DCS controller trusted state perception method and system

Abstract

The present invention provides a DCS controller trusted state perception method and system, which relates to the technical field of data perception and analysis, including obtaining multimodal heterogeneous data in multiple time windows of a DCS controller, adopting a multi-view subspace learning method to extract features, mapping different modal data to a common subspace, and obtaining corresponding multimodal fusion features; taking the multimodal fusion features as input, adopting an adaptive anomaly measurement algorithm, adaptively selecting cluster centers based on the product of local density and distance factors, determining normal clusters and abnormal clusters through clustering division and calculating anomaly scores, and obtaining controller state abnormality measurement; based on the controller state abnormality measurement, mapping to language value features through a preset membership function, constructing an evaluation cloud through calculating mathematical statistical features, determining a comprehensive evaluation cloud through weighted fusion, and performing similarity matching with a preset trusted state level standard cloud to determine the current trusted state of the DCS controller.

Description

Technical Field

[0001] The present invention relates to the technical field of data perception and analysis, and in particular to a DCS controller trusted state perception method and system. Background Art

[0002] As the complexity of industrial control systems continues to increase, the reliability and security of DCS (distributed control system) controllers, as their core components, have attracted much attention. Real-time perception of the trusted status of DCS controllers is of great significance to ensuring the safe operation of industrial control systems.

[0003] Traditional DCS controller status monitoring methods mainly rely on manual inspections and regular maintenance, which makes it difficult to achieve real-time and continuous status perception. At the same time, there is a lack of effective data fusion and analysis methods, resulting in insufficient accuracy and reliability of status perception. When machine learning algorithms are applied to the health status assessment of industrial equipment, they lack the comprehensive use of multi-source heterogeneous data and are difficult to be directly applied to the trusted status perception of DCS controllers.

[0004] To sum up, the existing technology lacks real-time and continuous status monitoring means, making it difficult to detect abnormal conditions of the controller in a timely manner, failing to fully utilize the multi-source heterogeneous data generated during the operation of the controller, insufficient information sources for status perception, lacking intelligent status assessment methods, and making it difficult to effectively mine status characteristics. The present invention can solve the problems in the existing technology. Summary of the invention

[0005] The embodiments of the present invention provide a DCS controller trusted state perception method and system, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A DCS controller trusted state perception method is provided, comprising:

[0008] Acquire multimodal heterogeneous data in multiple time windows of the DCS controller, use multi-view subspace learning method to extract features from the multimodal heterogeneous data in each time window, map different modal data to the common subspace, and obtain corresponding multimodal fusion features;

[0009] Taking the multimodal fusion features as input, an adaptive anomaly measurement algorithm is adopted to adaptively select cluster centers based on the product of local density and distance factor, and normal clusters and abnormal clusters are determined by clustering and calculating anomaly scores to obtain controller state abnormality measurement;

[0010] Based on the controller state abnormality measurement, it is mapped to language value features through a preset membership function, and an evaluation cloud is constructed by calculating mathematical statistical features. A comprehensive evaluation cloud is determined through weighted fusion, and similarity matching is performed with a preset trustworthy state level standard cloud to determine the current trustworthy state of the DCS controller.

[0011] In an optional embodiment,

[0012] The multi-view subspace learning method is used to extract features from the multimodal heterogeneous data in each time window, and the different modal data are mapped to the common subspace to obtain the corresponding multimodal fusion features including:

[0013] Preprocess the multi-modal heterogeneous operation data separately to determine the feature representation matrix corresponding to each mode;

[0014] Based on the multi-view subspace learning method, the feature representation matrix of each modality is mapped to a common subspace through a subspace mapping matrix to generate a subspace representation;

[0015] Based on the reconstruction error between the subspace representation of each modality and the preset consistency representation, measure the similarity between the subspace representation and the consistency representation, and determine the similarity term; based on the nuclear norm of the subspace mapping matrix, control the complexity of the subspace mapping matrix, and determine the complexity regularization term; based on the weighted distance of sample pairs in the same modality, maintain the global manifold structure, and determine the global manifold structure preservation term; based on the K nearest neighbor algorithm, determine similar sample pairs in the same modality, and based on the weighted distance of the similar sample pairs in the common subspace, maintain the local manifold structure, and determine the local manifold structure preservation term; construct an error function based on the similarity term, the complexity regularization term, the global manifold structure preservation term, and the local manifold structure preservation term;

[0016] Based on the alternating optimization strategy, the subspace mapping matrix is ​​optimized and the consistency representation is adjusted to minimize the error function, and the model parameters are iteratively updated until a preset number of iterations is reached to obtain the optimal subspace mapping matrix;

[0017] Based on the optimal subspace mapping matrix, the feature representation matrix of each modality is mapped to a common subspace to obtain a consistent optimal subspace representation;

[0018] Performing weighted fusion on the consistent optimal subspace representation to determine an intermediate fusion feature;

[0019] Based on multiple fully connected layers, an encoder sub-network is constructed, based on the encoder sub-network, a symmetric network structure is determined, and a decoder sub-network is constructed, and the encoder sub-network is combined with the decoder sub-network to construct an enhanced autoencoder network;

[0020] The intermediate fusion features are input into the enhanced autoencoder network, which reconstructs the input, learns the deep feature representation, and outputs the final multimodal fusion features.

[0021] In an optional embodiment,

[0022] According to the similarity term, the complexity regularization term, the global manifold structure preservation term and the local manifold structure preservation term, constructing an error function comprises:

[0023] The error function has the following formula:

[0024]

[0025] Among them, E sim represents the similarity term, m represents the modal ordinal number, M represents the total number of modalities, and Y m represents the subspace representation of the modality m in the common subspace, Y represents the consistency representation, ||·|| F represents the Frobenius norm, E complexity represents the complexity regularization term, W m represents the mapping matrix of mode m, ||·|| * represents the nuclear norm, E global represents the global manifold structure preservation term, tr(·) represents the diagonal elements of the corresponding matrix, and L m represents the Laplacian matrix of the graph corresponding to mode m, E local represents the local manifold structure preservation term, k represents the neighbor sequence number, K represents the total number of neighbors, and Y m (k) Indicates that Y m The subspace corresponding to the nearest neighbor k is represented, E represents the error function, λ1 represents the weight coefficient of the complexity regularization term, λ2 represents the weight coefficient of the global manifold structure preservation term, and λ3 represents the weight coefficient of the local manifold structure preservation term.

