into a training set and a test set; a double-constraint model is constructed, the double-constraint model is trained according to data in the training set, and constraint conditions of the double-constraint model comprise hard sparse neighborhood constraint and manifold structure regularization keeping constraint; testing the trained double-constraint model according to data in the test set; and carrying out industrial process fault diagnosis based on the tested double-constraint model. According to the method, extremely strict constraint conditions are constructed through hard sparse neighborhood constraint and manifold structure keeping regularization constraint in a double constraint model, the problem of small fault detection in a complex industrial process is solved, and high sensitivity and low false alarm rate in fault detection are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of process fault diagnosis, in particular to an industrial process fault diagnosis method, device, equipment and medium. BACKGROUND

[0002] With the modern industrial process becoming more and more complex, large-scale and automated, the safe and stable operation of production systems is facing severe challenges. Once a fault occurs in continuous production processes such as chemical industry, power, metallurgy and pharmaceutical industry, it will often cause serious economic losses, environmental pollution and even casualties. Therefore, establishing an efficient and accurate industrial process monitoring and fault diagnosis system has become a key technical requirement to ensure production safety and improve economic efficiency. Although the traditional method based on physical model has clear interpretability, it is often difficult to establish an accurate mathematical model when facing high-dimensional, strongly coupled and nonlinear complex industrial processes, and it has poor adaptability to process changes.

[0003] In recent years, data-driven process monitoring methods have gradually become a research hotspot in the field of industrial process monitoring because they do not rely on accurate mechanism models and can directly learn the normal and abnormal operation patterns of the process from massive historical data. Such methods extract statistical features of process data, construct monitoring statistics, and detect whether the process deviates from the normal condition in real time. Among them, multivariate statistical process monitoring methods such as principal component analysis (PCA), partial least squares (PLS), independent component analysis (ICA) have been widely applied. However, when dealing with complex industrial process data with nonlinear and multimodal characteristics, these traditional methods often assume that the data follows a Gaussian distribution or a linear relationship, and it is difficult to fully exploit the internal structural information in the data, resulting in insufficient fault detection sensitivity and high false alarm rate.

[0004] Therefore, in industrial field process control, how to ensure high sensitivity and low false alarm rate in fault detection is a technical problem to be solved at present. SUMMARY

[0005] In view of the above problems, the embodiments of the present application provide an industrial process fault diagnosis method, device, equipment and medium, which aims to solve the above problems or at least partially solve the above problems.

[0006] In a first aspect, the present application provides an industrial process fault diagnosis method, comprising: obtaining historical operation data, and dividing the historical operation data into a training set and a test set; building a double-constraint model, training the double-constraint model according to the data in the training set, and the constraint conditions of the double-constraint model including a hard sparse neighborhood constraint and a manifold structure preserving regularization constraint; testing the trained double-constraint model according to the data in the test set; Industrial process fault diagnosis based on the dual-constraint model after testing.

[0007] The historical operation data includes normal operation data and fault operation data. The data in the training set includes normal operation data. The data in the testing set includes normal operation data and fault operation data.

[0008] The objective function of the dual-constraint model is constructed, including: The affinity matrix is constructed according to the data in the training set, and the Laplacian matrix is calculated according to the affinity matrix; The objective function is constructed according to the Laplacian matrix based on the pre-set basis matrix and coefficient matrix, and the objective function includes the data reconstruction error term and the manifold regularization term.

[0009] The dual-constraint model is trained according to the data in the training set, including: The objective function in the dual-constraint model is solved based on the pre-set solving algorithm to determine the target basis matrix; The data in the training set is projected to the target basis matrix to obtain the reconstructed data; The statistics are constructed according to the data in the training set and the reconstructed data; The control limit is determined according to the statistics based on the pre-set threshold estimation algorithm.

[0010] The construction method of the hard sparse neighborhood constraint includes: Based on the pre-set number of centroids, a plurality of centroids are obtained by clustering the data in the training set to obtain a reference matrix; Based on the pre-set number of neighborhoods, the distance between each data in the training set and each centroid is calculated, and the plurality of distances are sorted to filter the target centroid corresponding to each data in the training set; According to each data in the training set and the target centroid corresponding thereto, a mask matrix is obtained, each row of the mask matrix corresponds to a centroid, each column corresponds to a data in the training set, and the position of the target centroid corresponding to each data in the training set is set to 1, and the other positions are set to 0; Based on the mask matrix, the sample coefficient matrix is calculated according to each data in the training set and the target centroid corresponding thereto; The hard sparse neighborhood constraint is constructed according to the mask matrix and the sample coefficient matrix.

[0011] The dual-constraint model is tested according to the data in the testing set, including: The data in the testing set is normalized, and the maximum and minimum values of the normalization are obtained according to the data in the training set. Based on the data in the normalized test set and the target basis matrix in the trained dual-constraint model, the reconstructed data of the data in the normalized test set is obtained; Calculate the statistics corresponding to the data in the test set based on the normalized test set data and its reconstructed data. The statistics corresponding to the data in the test set are compared with the control limits in the dual-constraint model, and fault diagnosis is performed based on the comparison results.

[0012] For example, a fault is determined to have occurred when the statistic corresponding to the data in the test set is greater than the control limit in the dual-constraint model; The system is considered to be running normally when the statistics corresponding to the data in the test set are less than or equal to the control limits in the dual-constraint model.

[0013] Secondly, this application provides an industrial process fault diagnosis device, comprising: The acquisition module is used to acquire historical running data and divide the historical running data into training set and test set; The training module is used to build a dual-constraint model. It trains the dual-constraint model based on the data in the training set. The constraints of the dual-constraint model include hard sparse neighborhood constraints and manifold structure preservation regularization constraints. The testing module is used to test the trained dual-constraint model based on data in the test set. The diagnostic module is used for industrial process fault diagnosis based on the tested dual-constraint model.

[0014] Thirdly, this application provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the industrial process fault diagnosis method of the first aspect.

[0015] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the industrial process fault diagnosis method of the first aspect.