[0026] In an optional embodiment,

[0027] Taking the multimodal fusion features as input, an adaptive anomaly measurement algorithm is used to adaptively select cluster centers based on the product of local density and distance factor, and normal clusters and abnormal clusters are determined by clustering and calculating anomaly scores. The controller state anomaly measurement includes:

[0028] Based on the multimodal fusion features, each multimodal fusion feature is taken as a sample point, a feature sample set is determined, the Euclidean distance between two sample points is calculated, a distance matrix is ​​generated, the median of all elements in the distance matrix is ​​selected as a cutoff distance threshold, and the local density of each sample point is calculated in combination with the distance matrix;

[0029] Iterate each sample point as the reference sample point, select all sample points with a local density greater than the reference sample point as comparison sample points, calculate the distance between each comparison sample point and the reference sample point, and select the minimum distance as the minimum distance factor; if the local density of the reference sample point is the largest, use all other sample points as comparison sample points, and select the maximum distance as the minimum distance factor;

[0030] Calculate the product of the local density corresponding to each sample point and the minimum distance factor, determine the multiplication factor of each sample point, arrange the multiplication factors in descending order, and calculate the multiplication factor difference between adjacent sample points;

[0031] Determine the median of the multiplication factor difference, determine the center threshold, extract the sample points whose multiplication factor difference is greater than the center threshold, and determine the cluster center;

[0032] Calculate the distances from all member sample points corresponding to non-cluster centers to each cluster center, assign the member sample points to the cluster center with the nearest distance, and generate clusters;

[0033] Based on each cluster, calculate the individual distance from each member sample point in the cluster to the cluster center, determine the maximum intra-cluster distance, and calculate the abnormal score of each member sample point based on the individual distance, maximum intra-cluster distance and local density;

[0034] All sample points are sorted by anomaly scores, and divided into normal clusters and abnormal clusters according to the preset anomaly score threshold. The controller state anomaly metric is determined based on the total number of sample points in the abnormal cluster and each anomaly score.

[0035] In an optional embodiment,

[0036] The minimum distance factor is formulated as follows:

[0037]

[0038] Among them, δ i represents the minimum distance factor of sample point i, i represents sample point i, j represents another sample point j, ρ i represents the local density of sample point i, ρ j Represents the local density of sample point j, max c ρ c represents the local density maximum, d ij Represents the distance between sample point i and sample point j;

[0039] The anomaly metric is formulated as follows:

[0040]

[0041] Among them, a represents the sample point in the abnormal cluster, X a represents the abnormality measure of sample point a, N represents the total number of all sample points, N A represents the total number of abnormal cluster sample points, A represents the abnormal cluster, s a Indicates the abnormal score of sample point a, max b∈A s b represents the maximum anomaly score in the anomaly cluster, q represents the anomaly score power index, and ρ a Represents the local density of sample point a.

[0042] In an optional embodiment,

[0043] Based on the controller state abnormality measurement, the preset membership function is used to map it to a language value feature, and the evaluation cloud is constructed by calculating the mathematical statistical features. The comprehensive evaluation cloud is determined by weighted fusion, and the similarity is matched with the preset trustworthy state level standard cloud to determine the current trustworthy state of the DCS controller, including:

[0044] Mapping the controller state abnormality metric to a preset language value feature set through a pre-constructed Gaussian membership function, calculating the membership of the abnormality metric to each language value feature, and determining a membership vector;

[0045] For each language value feature, based on the membership vector, the membership-weighted anomaly metric mean is calculated, the expectation is determined, the membership-weighted anomaly metric standard deviation is calculated, the entropy is determined, based on a preset proportionality coefficient, the product of the entropy and the proportionality coefficient is calculated, the hyperentropy is determined, and based on the expectation, the entropy and the hyperentropy, an evaluation cloud is constructed;

[0046] The weighted average cloud operation rule is adopted to weight the expectation, entropy and super entropy of each evaluation cloud to generate a comprehensive evaluation cloud of the overall trustworthy state of the controller;

[0047] Based on a set of pre-constructed credible status grade standard clouds, the similarity between the comprehensive evaluation cloud and each grade standard cloud is calculated, the adaptation grade standard cloud corresponding to the maximum similarity is extracted, and based on the credible status corresponding to the adaptation grade standard cloud, the current credible status of the DCS controller is determined.

[0048] In an optional embodiment,

[0049] Based on the expectation, the entropy and the super entropy, constructing an evaluation cloud comprises:

[0050] The expectation is as follows:

[0051]

[0052] Among them, B orepresents the expectation of the linguistic value feature o, ω oa Indicates the weight corresponding to the sample point a mapping the language value feature o, u oa Indicates the membership of the abnormal measure of sample point a to the language value feature o;

[0053] The entropy is formulated as follows:

[0054]

[0055] Among them, T o The entropy of the linguistic value feature o;

[0056] The super entropy is formulated as follows:

[0057] H o =η1·(T o ) r +η2·log(T o +1);

[0058] Among them, H o It represents the super entropy of the language value feature o, η1 represents the entropy influence degree control coefficient, r represents the entropy amplification degree index, and η2 represents the logarithmic term influence degree coefficient.

[0059] According to a second aspect of the embodiments of the present invention,

[0060] A DCS controller trusted state perception system is provided, comprising:

[0061] The first unit is used to obtain multimodal heterogeneous data in multiple time windows of the DCS controller, use a multi-view subspace learning method to extract features from the multimodal heterogeneous data in each time window, map different modal data to a common subspace, and obtain corresponding multimodal fusion features;

[0062] The second unit is used to use the multimodal fusion feature as input, adopt an adaptive anomaly measurement algorithm, adaptively select cluster centers based on the product of local density and distance factor, determine normal clusters and abnormal clusters through clustering and calculating anomaly scores, and obtain controller state abnormality measurement;

[0063] The third unit is used to map the controller state abnormality measurement into language value features through a preset membership function, construct an evaluation cloud by calculating mathematical statistical features, determine the comprehensive evaluation cloud through weighted fusion, and perform similarity matching with the preset trustworthy state level standard cloud to determine the current trustworthy state of the DCS controller.

[0064] According to a third aspect of the embodiments of the present invention,

[0065] An electronic device is provided, comprising:

[0066] processor;

[0067] a memory for storing processor-executable instructions;

[0068] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0069] A fourth aspect of the embodiments of the present invention is:

[0070] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0071] In an embodiment of the present invention, . BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 A schematic diagram of a flow chart of a DCS controller trusted state perception method according to an embodiment of the present invention;

[0073] Figure 2 Schematic diagram of the structure of a DCS controller trusted state perception system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0075] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0076] Figure 1 FIG. 1 is a flow chart of a method for sensing a DCS controller's trusted state according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0077] S101. Obtain multimodal heterogeneous data in multiple time windows of the DCS controller, use a multi-view subspace learning method to extract features from the multimodal heterogeneous data in each time window, map different modal data to a common subspace, and obtain corresponding multimodal fusion features;

[0078] The multimodal heterogeneous data specifically refers to operating data from multiple different sensors or data sources, and the operating data has different formats, structures and properties, including: numerical data, such as temperature, pressure, flow, current, voltage, etc. from sensors; image and video data, used to monitor the operating status of equipment or environmental changes; audio data, used to monitor the operating sound of the machine and identify abnormal noise; text data, such as log files and operation records; time series data, continuous records from sensors, such as vibration signals.