[0016] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: This application constructs extremely strict constraint conditions by using hard sparse neighborhood constraints and manifold structure regularization constraints in the dual constraint model, which solves the problem of detecting minute faults in complex industrial processes and ensures high sensitivity and low false alarm rate in fault detection. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of an application environment for an industrial process fault diagnosis method according to an embodiment of the present invention; Figure 2 This is a schematic flowchart of an industrial process fault diagnosis method according to an embodiment of the present invention; Figure 3 yes Figure 2 A flowchart illustrating a specific implementation method of step S2; Figure 4 This is a schematic diagram of the fault detection results of the industrial process fault diagnosis method d01 in one embodiment of the present invention; Figure 5 This is a schematic diagram of the fault detection results of the industrial process fault diagnosis method d02 in one embodiment of the present invention; Figure 6 This is a schematic diagram of the fault detection results of the industrial process fault diagnosis method d04 in one embodiment of the present invention; Figure 7 This is a schematic diagram of the fault detection results of the industrial process fault diagnosis method d06 in one embodiment of the present invention; Figure 8 This is a schematic diagram of the fault detection results of the industrial process fault diagnosis method d07 in one embodiment of the present invention; Figure 9 This is a schematic diagram of the fault detection results of the industrial process fault diagnosis method d08 in one embodiment of the present invention; Figure 10 This is a schematic diagram of the fault detection results of the industrial process fault diagnosis method d12 in one embodiment of the present invention; Figure 11 This is a schematic diagram of the fault detection results of the industrial process fault diagnosis method d13 in one embodiment of the present invention; Figure 12 This is a schematic diagram of the fault detection results of the industrial process fault diagnosis method d14 in one embodiment of the present invention; Figure 13 This is a schematic diagram of the fault detection results of the industrial process fault diagnosis method d17 in one embodiment of the present invention; Figure 14 This is a schematic diagram of the fault detection results of the industrial process fault diagnosis method d18 in one embodiment of the present invention; Figure 15 This is a schematic diagram of an industrial process fault diagnosis device according to an embodiment of the present invention; Figure 16 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 17 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such use can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the term "comprising" and its variations should be interpreted as open-ended terms meaning "including but not limited to."

[0020] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0021] As mentioned above, current industrial process monitoring methods suffer from insufficient fault detection sensitivity and a high false alarm rate. To address this technical problem, embodiments of this application provide an industrial process fault diagnosis method.

[0022] The industrial process fault diagnosis method provided in this invention can be applied to, for example, Figure 1 In this application environment, the device communicates with the server via a network. The server can obtain operational data from the device, divide the data into training and testing sets, construct a dual-constraint model, and train the model using data from the training set. The constraints of the dual-constraint model include hard sparse neighborhood constraints and manifold structure preservation regularization constraints. The trained dual-constraint model is then tested using data from the testing set. Finally, industrial process fault diagnosis is performed based on the tested dual-constraint model. This application constructs extremely stringent constraints through hard sparse neighborhood constraints and manifold structure preservation regularization constraints in the dual-constraint model, solving the problem of detecting minute faults in complex industrial processes and ensuring high sensitivity and low false alarm rate in fault detection.

[0023] The device side can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server side can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will now be described in detail through specific embodiments.

[0024] Please see Figure 2 As shown, Figure 2 A flowchart illustrating an industrial process fault diagnosis method provided in an embodiment of the present invention includes the following steps: S1: Obtain historical running data and divide the historical running data into training set and test set.

[0025] In one embodiment, this step is fundamental to building a high-quality model. Because the Non-negative Matrix Factorization (NMF) algorithm has strict requirements on the numerical range of the data (non-negativity), and industrial data often contains noise or dimensional differences, historical operating data must be standardized.

[0026] In one embodiment, the dataset is partitioned as follows: the collected historical process operation data is divided into a training set and a test set. The data in the training set includes normal operation data; the data in the test set includes both normal operation data and fault operation data. The training set must contain only normal operation data under normal operating conditions to learn the system's health benchmark; the test set should contain both normal operation data and fault operation data to verify the detection capabilities of the trained model.

[0027] In one embodiment, data normalization is performed to satisfy the non-negativity constraint of Non-negative Matrix Factorization (NMF). To eliminate the influence of different variable units (e.g., temperature and pressure have different units), maximum-min normalization is typically used. Let... For the first The first sample The normalization formula for the variables is:

[0028] The processed data in the training set is organized into a training set data matrix. ,in It is the dimension of the process variable. It refers to the number of samples.

[0029] It is important to note that during the testing phase, the statistics obtained during the training phase must be used. and Normalization should be performed, and the use of information from the test set itself is strictly prohibited to avoid data leakage.

[0030] Specifically, the Tennessee Eastman (TE) chemical process benchmark platform was used for verification. The TE process is a realistic industrial process simulation model, including core units such as reactors, separators, stripping towers, and compressors, involving 12 manipulated variables and 41 measured variables. In this embodiment, 52 continuously observed variables were selected as monitoring variables. Historical operating data were collected from the TE process simulation platform. Simulation data under "normal operating conditions" (usually labeled d00) was selected as the training set, with a total sample size of 960, to establish a normal behavior model of the process and determine control limits.

[0031] Specifically, data containing different types of faults were selected as the test set to verify the fault detection performance of the model. During data normalization, considering the significant differences in the physical dimensions and numerical ranges of the variables (such as temperature, pressure, and flow rate) in the TE process, and to meet the non-negativity requirement of the non-negative matrix factorization algorithm for the input data, the Min-Max Scaling method was used to process the data, as shown in formula (1). Furthermore, to verify the detection performance of the dual-constraint model under different abnormal operating conditions, this application used 21 sets of fault test datasets (corresponding to faults IDV1 to IDV21). Each test set also contained 52 monitoring variables (41 measurement variables and 11 operational variables), with a total of 960 samples. In each set of test data, the first 160 samples were in normal operating condition, and a specific type of fault was introduced starting from the 161st sample and continued until the end of the simulation. This setting was used to simulate the dynamic migration of the process from normal operating conditions to fault conditions to evaluate the model's fault detection rate (FDR) and false alarm rate (FAR).