[0079] The common subspace specifically refers to a low-dimensional space found in multimodal data analysis that can accommodate data features from different modalities; the common subspace is a mapping space, and all data of different modalities can be projected into this space through a certain mapping method, so that in the mapping space, data of different modalities can be uniformly represented and processed; by mapping heterogeneous multimodal data to the same low-dimensional common subspace, data features of different modalities can be effectively fused and compared, thereby improving the interpretability of the data and the performance of the model.

[0080] In this embodiment, by mapping data of different modes to a common subspace, data fusion and unified processing can be achieved, avoiding the differences and heterogeneity between data of each mode, making data analysis and processing simpler and more efficient; the multi-view subspace learning method can more accurately identify and detect abnormal situations by integrating data features of different modes; the multi-view subspace learning method can more effectively extract useful features from multi-modal data. By mapping to a common subspace, data features of different modes can complement and strengthen each other, extracting more effective feature representations; through the fusion of multi-modal data, a comprehensive analysis perspective can be provided, and the operating status of the system can be fully understood from different angles, which helps to analyze problems more comprehensively and discover potential hidden dangers and optimization opportunities.

[0081] In an optional embodiment, a multi-view subspace learning method is used to extract features from multimodal heterogeneous data in each time window, and different modal data are mapped to a common subspace to obtain corresponding multimodal fusion features including:

[0082] Preprocess the multi-modal heterogeneous operation data separately to determine the feature representation matrix corresponding to each mode;

[0083] Based on the multi-view subspace learning method, the feature representation matrix of each modality is mapped to a common subspace through a subspace mapping matrix to generate a subspace representation;

[0084] Based on the reconstruction error between the subspace representation of each modality and the preset consistency representation, measure the similarity between the subspace representation and the consistency representation, and determine the similarity term; based on the nuclear norm of the subspace mapping matrix, control the complexity of the subspace mapping matrix, and determine the complexity regularization term; based on the weighted distance of sample pairs in the same modality, maintain the global manifold structure, and determine the global manifold structure preservation term; based on the K nearest neighbor algorithm, determine similar sample pairs in the same modality, and based on the weighted distance of the similar sample pairs in the common subspace, maintain the local manifold structure, and determine the local manifold structure preservation term; construct an error function based on the similarity term, the complexity regularization term, the global manifold structure preservation term, and the local manifold structure preservation term;

[0085] Based on the alternating optimization strategy, the subspace mapping matrix is ​​optimized and the consistency representation is adjusted to minimize the error function, and the model parameters are iteratively updated until a preset number of iterations is reached to obtain the optimal subspace mapping matrix;

[0086] Based on the optimal subspace mapping matrix, the feature representation matrix of each modality is mapped to a common subspace to obtain a consistent optimal subspace representation;

[0087] Performing weighted fusion on the consistent optimal subspace representation to determine an intermediate fusion feature;

[0088] Based on multiple fully connected layers, an encoder sub-network is constructed, based on the encoder sub-network, a symmetric network structure is determined, and a decoder sub-network is constructed, and the encoder sub-network is combined with the decoder sub-network to construct an enhanced autoencoder network;

[0089] The intermediate fusion features are input into the enhanced autoencoder network, which reconstructs the input, learns the deep feature representation, and outputs the final multimodal fusion features.

[0090] The sample pair in the same modality specifically refers to a pair of two samples selected in the same data modality. By calculating the similarity between the sample pairs, the internal structure of the data can be better understood and preserved;

[0091] The similar samples specifically refer to samples that are close in distance in the feature space and have similar attributes or characteristics. By identifying and utilizing similar samples, the accuracy and robustness of the model can be enhanced during data analysis and processing;

[0092] The global manifold structure specifically refers to the overall distribution and structural form of data in high-dimensional space. Maintaining the global manifold structure means that in the process of data mapping and dimensionality reduction, the global geometric features and topological relationships of the original data in high-dimensional space should be preserved as much as possible.

[0093] The local manifold structure specifically refers to the distribution and structural characteristics of data in a local area. Maintaining the local manifold structure means that in the process of data mapping and dimensionality reduction, the local relationship and relative position between adjacent or similar sample points should be preserved to ensure the integrity of local features.

[0094] The multimodal heterogeneous operating data collected by the DCS controller are preprocessed. For each mode of data, corresponding data cleaning, normalization, feature extraction and other methods are used according to the corresponding data characteristics and data format to convert the original data into a feature representation matrix suitable for subsequent analysis, where the feature representation matrices of different modes have different dimensions and scales.

[0095] Using the multi-view subspace learning method, the feature representation matrix of each modality is mapped to a common subspace, so that the data of different modalities have a consistent representation in the subspace. Specifically, for each modality, a subspace mapping matrix is ​​designed, and the feature representation matrix of the corresponding modality is linearly transformed to the common subspace through the subspace mapping matrix to obtain the representation of the corresponding modality in the subspace. The parameters of the subspace mapping matrix need to be obtained through optimization.

[0096] In the common subspace, a preset consistent representation is introduced as the common representation target of different modal data in the subspace. By measuring the reconstruction error between the subspace representation of each modality and the consistent representation, the similarity between the subspace representation of each modality and the consistent representation is evaluated, and a similarity term is constructed. The reconstruction error is preferably calculated using the Euclidean distance, and the goal of the similarity term is to make the subspace representation of each modality as close to the consistent representation as possible.

[0097] In order to control the complexity of the subspace mapping matrix and avoid overfitting, the nuclear norm of the subspace mapping matrix is ​​introduced as a complexity regularization term in the optimization objective. The nuclear norm is the L1 norm of the singular values ​​of the matrix, which can measure the rank or complexity of the matrix. By minimizing the nuclear norm, a simpler and more robust subspace mapping matrix can be obtained.

[0098] In order to maintain the global manifold structure of the internal data of each modality, a global manifold structure preservation term is constructed. Specifically, for each pair of samples within the same modality, the weighted distance of each pair of samples in the original feature space is calculated, and the weighted distance relationship is introduced into the common subspace. It is required that the weighted distance in the subspace is as close as possible to the weighted distance in the original space, and the global manifold structure within the modality is maintained during the subspace learning process.