[0032] S2: Construct a dual-constraint model. Train the dual-constraint model based on the data in the training set. The constraints of the dual-constraint model include hard sparse neighborhood constraints and manifold structure preservation regularization constraints.

[0033] In one embodiment, such as Figure 3 As shown, the objective function for constructing the dual-constraint model in step S2 includes: S21: Construct an affinity matrix based on the data in the training set, and calculate the Laplace matrix based on the affinity matrix.

[0034] In one embodiment, this is the core algorithm implementation phase. The goal is to decompose the training set data matrix. Learn the basis matrix that can represent the health status of the system. At the same time, it preserves the local manifold structure of the data.

[0035] In one embodiment, the K-Nearest Neighbors (KNN) algorithm is used to compute the training set data matrix. Affinity matrix And based on the affinity matrix Calculate the degree matrix and Laplace matrix This step is to capture the local geometric structure of the data in high-dimensional space.

[0036] Specifically, when applied to the TE process, in order to capture the nonlinear manifold structure of the TE process data in the high-dimensional state space and provide a topological basis for subsequent manifold regularization, a nearest neighbor graph is first constructed. Using the K-Nearest Neighbors (KNN) algorithm, for each normal sample in the training set... Find the one with the closest Euclidean distance. One neighbor (in this embodiment) Constructing an affinity matrix based on neighborhood relationships. If the sample and If they are close neighbors, then (Or use Gaussian kernel weights), otherwise 0. Subsequently, based on the constructed affinity matrix... Calculate the degree matrix degree matrix Affinity matrix A diagonal matrix, diagonal elements Equal to affinity matrix No. i The sum of all elements in a row. Laplace matrix. L according to Calculate the matrix. It contains local topology information under normal operating conditions.

[0037] S22: Based on the pre-set basis matrix and coefficient matrix, construct the objective function according to the Laplacian matrix. The objective function includes a data reconstruction error term and a manifold regularization term.

[0038] In one embodiment, the dual constraint model is a dual constraint model based on sparse locality and manifold regularization, which can be expressed as MR-HTCNMF (Manifold Regularized Hard-Threshold Constrained Non-negative Matrix Factorization).

[0039] In one embodiment, combining the baseline framework of nonnegative matrix factorization (NMF), and introducing both a "hard sparsity constraint" describing the local affiliation of samples and a "manifold regularization" objective function describing the topological relationships between samples, the loss is minimized as follows:

[0040] In this formula, express F The square of the norm; It represents the Hadamardi (or Hadama) stack; Represents the trace operator; Represents the training set data matrix; L Represents the Laplace matrix; Represents the basis matrix; Represents the sparse coefficient matrix (derived from the sample coefficient matrix) and mask matrix (Composition). Data reconstruction error term: This term measures the basis matrix. With sparse coefficient matrix For the training set data matrix The degree of approximation is ensured to guarantee that the model can learn the main features under normal operating conditions.

[0041] This application constructs a mathematical framework that deeply integrates "geometric structure perception" and "hard threshold sparse representation", aiming to solve the problem of detecting minute faults in complex industrial processes through extremely strict constraints.

[0042] Specifically, constraint one: hard sparse neighborhood constraint ( ).in It is a pre-calculated binary mask matrix. After obtaining the centroids by clustering normal operating data samples, the nearest centroid is selected for each sample. The system generates a matrix of basis vectors (centroids) by setting 1 to the corresponding position and 0 to the remaining positions. This constraint forces the representation of each sample to have a sparse coefficient matrix. Only in the context of the nearest The positions of the basis vectors (centroids) are non-zero. This localized sparsity strategy not only gives the model clear physical interpretability, but also creates a "reconstruction barrier" for fault samples—that is, fault samples are difficult to be accurately reconstructed under strict sparsity constraints because they deviate from the normal centroids, thus amplifying the fault signal.

[0043] Specifically, the method for constructing hard sparse neighborhood constraints includes: obtaining multiple centroids by clustering the data in the training set based on a pre-set number of centroids, thus obtaining a baseline matrix; calculating the distance between each data point in the training set and each centroid based on a pre-set number of neighborhoods, sorting the obtained distances, and filtering to obtain the target centroid corresponding to each data point in the training set; obtaining a mask matrix based on each data point in the training set and its corresponding target centroid, where each row of the mask matrix corresponds to a centroid, each column corresponds to a data point in the training set, and the position of the target centroid corresponding to each data point in the training set is set to 1, while other positions are set to 0; calculating the sample coefficient matrix based on the mask matrix and its corresponding target centroid; and constructing hard sparse neighborhood constraints based on the mask matrix and the sample coefficient matrix. For example, in a chemical process, five normal samples are collected, each with three features (such as temperature, pressure, and flow rate). The training set data matrix is... One column corresponds to one sample; the first row contains temperature data, the second row contains pressure data, and the third row contains flow rate data; the preset number of centroids is also included. k The number of neighbors is 2. Given a value of 1, K-Means clustering yields two centroids: centroid 1. Centroid 2 The baseline matrix is ​​obtained. Calculate the Euclidean distance between each data point and centroids 1 and 2 to obtain distance 1 and distance 2 for each data point. For example, the Euclidean distance between sample 1 and centroid 1 is distance 1, and the Euclidean distance between sample 1 and centroid 2 is distance 2. Since distance 1 is less than distance 2, the target centroid of sample 1 is determined to be centroid 1. Similarly, the target centroids of samples 2 and 5 are determined to be centroid 1, and the target centroids of samples 3 and 4 are determined to be centroid 2. The mask matrix consists of 2 rows and 5 columns. The first row corresponds to centroid 1, the second row corresponds to centroid 2, and sample 1 corresponds to the first column of the mask matrix. The target centroid of sample 1 is centroid 1, located in the first row and first column of the mask matrix. Therefore, set this column to 1, set the first column of the second row to 0, and so on, to determine the target centroid of the mask matrix. Sample 1 x Sample coefficient of 1 The sample coefficients are calculated and then calculated for each sample in the same way. The calculated sample coefficients are used to replace the "1"s in the mask matrix to obtain the sample coefficient matrix. Hard sparse neighborhood constraints are constructed based on the mask matrix and the sample coefficient matrix.