[0099] While maintaining the global manifold structure, in order to further maintain the local manifold structure within the modality, a local manifold structure preservation term is constructed. The K nearest neighbor algorithm is used to find the K most similar neighboring samples corresponding to each sample in each modality, and form similar sample pairs with the selected samples. Based on the similar sample pairs, their weighted distances in the common subspace are calculated, requiring the weighted distance in the subspace to be as close as possible to the weighted distance in the original space, and maintaining the local manifold structure within the modality during the subspace learning process.

[0100] By integrating the above four aspects, a unified error function is constructed as the optimization objective of multi-view subspace learning. The error function is a weighted combination of the similarity term, the complexity regularization term, the global manifold structure preservation term and the local manifold structure preservation term. By minimizing the error function, the optimal subspace mapping matrix and consistency representation can be obtained.

[0101] When solving the optimal subspace mapping matrix and consistent representation, an alternating optimization strategy is adopted. The consistent representation is fixed and the subspace mapping matrix is ​​optimized. Then the subspace mapping matrix is ​​fixed and the consistent representation is optimized. The alternating optimization steps are repeated until the preset number of iterations is reached or the preset convergence standard is reached. Early stopping is performed in advance to obtain the optimal subspace mapping matrix and consistent representation.

[0102] The obtained optimal subspace mapping matrix is ​​used to map the feature representation matrix of each modality to the common subspace to obtain the consistent optimal representation in the subspace. The consistent optimal representations of different modalities in the subspace are weightedly fused to generate intermediate fusion features.

[0103] In order to further explore the deep feature representation in the intermediate fusion features, an enhanced autoencoder network is constructed. The enhanced autoencoder network consists of an encoder subnetwork and a decoder subnetwork. The encoder subnetwork is stacked with multiple fully connected layers. The intermediate fusion features are mapped to a more abstract and advanced feature space through the encoder subnetwork; the decoder subnetwork is symmetrical with the encoder subnetwork and is used to reconstruct the original intermediate fusion features from the deep feature space. By reconstructing the input samples, the enhanced autoencoder network can adaptively learn the intrinsic deep feature representation of the data. The intermediate fusion features are input into the trained enhanced autoencoder network, and the output of its encoder subnetwork is extracted as the final multimodal fusion feature.

[0104] In this embodiment, through the preprocessing step, the data of different modalities are cleaned, normalized and feature extracted to ensure data quality and improve the accuracy of subsequent feature extraction; the multi-view subspace learning method is used to map the data of each modality to a common subspace, uniformly represent the data of different modalities, reduce the differences in data dimensions and scales, and improve the consistency and accuracy of feature extraction; by introducing the nuclear norm regularization term, the complexity of the subspace mapping matrix is ​​controlled to avoid model overfitting and enhance the robustness and generalization ability of the model; by maintaining the global and local manifold structure of each modal data, it is ensured that the data retains the original structural characteristics in the common subspace, Improve the consistency and robustness of the model on data of different modalities; introduce consistent representation as a common goal, optimize the feature fusion effect by measuring the reconstruction error between the subspace representation and the consistent representation, and make the data of different modalities have a consistent representation in the common subspace; gradually reduce the error function value by alternately optimizing the subspace mapping matrix and the consistent representation to ensure the efficiency of the optimization process and the optimality of the results; construct an enhanced autoencoder network, and through the encoder subnetwork and the decoder subnetwork, further explore the deep feature representation of the intermediate fusion features, enhance the abstraction and high-level level of the feature representation, and improve the quality of the final multimodal fusion features.

[0105] In an optional embodiment, constructing an error function according to the similarity term, the complexity regularization term, the global manifold structure preservation term and the local manifold structure preservation term includes:

[0106] The error function has the following formula:

[0107]

[0108] Among them, E sim represents the similarity term, m represents the modal ordinal number, M represents the total number of modalities, and Y m represents the subspace representation of the modality m in the common subspace, Y represents the consistency representation, ||·|| F represents the Frobenius norm, E complexity represents the complexity regularization term, W m represents the mapping matrix of mode m, ||·|| * represents the nuclear norm, E global represents the global manifold structure preservation term, tr(·) represents the diagonal elements of the corresponding matrix, and L m represents the Laplacian matrix of the graph corresponding to mode m, E local represents the local manifold structure preservation term, k represents the neighbor sequence number, K represents the total number of neighbors, and Y m (k) Indicates that Y mThe subspace corresponding to the nearest neighbor k is represented, E represents the error function, λ1 represents the weight coefficient of the complexity regularization term, λ2 represents the weight coefficient of the global manifold structure preservation term, and λ3 represents the weight coefficient of the local manifold structure preservation term.

[0109] According to the formula, the similarity term is used to ensure that the subspace representation of different modes is as close as possible to the consistency representation, so as to achieve a unified representation of multimodal data in a common subspace. The complexity of the subspace mapping matrix is ​​limited by the complexity regularization term, so that the model is more robust when dealing with new data and the generalization ability of the model is enhanced; the global manifold structure preservation term is used to maintain the global manifold structure of each modal data, so that the subspace representation can reflect the global characteristics of the original data and improve the effectiveness of feature extraction; the local manifold structure preservation term is used to maintain the local manifold structure of each modal data, so that the subspace representation can retain the local characteristics of the original data and further improve the accuracy of feature extraction; the optimal subspace mapping matrix and consistency representation are ensured by constructing a comprehensive error function and optimizing it, and finally high-quality multimodal fusion features are generated to improve the performance of DCS controller related tasks.

[0110] S102. Taking the multimodal fusion feature as input, an adaptive anomaly measurement algorithm is used to adaptively select cluster centers based on the product of local density and distance factor, and normal clusters and abnormal clusters are determined by clustering and calculating anomaly scores to obtain controller state abnormality measurement;

[0111] The local density specifically refers to the density of a data point in its local neighborhood, which is usually expressed by counting the number of neighbors of the point within a certain distance range. Data points with high local density are located in dense areas of data distribution, while data points with low local density are located in sparse areas. Local density is used to identify whether a data point is in the center or edge of a cluster.

[0112] The distance factor specifically refers to the distance between a data point and its neighbor or cluster center. The distance factor can help distinguish the degree of separation between different clusters. Data points with large distance factors are more likely to be located in the boundary area between different clusters. By comprehensively considering the distance factor and local density, the distribution of cluster centers and clusters can be determined more effectively.

[0113] The anomaly score specifically refers to a numerical value used to quantify the degree of abnormality of a data point, which is calculated based on the local density and distance factor of the data point. Data points with high anomaly scores have significantly different relationships with their neighbors and may be abnormal points or outliers. The calculation of anomaly scores can help identify potential abnormal situations or abnormal behaviors in the data.

[0114] The controller state abnormality metric specifically refers to an indicator or value used to evaluate whether the operating state of the DCS controller is abnormal. By analyzing the controller operation data, it is determined whether the working state of the controller is abnormal. The level of the abnormality metric is used to detect and diagnose potential faults or abnormal behaviors of the controller to ensure the stability and reliability of the system.