[0044] Specifically, constraint two: the manifold structure is kept regularized. This term utilizes the Laplace matrix. For sparse coefficient matrix Apply a smoothing constraint. Minimizing this term means that if two samples are neighbors in the original space, their low-dimensional representations must also remain close. This geometry-aware strategy ensures that the model "remembers" the local topological relationships of normal data during the feature learning phase, making it sensitive to any faults that disrupt the local neighborhood structure (such as intermittent faults).

[0045] Specifically, the synergistic effect: the two constraints mentioned above complement each other in the optimization process. The "sparse constraint" limits the range of the reconstructed basis of the samples, while the "manifold regularization" limits the relative positional relationship between the samples. Together, they achieve an accurate "imprint" of the manifold under normal operating conditions.

[0046] Specifically, the MR-HTCNMF method is designed based on a deep understanding of the manifold structure under normal operating conditions: normal data is not randomly distributed, but rather tightly clustered around several stable operating modes (centroids). To accurately pinpoint this characteristic, this application introduces highly discriminative dual constraints on top of nonnegative matrix factorization: ① Hard-Threshold Sparse Constraint: Unlike traditional... Soft regularization: This application employs a "hard thresholding" strategy, which forces each sample's representation coefficient vector to retain only those samples with the closest geometric distance. The non-zero values ​​of the elements are forcibly truncated to zero at the remaining positions. This constraint not only achieves strict localization representation but also sets stringent "admission criteria" for fault detection: any fault sample deviating from the normal centroid will produce a drastic reconstruction error because it cannot meet this strict neighborhood matching requirement, thus being sensitively identified. ② Manifold Regularization: By introducing a graph Laplacian regularization term, the local topological relationships (such as nearest neighbor connections) of the data in the high-dimensional space are forcibly mapped to the low-dimensional feature space. This constraint ensures that the model can "freeze" and memorize the geometric structure between normal samples during the learning process, making intermittent or dynamic faults that disrupt this local neighborhood relationship nowhere to hide.

[0047] In one embodiment, such as Figure 3 As shown, step S2 involves training the dual-constraint model based on the data in the training set, including: S23: Based on the pre-set solution algorithm, solve the objective function in the dual-constraint model and determine the objective basis matrix.

[0048] In one embodiment, to overcome the limitations of traditional multiplicative updates in handling manifold regularization terms, this step employs the Alternating Direction Multiplier Method (ADMM) for optimization to obtain the target basis matrix. .

[0049] In one embodiment, this step is the computational core of the entire fault monitoring method. Its goal is to efficiently solve the non-convex, non-smooth optimization problem constructed in step S2, which contains complex manifold regularization terms, using the Alternating Direction Multiplier Method (ADMM). ADMM decouples the original problem into several easily tractable subproblems by introducing auxiliary variables, thus overcoming the convergence difficulty of traditional gradient descent methods when dealing with graph Laplace terms. Ultimately, it learns a basis matrix that accurately represents normal operating conditions. The specific implementation process is as follows: S231: Parameter initialization and variable configuration.

[0050] First, set the model hyperparameters: the number of basis vectors (centroids). (For example Hard-constrained neighborhood number (For example manifold regularization intensity And the penalty parameters unique to the ADMM algorithm Initialize the basis matrix. and sample coefficient matrix (Random initialization is usually used), and the auxiliary variables required for ADMM are initialized. , and dual variables , The projection matrix is ​​initialized to zero. For random orthogonal matrices, latent variables For random matrices, noise and Lagrange multipliers This is an all-zero matrix. Set the penalty parameter. and its growth rate .

[0051] S232: Alternating Multivariable Iterative Optimization (ADMMLoop). Enter the iterative loop and execute the following three sub-steps sequentially until the dual-constraint model converges. In each iteration, update the variables in the following order: Step 1: Update the hard sparse neighborhood mask matrix (Geometric distribution).

[0052] Fixed basis matrix Calculate the training set data matrix The Euclidean distance between all sample data and each basis vector. For each sample... Only keep the closest one Based on the basis vectors, apply them to the mask matrix. No. The corresponding position in the column is set to 1, and the remaining positions are forcibly set to 0. This step achieves a "hard assignment" of the local pattern to which the sample belongs.

[0053] Step 2: Update the sample coefficient matrix (Solve the Sylvester equation).

[0054] Fixed basis matrix and mask matrix Solve for terms containing manifold regularization. of Minimize the subproblem. Within the ADMM framework, the condition for the derivative of this subproblem to be zero is derived as a standard Sylvester matrix equation: The equation is in the form of This application utilizes a specialized linear algebra solver (such as the Bartels-Stewart algorithm) to directly and accurately solve the equation, avoiding gradient approximation in traditional methods. This ensures computational efficiency while perfectly preserving the mathematical constraints of the manifold structure. Subsequently, auxiliary variables are used... The solution results are processed using nonnegative projection and sparse masking: This ensures that the coefficients are non-negativity and strictly adhere to hard sparsity constraints.

[0055] Step 3: Update the basis matrix (Regularized least squares).

[0056] Fixed updated sparse coefficient matrix .at this time The solution degenerates into a least squares problem with ADMM regularization terms.

[0057] Update directly by solving the system of linear equations : And apply nonnegative projection to it as well: .

[0058] Step 4: Update the dual variable.

[0059] Update the dual variable according to the standard rules of ADMM. and This is done by accumulating the original residuals to drive the satisfaction of the constraints. ; .

[0060] S233: Convergence determination and model output.

[0061] At the end of each iteration, the data reconstruction error is calculated. The rate of change. When the error change is less than a preset threshold (e.g., ... Training stops when the maximum number of iterations is reached.

[0062] S234: Output.

[0063] Save the target basis matrix that eventually converges. This matrix constitutes the "manifold skeleton" of normal operating conditions, containing all typical local feature patterns of the system during stable operation, and serves as the benchmark for subsequent online fault detection.

[0064] At this point, the basic model training for the dual-constraint model is complete.

[0065] Given the discontinuity caused by the "hard threshold" and the variable coupling caused by the manifold regularization term, traditional optimization algorithms are difficult to solve directly. This application uses the Alternating Direction Multiplier Method (ADMM) to optimize the objective function step by step. The ADMM algorithm decouples the complex original problem by introducing auxiliary and dual variables, cleverly transforming the manifold regularization term into a Sylvester matrix equation for accurate solution, and combining projection operators to handle nonnegativity and hard threshold truncation. This "divide and conquer" strategy ensures the convergence and numerical stability of the algorithm under non-convex constraints, and successfully learns a "manifold dictionary" that can accurately imprint normal operating conditions.

[0066] From a mechanistic perspective, the dual constraints optimized by the Alternating Directional Multiplier Method (ADMM) achieve "maximization of the difference" between normal and fault modes. During online monitoring, normal samples can be accurately reconstructed due to their conformity to the manifold distribution and hard threshold conditions; however, fault samples face significant reconstruction obstacles due to the truncation mechanism of the "hard threshold" and the topological constraints of the "manifold," resulting in a significant increase in residuals. Experiments show that in high-dimensional nonlinear scenarios such as chemical engineering and power generation, this application, with its unique hard threshold mechanism, significantly improves the detection rate of minor and early-stage faults while maintaining an extremely low false alarm rate, providing solid and reliable algorithmic support for intelligent industrial operation and maintenance.

[0067] S24: Project the data in the training set onto the target basis matrix to obtain the reconstructed data.

[0068] In one embodiment, after the basic model of the dual-constraint model is trained, a statistical standard needs to be established to define the boundary between "normal" and "abnormal" so that the final dual-constraint model can output fault judgment results.

[0069] In one embodiment, the training set data matrix Projecting back the learned target basis matrix The above yields reconstructed data. During this process, the target basis matrix is ​​fixed. The coefficient matrix of the training set data matrix remains unchanged, but the ADMM steps are simplified to solve for it. .

[0070] S25: Construct statistics based on the data in the training set and the reconstructed data.

[0071] In one embodiment, using SPE A statistic is used to measure the degree to which a sample deviates from the normal model. For the... Training set data samples The statistical measure is calculated as follows: ,in, Represents the sparse coefficient matrix The Column vectors.

[0072] S26: Based on a pre-set threshold estimation algorithm, determine the control limits according to the statistics.

[0073] In one embodiment, the threshold estimation algorithm is a nonparametric kernel density estimation (KDE) or percentile method.

[0074] In one embodiment, based on all sample data from the training set SPE Value distribution, determining control limits :

[0075] in, This is the significance level (usually taken as 0.01 or 0.05). This represents the percentile function. The specific operation is: [The function is described in the original text]. The values ​​in the table are arranged in descending order, and the first value is selected. The numerical value of the position (i.e., corresponding to) The boundary values ​​of the confidence interval are used as control limits. Specifically, the quantiles at a 99% confidence level are selected as the control limits for fault detection.

[0076] S3: Test the trained dual-constraint model using data from the test set, including: the maximum value of the normalized data when normalizing the data in the test set. and minimum value Data is acquired from the training set; the normalized test set data is input into the trained dual-constraint model, and the test set data is projected onto the target basis matrix to obtain reconstructed test set data; based on the test set data and its reconstructed data, the corresponding statistics for the test set data are calculated; the statistics for the test set data are compared with the control limits in the dual-constraint model, and fault diagnosis is performed based on the comparison results; the fault diagnosis results are compared with the data labels in the test set data to determine whether the dual-constraint model has given false alarms. Furthermore, when the statistics for the data to be detected are greater than the control limits in the dual-constraint model, a fault is determined to have occurred; when the statistics for the data to be detected are less than or equal to the control limits in the dual-constraint model, normal operation is determined.

[0077] In one embodiment, the test set includes normal operation data and fault operation data. Understandably, normal operation data is labeled "normal" and fault operation data is labeled "fault". Furthermore, fault operation data can be divided according to type and further fault category labels can be set.

[0078] In one embodiment, during the test, the statistics corresponding to the normal operation data in the test set should be less than or equal to the control limits in the dual constraint model; the statistics corresponding to the fault operation data in the test set should be greater than the control limits in the dual constraint model.

[0079] Specifically, during the testing phase, once the statistical quantity corresponding to the data in the test set exceeds the preset control limit, an anomaly is determined to have occurred at that moment, and the diagnosis result is a fault. To quantitatively evaluate the monitoring effectiveness of the method in this application for different types of faults, it is necessary to perform sample-by-sample discrimination on the test set data, statistically analyze the response of the dual-constraint model in the fault interval, and calculate the fault detection rate (FDR) and false alarm rate (FAR) as core evaluation indicators. The calculation formulas are as follows:

[0080] The variables in the formula are defined as follows: TP (True Case): the number of samples that are actually in a fault state and whose monitoring statistic is correctly exceeding the limit, thus being successfully identified as "fault"; FN (False Negative Case): the number of samples that are actually in a fault state but whose monitoring statistic is not exceeding the limit, resulting in being incorrectly judged as "normal" (i.e., missed reports); the sum of TP and FN constitutes all fault samples in the test set; FP (False Positive Case): the number of samples that are actually in a normal state but whose monitoring statistic is abnormally exceeding the limit, resulting in being incorrectly judged as "fault" (i.e., false reports); TN (True Negative Case): the number of samples that are actually in a normal state and whose monitoring statistic is within the limit, resulting in being correctly judged as "normal"; the sum of FP and TN constitutes all normal samples in the test set.

[0081] Specifically, in order to verify the effectiveness and technical advantages of the dual constraint model based on sparse locality and manifold regularization proposed in this application based on ADMM optimization, the following will be described in detail with specific industrial process examples.

[0082] It should be understood that the listed embodiments are intended to help understand the technical content and beneficial effects of the present invention, and do not constitute a limitation on the scope of protection of the present invention.

[0083] Specifically, this application was applied to 21 standard fault test sets for TE processes, and the fault detection rate (FDR) and false alarm rate (FAR) were statistically analyzed. Experimental results show that, thanks to the preservation of local structure by manifold regularization and the accurate solution of ADMM, this method significantly improves the FDR compared to traditional PCA and standard NMF methods when detecting minor faults and strongly nonlinear faults, while maintaining an extremely low FAR (<1%), verifying the effectiveness and robustness of this method in monitoring complex chemical processes.