[0115] In this embodiment, by considering the local density and distance factor of the data points, the algorithm can adaptively select the cluster center and divide the data points into normal clusters and abnormal clusters, which helps to accurately identify normal and abnormal situations and improve the sensitivity and accuracy of the system to abnormal conditions; the adaptive anomaly measurement algorithm can dynamically select the cluster center according to the actual distribution of the data, without presetting fixed thresholds or parameters, so that the algorithm has strong adaptability and can handle different data characteristics and distribution situations; by comprehensively considering the local density and distance factors, the algorithm can more accurately determine the anomaly score, reduce the false alarm rate, help reduce the occurrence of false alarms, and improve the reliability and stability of the system; the use of the adaptive anomaly measurement algorithm can quickly process multimodal fusion features, and through clustering division and anomaly score calculation, it can timely discover and respond to potential abnormal situations. It helps to improve the real-time monitoring capability of the system and reduce the impact of abnormal conditions on system performance.

[0116] In an optional embodiment, the multimodal fusion feature is used as input, an adaptive anomaly measurement algorithm is used, a cluster center is adaptively selected based on the product of local density and distance factor, normal clusters and abnormal clusters are determined by clustering and calculating anomaly scores, and the controller state anomaly measurement is obtained, including:

[0117] Based on the multimodal fusion features, each multimodal fusion feature is taken as a sample point, a feature sample set is determined, the Euclidean distance between two sample points is calculated, a distance matrix is ​​generated, the median of all elements in the distance matrix is ​​selected as a cutoff distance threshold, and the local density of each sample point is calculated in combination with the distance matrix;

[0118] Iterate each sample point as the reference sample point, select all sample points with a local density greater than the reference sample point as comparison sample points, calculate the distance between each comparison sample point and the reference sample point, and select the minimum distance as the minimum distance factor; if the local density of the reference sample point is the largest, use all other sample points as comparison sample points, and select the maximum distance as the minimum distance factor;

[0119] Calculate the product of the local density corresponding to each sample point and the minimum distance factor, determine the multiplication factor of each sample point, arrange the multiplication factors in descending order, and calculate the multiplication factor difference between adjacent sample points;

[0120] Determine the median of the multiplication factor difference, determine the center threshold, extract the sample points whose multiplication factor difference is greater than the center threshold, and determine the cluster center;

[0121] Calculate the distances from all member sample points corresponding to non-cluster centers to each cluster center, assign the member sample points to the cluster center with the nearest distance, and generate clusters;

[0122] Based on each cluster, calculate the individual distance from each member sample point in the cluster to the cluster center, determine the maximum intra-cluster distance, and calculate the abnormal score of each member sample point based on the individual distance, maximum intra-cluster distance and local density;

[0123] All sample points are sorted by anomaly scores, and divided into normal clusters and abnormal clusters according to the preset anomaly score threshold. The controller state anomaly metric is determined based on the total number of sample points in the abnormal cluster and each anomaly score.

[0124] The cluster specifically refers to a subset of a group of data points with high similarity to each other in cluster analysis, and each cluster represents a natural aggregation area in the data set; the cluster uses an algorithm to group data points according to a certain similarity measure, so that the data points in the same cluster are as similar as possible, and the data points between different clusters are as different as possible;

[0125] The normal cluster specifically refers to a set of data points in the cluster whose characteristics show normal patterns and behaviors, usually with a lower anomaly score, indicating an expected operating state or behavioral characteristics. The normal cluster is used to represent the normal operating state of the system or the common characteristics of most data points.

[0126] The abnormal cluster specifically refers to a set of data points whose characteristics in the cluster show abnormal patterns and behaviors, usually with a high abnormality score, indicating that they deviate from the expected operating state or behavior characteristics. The abnormal cluster is used to identify abnormal operating states or abnormal behavior data points that may have problems in the system.

[0127] The features obtained by multimodal fusion are taken as input, and each multimodal fusion feature is regarded as a sample point in the feature space to form a feature sample set. The Euclidean distance between any two sample points in the feature sample set is calculated to generate a distance matrix. The distance matrix is ​​a symmetric matrix. For example, the element in the third row and the second column represents the Euclidean distance between the third sample point and the second sample point.

[0128] The median of all elements in the distance matrix is ​​selected as the cutoff distance threshold. For each sample point, the number of sample points whose distances between the corresponding sample point and other sample points are less than or equal to the cutoff distance threshold is counted to obtain the local density of the sample point. The local density reflects the compactness of the sample point in its neighborhood. The larger the local density, the more neighborhood sample points there are around the sample point, and the more likely the sample point is to be the cluster center.

[0129] Traverse each sample point, determine it as a reference sample point, and select all sample points with a local density greater than the reference sample point as comparison sample points. For each comparison sample point, calculate the distance between the comparison sample point and the reference sample point, and select the minimum distance value as the minimum distance factor of the reference sample point. The minimum distance factor reflects the minimum distance between the reference sample point and other sample points with a higher local density. If the local density of the reference sample point is the largest among all sample points, all other sample points are used as comparison sample points, and the maximum distance value is selected as the minimum distance factor.

[0130] For each sample point, the product of its local density and the minimum distance factor is calculated to obtain the multiplication factor. The multiplication factor comprehensively considers the local density of the sample point and the distance from other high-density sample points. The larger the multiplication factor, the higher the local density of the sample point and the farther away from other high-density sample points, and the more likely it is to become a cluster center.

[0131] Arrange all sample points in descending order according to the multiplication factor, and calculate the multiplication factor difference between adjacent sample points. The multiplication factor difference reflects the degree of change of the multiplication factor between sample points.

[0132] The median of the product factor difference is determined as the center threshold. Sample points whose product factor difference is greater than the center threshold are extracted as cluster centers. The product factor of the sample point corresponding to the cluster center is significantly higher than that of other sample points, indicating that the sample point corresponding to the cluster center has significant advantages in local density and distance, and is suitable as the center point of the cluster.

[0133] All sample points that are not cluster centers are regarded as member sample points, the distance from each member sample point to each cluster center is calculated, and the member sample points are assigned to the cluster cluster with the nearest cluster center. As described above, all sample points are divided into different cluster clusters according to the cluster center.

[0134] For each cluster, the individual distance from each member sample point in the cluster to the cluster center is calculated, and the maximum individual distance in the cluster is determined as the maximum intra-cluster distance. The maximum intra-cluster distance reflects the range and boundary of the cluster.

[0135] The anomaly score of each member sample point is calculated using individual distance, maximum intra-cluster distance and local density. The anomaly score comprehensively considers the relative position of the sample point within the cluster, the distance from the cluster center and the local density factors. The higher the anomaly score, the more likely the sample point is to be an anomaly.