[0084] Specifically, the testing process will be described as follows: Specifically, regarding the dataset: This experiment uses the classic Tennessee Eastman (TE) chemical process dataset as a benchmark validation platform for industrial process monitoring. A TE process is a continuous chemical production system containing multiple unit operations such as reaction, condensation, separation, and circulation, exhibiting typical characteristics of strong multivariate coupling, nonlinearity, and significant dynamics. Due to its clear process mechanism, rich fault types, and large data scale, the TE dataset has become one of the most commonly used public benchmarks in the field of industrial process fault detection and diagnosis.

[0085] 1) The TE process dataset provides multivariate time-series sampling data, covering key process measurements and manipulated variables, which can comprehensively reflect the operating status of the plant. The commonly used variable composition includes 41 process measurement variables and 11 manipulated variables.

[0086] 2) TE preset faults (d01-d21 / IDV1-IDV21).

[0087] The TE benchmark provides 21 typical disturbances / faults (corresponding to faults IDV1 to IDV21), covering categories such as step disturbances, random fluctuations, slow drift, valve jamming / fixed position, and unknown disturbances, as shown in Table 1. Among them, IDV16–IDV20 are explicitly marked as “unknown type disturbances” in the public benchmark, used to verify the generalization ability of the detection method; IDV21 is a typical valve fixed position fault.

[0088] Table 1 Introduction to TE process datasets

[0089] The test process and output results are described below: Step 1: Normalize the data in the test set, and find the maximum value after normalization. and minimum value Obtained from data in the training set.

[0090] Specifically, regarding the data in the test set (in ), using the training set statistics (maximum value) saved during the training phase and minimum value The model is standardized and mapped to the [0,1] interval to ensure consistency with the input space of the dual-constraint model.

[0091] Step 2: Input the normalized test set data into the trained dual-constraint model, project the test set data onto the target basis matrix, and obtain the reconstructed data of the test set data.

[0092] Specifically, the system call target basis matrix and maintain The inference logic is based on the ADMM algorithm, solving for the data in the test set. In the basis matrix The optimal sparse coefficient matrix This process essentially involves solving the following constrained optimization problem:

[0093] in, card This represents a hard sparse constraint, which forces the coefficient vector to retain only the values ​​of the specified values. non-zero elements (corresponding to the nearest) (basic vectors).

[0094] Specifically, using the solved optimal sparse coefficient matrix and target basis matrix Calculate the reconstructed data from the test set. .

[0095] Step 3: Calculate the statistics corresponding to the test set data based on the data in the test set and its reconstructed data.

[0096] Specifically, calculate the statistics corresponding to the test set data to quantify the degree to which the test set data samples deviate from the normal manifold space: .

[0097] Step 4: Fault Determination and Alarm. Calculate the statistics corresponding to the test set data. Control limits in dual-constraint models The comparison is used to diagnose faults. The fault diagnosis results are compared with the data labels in the test set data to determine whether the dual-constraint model has given false alarms. When the statistic corresponding to the data in the test set is greater than the control limit in the dual-constraint model, a fault is determined to have occurred; when the statistic corresponding to the data in the test set is less than or equal to the control limit in the dual-constraint model, the operation is determined to be normal.

[0098] like The diagnostic result is that the test set data is in a normal control state. If the data label is "normal", the dual constraint model is accurate. If the data label is "fault", the dual constraint model is a false alarm.

[0099] like The diagnostic result indicates that the test set data is faulty or abnormal. If the data label is "normal", the dual constraint model is a false alarm; if the data label is "faulty", the dual constraint model is accurate.

[0100] Through the above steps, the system can continuously and in real-time monitor the TE process and assess the health status of the process operation based on the changing trends of statistics. Table 2 shows the monitoring results for 11 types of faults; the specific fault detection results are as follows: Figures 4 to 14 As shown. Figures 4 to 14 The diagram shows the fault detection rate and false alarm rate for this type of fault. The fault detection rate represents the true faults that the model can identify, while the false alarm rate represents the probability that normal data is misclassified as faulty. Figure 4 The dual-constraint model shown has a fault detection rate of 99.75% and a false alarm rate of 1.88% for the d01 fault; Figure 5 The dual-constraint model shown has a fault detection rate of 98.38% and a false alarm rate of 0% for the d02 fault; Figure 6 The dual-constraint model shown has a fault detection rate of 5.62% and a false alarm rate of 2.50% for the d04 fault; Figure 7 The dual-constraint model shown has a 100% fault detection rate and a false alarm rate of 0.62% for fault d06; Figure 8 The dual-constraint model shown has a 100% fault detection rate and a 0% false alarm rate for fault d07; Figure 9 The dual-constraint model shown has a fault detection rate of 97.62% and a false alarm rate of 0% for fault d08; Figure 10 The dual-constraint model shown has a fault detection rate of 99% and a false alarm rate of 6.88% for the d12 fault; Figure 11 The dual-constraint model shown has a fault detection rate of 95.62% and a false alarm rate of 0.62% for the d13 fault; as Figure 12 The dual-constraint model shown has a 100% fault detection rate and a false alarm rate of 3.12% for fault d14; Figure 13 The dual-constraint model shown has a fault detection rate of 92.38% and a false alarm rate of 0.62% for fault d17; Figure 14 The dual-constraint model shown has a fault detection rate of 90% and a false alarm rate of 2.50% for d18 faults. Figures 4 to 14 This includes control thresholds: control limits at a 99% confidence level, represented by the horizontal dashed line in the graph, which indicates the maximum permissible deviation of normal data; and statistics: the percentage of each sample... SPEThe statistics are calculated in real time, represented by solid lines in the graph. Larger values ​​indicate a greater deviation from the normal model. Fault introduction time: the sample index position where the fault begins (approximately 160), represented by the vertical dashed line in the graph. From this point, the system enters a fault state. X-axis: Sample index, the sequential number of sample collection, from 0 to 1000, representing the time or order of data collection. Y-axis: Statistics. SPE The numerical value of the statistic indicates a greater deviation from the normal state; a larger value indicates a more severe deviation. Normal phase (sample 0-160): SPE If the value remains stable below the threshold, it indicates that the system is operating normally and the data perfectly matches the normal model; a fault occurs (around sample 160): SPE A sudden and sharp increase in the value, exceeding the threshold and reaching a peak, indicates a sudden change in the system state and the onset of a fault. Fault duration phase (samples 160-1000): SPE The value remained above the threshold, and although it fluctuated, it never returned to the normal range, indicating that the fault persisted and the model continued to accurately detect the abnormal state. Figures 4 to 14 This demonstrates the successful application of fault diagnosis algorithms based on statistical process control in industrial scenarios, through monitoring... SPE Changes in statistical quantities can detect system anomalies in real time and accurately, providing reliable technical support for preventative maintenance and fault early warning of equipment. This method is widely used in production process monitoring in industries such as chemical, power, and manufacturing.