[0136] All sample points are sorted according to the anomaly score, and according to the preset anomaly score threshold, the sample points with anomaly scores higher than the threshold are divided into anomaly clusters, and the remaining sample points are divided into normal clusters. The anomaly cluster contains sample points with abnormal controller status, and the normal cluster contains sample points with normal controller status.

[0137] According to the total number of sample points in the abnormal cluster and the abnormal score of each sample point, the abnormality measure of the controller state is determined. The abnormality measure is obtained by weighted average of the number of abnormal sample points and the degree of abnormality, which reflects the overall degree to which the controller state deviates from the normal state.

[0138] In this embodiment, by calculating the local density of each sample point, the dense area and the sparse area in the data set can be effectively distinguished, and the accuracy of anomaly detection can be improved; the product of the local density and the minimum distance factor is used to select the cluster center, so that the selected cluster center has significant advantages in both local density and distance, and the reliability of the clustering result is improved; the cluster center is adaptively selected based on the product of the local density and the minimum distance factor, which can avoid the problem of presetting the number and position of cluster centers in traditional clustering methods, making the selection of cluster centers more flexible and adaptive, and adapting to the actual distribution of data; by calculating the anomaly score and dividing the normal cluster and the abnormal cluster according to the preset anomaly score threshold, the normal samples and the abnormal samples of the controller state can be effectively distinguished, and the monitoring and diagnosis capabilities of the controller state can be improved; based on the total number of sample points in the abnormal cluster and the abnormal score of each sample point, the abnormal measurement of the controller state is determined by the weighted average method, providing a quantitative indicator that comprehensively reflects the degree of abnormality of the controller state, which is helpful for further analysis and decision-making.

[0139] In an optional embodiment, the minimum distance factor is expressed as follows:

[0140]

[0141] Among them, δ i represents the minimum distance factor of sample point i, i represents sample point i, j represents another sample point j, ρ i represents the local density of sample point i, ρ j Represents the local density of sample point j, max c ρ c represents the local density maximum, d ij Represents the distance between sample point i and sample point j;

[0142] The anomaly metric is formulated as follows:

[0143]

[0144] Among them, a represents the sample point in the abnormal cluster, X a represents the abnormality measure of sample point a, N represents the total number of all sample points, N A represents the total number of abnormal cluster sample points, A represents the abnormal cluster, s a Indicates the abnormal score of sample point a, max b∈A s b represents the maximum anomaly score in the anomaly cluster, q represents the anomaly score power index, and ρ a Represents the local density of sample point a.

[0145] Based on the formula, when the local density of a sample point is less than the maximum local density of all sample points, the closest distance value among the sample points with a larger local density than the corresponding sample point is selected as the minimum distance factor; when the local density of a sample point is equal to the maximum local density of all sample points, the maximum distance value between the corresponding sample point and all other sample points is selected as the minimum distance factor;

[0146] The anomaly metric is based on the average number of all sample points and the number of sample points in the anomaly cluster. The anomaly metric of each sample point is determined by the ratio of the corresponding anomaly score to the maximum anomaly score and the exponential decay of its local density relative to the maximum local density.

[0147] According to the formula, the minimum distance factor formula accurately identifies abnormal sample points in the data set by combining local density and the distance between sample points, thereby improving the accuracy of anomaly detection; by adaptively selecting cluster centers and calculating anomaly metrics, the algorithm can adapt to the characteristics and structures of different data sets, reduce dependence on parameter tuning, and improve the applicability and robustness of the algorithm; the anomaly metric formula comprehensively considers the anomaly score of the sample point and the local density factor, which can more comprehensively reflect the degree of abnormality of the sample point and more effectively identify abnormal conditions of the controller state.

[0148] S103. Based on the controller state abnormality measurement, it is mapped to language value features through a preset membership function, and an evaluation cloud is constructed by calculating mathematical statistical features. A comprehensive evaluation cloud is determined through weighted fusion, and a similarity match is performed with a preset trustworthy state level standard cloud to determine the current trustworthy state of the DCS controller.

[0149] The membership function specifically refers to the degree to which an element belongs to a fuzzy set, and its value is between 0 and 1; in the controller state anomaly measurement, the membership function maps the anomaly measurement value to a fuzzy set, corresponding to a language value feature, indicating the degree to which the state belongs to a specific language description.

[0150] The language value feature specifically refers to a fuzzy description of a numerical feature, using natural language to express the degree of a certain feature, for example, "extremely low degree of abnormality", "low degree of abnormality", "medium degree of abnormality", "high degree of abnormality", "extremely high degree of abnormality", etc.; the quantitative abnormality measurement is mapped to a qualitative language description through a membership function, making complex numerical information easier to understand and analyze.

[0151] The evaluation cloud specifically refers to a concept based on a cloud model, which is used to describe the evaluation results of uncertainty and fuzziness, and is usually constructed through mathematical statistical characteristics such as expected value, entropy, and hyperentropy. In the controller state evaluation, the evaluation cloud forms an overall evaluation result by weighted fusion of the mathematical statistical characteristics of each language value feature, reflecting the uncertainty and volatility of the system state.

[0152] The trust status level standard cloud specifically refers to a preset standard model, which is used to describe cloud models of different trust levels, usually including multiple levels, such as "high trust", "medium trust", "low trust", etc.; as a benchmark, the trust status of the current DCS controller is determined by similarity matching with the comprehensive evaluation cloud. The standard cloud provides reference standards for different trust states, making the evaluation results comparable and referenceable.

[0153] In this embodiment, through the use of membership functions, complex numerical information is converted into intuitive language descriptions, enhancing the interpretability and comprehensibility of the results; the language value feature combines fuzzy descriptions and quantitative analysis to make the evaluation results closer to the actual situation, which helps users make intuitive judgments and decisions; the evaluation cloud model can effectively process and express the uncertainty of the system state, comprehensively reflect the overall state of the system through mathematical statistical features, and improve the accuracy and comprehensiveness of the evaluation; the standard cloud provides a unified evaluation standard, and through similarity matching with the comprehensive evaluation cloud, it can accurately determine the current trusted state of the controller, ensuring the reliability and consistency of the evaluation results.