[0101] Table 2 Fault Detection Results

[0102] In one embodiment, the method further includes: S4: Performing industrial process fault diagnosis based on the tested dual-constraint model. This enables online real-time monitoring and fault diagnosis.

[0103] The online monitoring process and output results are described below: Step 1: Obtain the running data to be tested. For The online monitoring data collected in real time is normalized, and the maximum value of the normalized data is... and minimum value Obtained from data in the training set.

[0104] Specifically, Real-time online monitoring of operational data to be tested (in ), using the training set statistics (maximum value) saved during the training phase and minimum value The model is standardized and mapped to the [0,1] interval to ensure consistency with the input space of the dual-constraint model.

[0105] Step 2: Input the normalized running data to be detected into the trained dual-constraint model, and project the running data to be detected onto the target basis matrix to obtain the reconstructed data of the running data to be detected.

[0106] Specifically, the system call target basis matrix and maintain Fixed and unchanging. Based on the inference logic of the ADMM algorithm, the running data to be detected is solved. In the basis matrix The optimal sparse coefficient matrix This process essentially involves solving the following constrained optimization problem:

[0107] in, card This represents a hard sparse constraint, which forces the coefficient vector to retain only the values ​​of the specified values. non-zero elements (corresponding to the nearest) (basic vectors).

[0108] Specifically, using the solved optimal sparse coefficient matrix and target basis matrix Calculate the reconstructed data of the running data to be detected. .

[0109] Step 3: Calculate the statistics corresponding to the running data to be tested based on the running data to be tested and its reconstructed data.

[0110] Specifically, the statistics corresponding to the data to be tested are calculated to quantify the degree to which the data deviates from the normal manifold space: .

[0111] Step 4: Fault Judgment and Alarm. Calculate the statistical quantities corresponding to the operational data to be tested. Control limits in dual-constraint models The system compares the data and diagnoses faults based on the comparison results. When the statistic corresponding to the data to be tested is greater than the control limit in the dual constraint model, a fault is determined to have occurred and an alarm is triggered. When the statistic corresponding to the data to be tested is less than or equal to the control limit in the dual constraint model, the system is determined to be operating normally.

[0112] like The process is determined to be under normal control.

[0113] like If a fault or abnormality is detected in the current process, the system will immediately trigger an alarm signal.

[0114] As can be seen, in the above scheme, this application constructs extremely strict constraint conditions by using hard sparse neighborhood constraints and manifold structure regularization constraints in the dual constraint model, which solves the problem of detecting minute faults in complex industrial processes and ensures high sensitivity and low false alarm rate in fault detection.

[0115] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0116] In one embodiment, an industrial process fault diagnosis device is provided, which corresponds one-to-one with the industrial process fault diagnosis methods described in the above embodiments. For example... Figure 15 As shown, the industrial process fault diagnosis device includes an acquisition module 101, a training module 102, a testing module 103, and a diagnosis module 104. Detailed descriptions of each functional module are as follows: The acquisition module 101 is used to acquire historical running data and divide the historical running data into training set and test set; Training module 102 is used to construct a dual-constraint model. The dual-constraint model is trained based on the data in the training set. The constraints of the dual-constraint model include hard sparse neighborhood constraints and manifold structure preservation regularization constraints. Test module 103 is used to test the trained dual-constraint model based on the data in the test set; Diagnostic module 104 is used for industrial process fault diagnosis based on a tested dual-constraint model.

[0117] The historical operating data obtained from module 101 includes normal operating data and fault operating data; the data in the training set includes normal operating data; and the data in the test set includes normal operating data and fault operating data.

[0118] The training module 102 is also used to construct an affinity matrix based on the data in the training set and to calculate the Laplacian matrix based on the affinity matrix; Based on the pre-set basis matrix and coefficient matrix, an objective function is constructed according to the Laplacian matrix. The objective function includes a data reconstruction error term and a manifold regularization term.

[0119] Training module 102 is also used to solve the objective function in the dual-constraint model based on a pre-set solution algorithm and determine the objective basis matrix; The data in the training set is projected onto the target basis matrix to obtain the reconstructed data; Based on the data in the training set and the reconstructed data, construct statistics; Based on a pre-set threshold estimation algorithm, control limits are determined according to statistics.

[0120] Training module 102 is also used to obtain multiple centroids by clustering the data in the training set based on a pre-set number of centroids, and to obtain a reference matrix; Based on a pre-set number of neighbors, the distance between each data point in the training set and each centroid is calculated, and the obtained distances are sorted to select the target centroid corresponding to each data point in the training set. Based on each data point in the training set and its corresponding target centroid, a mask matrix is ​​obtained. Each row in the mask matrix corresponds to a centroid, and each column corresponds to a data point in the training set. The position of the target centroid corresponding to each data point in the training set is set to 1, and the other positions are set to 0. Based on the mask matrix, the sample coefficient matrix is ​​calculated according to each data point in the training set and its corresponding target centroid. Hard sparse neighborhood constraints are constructed based on the mask matrix and the sample coefficient matrix.