[0154] In an optional embodiment, based on the controller state abnormality measurement, a preset membership function is used to map it to a language value feature, and an evaluation cloud is constructed by calculating mathematical statistical features. After weighted fusion, a comprehensive evaluation cloud is determined, and similarity matching is performed with a preset trustworthy state level standard cloud. Determining the current trustworthy state of the DCS controller includes:

[0155] Mapping the controller state abnormality metric to a preset language value feature set through a pre-constructed Gaussian membership function, calculating the membership of the abnormality metric to each language value feature, and determining a membership vector;

[0156] For each language value feature, based on the membership vector, the membership-weighted anomaly metric mean is calculated, the expectation is determined, the membership-weighted anomaly metric standard deviation is calculated, the entropy is determined, based on a preset proportionality coefficient, the product of the entropy and the proportionality coefficient is calculated, the hyperentropy is determined, and based on the expectation, the entropy and the hyperentropy, an evaluation cloud is constructed;

[0157] The weighted average cloud operation rule is adopted to weight the expectation, entropy and super entropy of each evaluation cloud to generate a comprehensive evaluation cloud of the overall trustworthy state of the controller;

[0158] Based on a set of pre-constructed credible status grade standard clouds, the similarity between the comprehensive evaluation cloud and each grade standard cloud is calculated, the adaptation grade standard cloud corresponding to the maximum similarity is extracted, and based on the credible status corresponding to the adaptation grade standard cloud, the current credible status of the DCS controller is determined.

[0159] The controller state anomaly measurement obtained previously is used as input and processed by a pre-constructed Gaussian membership function. The Gaussian membership function is a commonly used fuzzy membership function that can map the anomaly measurement to a preset language value feature set. The language value feature is a qualitative description of the degree of anomaly. For each language value feature, the membership of the anomaly measurement to each language value feature is calculated using the Gaussian membership function to obtain a membership vector. The membership vector represents the degree of membership of the anomaly measurement on each language value feature.

[0160] For each language value feature, further processing is performed using the membership vector. The membership-weighted anomaly metric mean is calculated and used as the expectation of the evaluation cloud corresponding to the language value feature. The expectation reflects the central position of the anomaly metric on the corresponding language value feature. The membership-weighted anomaly metric standard deviation is calculated and the weighted anomaly metric standard deviation is used as the entropy of the evaluation cloud of the corresponding language value feature. The entropy reflects the degree of discreteness or uncertainty of the anomaly metric on the corresponding language value feature. Based on a preset proportional coefficient, the product of the entropy and the proportional coefficient is calculated and used as the super entropy of the evaluation cloud of the corresponding language value feature. The super entropy reflects the uncertainty or fuzziness of the evaluation cloud itself. The evaluation cloud corresponding to each language value feature is constructed through the three digital features of expectation, entropy and super entropy.

[0161] The cloud operation rule of weighted average is adopted to fuse the evaluation clouds corresponding to all language value features. Specifically, the expectation, entropy and super entropy of each evaluation cloud are weighted averaged to obtain the expectation, entropy and super entropy of the comprehensive evaluation cloud. The weight of the weighted average can be pre-set according to the importance or credibility of different language value features. The comprehensive evaluation cloud represents the fuzzy distribution characteristics of the overall trustworthy state of the controller.

[0162] The comprehensive evaluation cloud is matched and evaluated using a pre-built set of trusted state level standard clouds. The level standard cloud is pre-built based on domain knowledge or historical data, and each level standard cloud corresponds to a specific trusted state level. The similarity between the comprehensive evaluation cloud and each level standard cloud is calculated. The similarity can be measured using Euclidean distance, cosine similarity, etc. Preferably, Euclidean distance is used. The level standard cloud with the greatest similarity to the comprehensive evaluation cloud is selected, and its corresponding trusted state level is used as the current trusted state of the DCS controller.

[0163] In this embodiment, the Gaussian membership function is used to map complex numerical anomaly metrics to a preset language value feature set, which improves the interpretability and comprehensibility of the data and helps to detect and identify anomalies more accurately; based on the membership-weighted anomaly metric mean, standard deviation and hyperentropy, the anomaly features are fully described, not only by considering the central trend through expectation, but also by considering the dispersion and uncertainty of the data through entropy and hyperentropy, providing a comprehensive anomaly assessment; the cloud operation rule of weighted average is adopted to fuse the expectation, entropy and hyperentropy of different evaluation clouds to generate a comprehensive evaluation cloud of the overall trustworthy state. Through the fusion of multiple features, the evaluation results are ensured to be more accurate and reliable; by calculating the similarity between the comprehensive evaluation cloud and the preset trustworthy state level standard cloud, the adaptation level standard cloud corresponding to the maximum similarity is extracted to determine the current trustworthy state. This method provides a clear level division, which is convenient for users to understand and use; by adjusting the preset proportional coefficient, it can flexibly adapt to different application scenarios and needs, and improve the adaptability and flexibility of the system.

[0164] In an optional embodiment, based on the expectation, the entropy and the super entropy, constructing an evaluation cloud includes:

[0165] The expectation is as follows:

[0166]

[0167] Among them, B o represents the expectation of the linguistic value feature o, ω oa Indicates the weight corresponding to the sample point a mapping the language value feature o, u oa Indicates the membership of the abnormal measure of sample point a to the language value feature o;

[0168] The entropy is formulated as follows:

[0169]

[0170] Among them, T o The entropy of the linguistic value feature o;

[0171] The super entropy is formulated as follows:

[0172] H o =η1·(T o ) r +η2·log(T o +1);

[0173] Among them, H o It represents the super entropy of the language value feature o, η1 represents the entropy influence degree control coefficient, r represents the entropy amplification degree index, and η2 represents the logarithmic term influence degree coefficient.

[0174] According to the formula, expectation represents the expectation of each language value feature, which is obtained by multiplying the anomaly measure of each sample point by the weight and membership degree of the sample point mapped to the language value feature, and then performing weighted averaging; the expectation of the anomaly measure is calculated by weighted averaging to accurately reflect the central trend of each language value feature, which helps to describe the abnormal state more accurately; entropy represents the entropy of the language value feature, which is obtained by calculating the fourth power of the difference between the anomaly measure and the expectation of each sample point, multiplying it by the membership degree, and then performing weighted averaging and taking the fourth root; by calculating the fourth power of the difference between the anomaly measure and the expectation and taking the fourth root, the dispersion of the data is comprehensively measured to capture the extensiveness and variability of the abnormal data; super entropy represents the super entropy of the language value feature, which is obtained by multiplying the entropy by an amplification coefficient and an exponent and adding the influence of the logarithmic term; by introducing the amplification of entropy and the influence of the logarithmic term, the complexity and uncertainty of the anomaly measure are comprehensively considered, the sensitivity and robustness of the model to abnormal data are enhanced, and the instability caused by the fluctuation of a single feature is avoided; by comprehensively considering expectation, entropy and super entropy, all aspects of the controller state are comprehensively evaluated to ensure the comprehensiveness and accuracy of the evaluation results, which helps to timely discover potential anomalies and problems.