[0121] Test module 103 is also used to normalize the data in the test set. The maximum and minimum values ​​of the normalization are obtained from the data in the training set. Based on the data in the normalized test set and the target basis matrix in the trained dual-constraint model, the reconstructed data of the data in the normalized test set is obtained; Calculate the statistics corresponding to the data in the test set based on the normalized test set data and its reconstructed data. The statistics corresponding to the data in the test set are compared with the control limits in the dual-constraint model, and fault diagnosis is performed based on the comparison results.

[0122] Test module 103 is also used to determine a fault when the statistic corresponding to the data in the test set is greater than the control limit in the dual constraint model; The system is considered to be running normally when the statistics corresponding to the data in the test set are less than or equal to the control limits in the dual-constraint model.

[0123] This invention provides an industrial process fault diagnosis device. This application constructs extremely strict constraint conditions by using hard sparse neighborhood constraints and manifold structure maintaining regularization constraints in a dual constraint model, which solves the problem of detecting minute faults in complex industrial processes and ensures high sensitivity and low false alarm rate in fault detection.

[0124] Specific limitations regarding industrial process fault diagnosis devices can be found in the limitations of industrial process fault diagnosis methods described above, and will not be repeated here. Each module in the aforementioned industrial process fault diagnosis device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0125] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 16 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external devices via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side method for industrial process fault diagnosis.

[0126] In one embodiment, a computer device is provided, which may be a device terminal, and its internal structure diagram may be as follows: Figure 17 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of an industrial process fault diagnosis method on the device side.

[0127] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Obtain historical operational data and divide the historical operational data into training set and test set; A dual-constraint model is constructed and trained on the data in the training set. The constraints of the dual-constraint model include hard sparse neighborhood constraints and manifold structure preservation regularization constraints. The trained dual-constraint model was tested based on data from the test set. Industrial process fault diagnosis based on the tested dual-constraint model.

[0128] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain historical operational data and divide the historical operational data into training set and test set; A dual-constraint model is constructed and trained on the data in the training set. The constraints of the dual-constraint model include hard sparse neighborhood constraints and manifold structure preservation regularization constraints. The trained dual-constraint model was tested based on data from the test set. Industrial process fault diagnosis based on the tested dual-constraint model.

[0129] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and device side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0130] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0131] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0132] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for diagnosing faults in an industrial process, characterized in that, include: Obtain historical operational data and divide the historical operational data into training set and test set; A dual-constraint model is constructed and trained on the data in the training set. The constraints of the dual-constraint model include hard sparse neighborhood constraints and manifold structure preservation regularization constraints. The trained dual-constraint model was tested using data from the test set. Industrial process fault diagnosis based on the tested dual-constraint model.

2. The industrial process fault diagnosis method according to claim 1, characterized in that, Historical operational data includes normal operation data and fault operation data; The data in the training set includes normal operation data; The data in the test set includes normal operation data and fault operation data.

3. The industrial process fault diagnosis method according to claim 1, characterized in that, The objective function for constructing the dual-constraint model includes: Construct an affinity matrix based on the data in the training set, and calculate the Laplace matrix based on the affinity matrix; Based on the pre-set basis matrix and coefficient matrix, an objective function is constructed according to the Laplacian matrix. The objective function includes a data reconstruction error term and a manifold regularization term.

4. The industrial process fault diagnosis method according to claim 1, characterized in that, The dual-constraint model is trained based on the data in the training set, including: Based on a pre-set solution algorithm, the objective function in the dual-constraint model is solved to determine the objective basis matrix; The data in the training set is projected onto the target basis matrix to obtain the reconstructed data; Based on the data in the training set and the reconstructed data, construct statistics; Based on a pre-set threshold estimation algorithm, control limits are determined according to statistics.

5. The industrial process fault diagnosis method according to claim 1, characterized in that, Methods for constructing hard sparse neighborhood constraints include: Based on a pre-set number of centroids, multiple centroids are obtained by clustering the data in the training set, thus obtaining the baseline matrix; Based on a pre-set number of neighbors, the distance between each data point in the training set and each centroid is calculated, and the obtained distances are sorted and filtered to obtain the target centroid corresponding to each data point in the training set. Based on each data point in the training set and its corresponding target centroid, a mask matrix is ​​obtained. Each row in the mask matrix corresponds to a centroid, and each column corresponds to a data point in the training set. The position of the target centroid corresponding to each data point in the training set is set to 1, and the other positions are set to 0. Based on the mask matrix, the sample coefficient matrix is ​​calculated according to each data point in the training set and its corresponding target centroid. Hard sparse neighborhood constraints are constructed based on the mask matrix and the sample coefficient matrix.

6. The industrial process fault diagnosis method according to claim 1, characterized in that, The dual-constraint model was tested based on data from the test set, including: The data in the test set is normalized, and the maximum and minimum values ​​of the normalization are obtained from the data in the training set. Based on the data in the normalized test set and the target basis matrix in the trained dual-constraint model, the reconstructed data of the data in the normalized test set is obtained; Calculate the statistics corresponding to the data in the test set based on the normalized test set data and its reconstructed data. The statistics corresponding to the data in the test set are compared with the control limits in the dual-constraint model, and fault diagnosis is performed based on the comparison results.

7. The industrial process fault diagnosis method according to claim 6, characterized in that, A fault is determined to have occurred when the statistic corresponding to the data in the test set is greater than the control limit in the dual-constraint model. The system is considered to be running normally when the statistics corresponding to the data in the test set are less than or equal to the control limits in the dual-constraint model.

8. An industrial process fault diagnosis device, characterized in that, include: The acquisition module is used to acquire historical running data and divide the historical running data into training set and test set; The training module is used to build a dual-constraint model. It trains the dual-constraint model based on the data in the training set. The constraints of the dual-constraint model include hard sparse neighborhood constraints and manifold structure preservation regularization constraints. The testing module is used to test the trained dual-constraint model based on data in the test set. The diagnostic module is used for industrial process fault diagnosis based on the tested dual-constraint model.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes a computer program, it implements the steps of the industrial process fault diagnosis method as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of any one of the industrial process fault diagnosis methods as claimed in claims 1 to 7.

Images