[0175] Figure 2 FIG. 1 is a schematic diagram of the structure of a DCS controller trusted state perception system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0176] The first unit is used to obtain multimodal heterogeneous data in multiple time windows of the DCS controller, use a multi-view subspace learning method to extract features from the multimodal heterogeneous data in each time window, map different modal data to a common subspace, and obtain corresponding multimodal fusion features;

[0177] The second unit is used to use the multimodal fusion feature as input, adopt an adaptive anomaly measurement algorithm, adaptively select cluster centers based on the product of local density and distance factor, determine normal clusters and abnormal clusters through clustering and calculating anomaly scores, and obtain controller state abnormality measurement;

[0178] The third unit is used to map the controller state abnormality measurement into language value features through a preset membership function, construct an evaluation cloud by calculating mathematical statistical features, determine the comprehensive evaluation cloud through weighted fusion, and perform similarity matching with the preset trustworthy state level standard cloud to determine the current trustworthy state of the DCS controller.

[0179] According to a third aspect of the embodiments of the present invention,

[0180] An electronic device is provided, comprising:

[0181] processor;

[0182] a memory for storing processor-executable instructions;

[0183] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0184] A fourth aspect of the embodiments of the present invention is:

[0185] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0186] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0187] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A DCS controller trusted state perception method, characterized in that: include: Acquire multimodal heterogeneous data in multiple time windows of the DCS controller, use multi-view subspace learning method to extract features from the multimodal heterogeneous data in each time window, map different modal data to the common subspace, and obtain corresponding multimodal fusion features; Taking the multimodal fusion features as input, an adaptive anomaly measurement algorithm is adopted to adaptively select cluster centers based on the product of local density and distance factor, and normal clusters and abnormal clusters are determined by clustering and calculating anomaly scores to obtain controller state abnormality measurement; Based on the controller state abnormality measurement, it is mapped to a language value feature through a preset membership function, and an evaluation cloud is constructed by calculating mathematical statistical features. After weighted fusion, a comprehensive evaluation cloud is determined, and similarity matching is performed with a preset trustworthy state level standard cloud to determine the current trustworthy state of the DCS controller; The multi-view subspace learning method is used to extract features from the multimodal heterogeneous data in each time window, and the different modal data are mapped to the common subspace to obtain the corresponding multimodal fusion features including: Preprocess the multi-modal heterogeneous operation data separately to determine the feature representation matrix corresponding to each mode; Based on the multi-view subspace learning method, the feature representation matrix of each modality is mapped to a common subspace through a subspace mapping matrix to generate a subspace representation; Based on the reconstruction error between the subspace representation of each modality and the preset consistency representation, measure the similarity between the subspace representation and the consistency representation, and determine the similarity term; based on the nuclear norm of the subspace mapping matrix, control the complexity of the subspace mapping matrix, and determine the complexity regularization term; based on the weighted distance of sample pairs in the same modality, maintain the global manifold structure, and determine the global manifold structure preservation term; based on the K nearest neighbor algorithm, determine similar sample pairs in the same modality, and based on the weighted distance of the similar sample pairs in the common subspace, maintain the local manifold structure, and determine the local manifold structure preservation term; construct an error function based on the similarity term, the complexity regularization term, the global manifold structure preservation term, and the local manifold structure preservation term; Based on the alternating optimization strategy, the subspace mapping matrix is ​​optimized and the consistency representation is adjusted to minimize the error function, and the model parameters are iteratively updated until a preset number of iterations is reached to obtain the optimal subspace mapping matrix; Based on the optimal subspace mapping matrix, the feature representation matrix of each modality is mapped to a common subspace to obtain a consistent optimal subspace representation; Performing weighted fusion on the consistent optimal subspace representation to determine an intermediate fusion feature; Based on multiple fully connected layers, an encoder sub-network is constructed, based on the encoder sub-network, a symmetric network structure is determined, and a decoder sub-network is constructed, and the encoder sub-network is combined with the decoder sub-network to construct an enhanced autoencoder network; The intermediate fusion features are input into the enhanced autoencoder network, which reconstructs the input, learns the deep feature representation, and outputs the final multimodal fusion features.

2. The method according to claim 1, characterized in that According to the similarity term, the complexity regularization term, the global manifold structure preservation term and the local manifold structure preservation term, constructing an error function comprises: The error function has the following formula: Among them, E sim represents the similarity term, m represents the modal ordinal number, M represents the total number of modalities, and Y m represents the subspace representation of the modality m in the common subspace, Y represents the consistency representation, ||·|| F represents the Frobenius norm, E complexity represents the complexity regularization term, W m represents the mapping matrix of mode m, ||·||* represents the nuclear norm, E global represents the global manifold structure preservation term, tr(·) represents the diagonal elements of the corresponding matrix, and L m represents the Laplacian matrix of the graph corresponding to mode m, E local represents the local manifold structure preservation term, k represents the neighbor sequence number, K represents the total number of neighbors, and Y m (k) Indicates that Y m The subspace corresponding to the nearest neighbor k is represented, E represents the error function, λ1 represents the weight coefficient of the complexity regularization term, λ2 represents the weight coefficient of the global manifold structure preservation term, and λ3 represents the weight coefficient of the local manifold structure preservation term.

3. The method according to claim 1, characterized in that Taking the multimodal fusion features as input, an adaptive anomaly measurement algorithm is used to adaptively select cluster centers based on the product of local density and distance factor, and normal clusters and abnormal clusters are determined by clustering and calculating anomaly scores. The controller state anomaly measurement includes: Based on the multimodal fusion features, each multimodal fusion feature is taken as a sample point, a feature sample set is determined, the Euclidean distance between two sample points is calculated, a distance matrix is ​​generated, the median of all elements in the distance matrix is ​​selected as a cutoff distance threshold, and the local density of each sample point is calculated in combination with the distance matrix; Iterate each sample point as the reference sample point, select all sample points with a local density greater than the reference sample point as comparison sample points, calculate the distance between each comparison sample point and the reference sample point, and select the minimum distance as the minimum distance factor; if the local density of the reference sample point is the largest, use all other sample points as comparison sample points, and select the maximum distance as the minimum distance factor; Calculate the product of the local density corresponding to each sample point and the minimum distance factor, determine the multiplication factor of each sample point, arrange the multiplication factors in descending order, and calculate the multiplication factor difference between adjacent sample points; Determine the median of the multiplication factor difference, determine the center threshold, extract sample points whose multiplication factor difference is greater than the center threshold, and determine the cluster center; Calculate the distances from all member sample points corresponding to non-cluster centers to each cluster center, assign the member sample points to the cluster center with the nearest distance, and generate clusters; Based on each cluster, calculate the individual distance from each member sample point in the cluster to the cluster center, determine the maximum intra-cluster distance, and calculate the abnormal score of each member sample point based on the individual distance, maximum intra-cluster distance and local density; All sample points are sorted by anomaly scores, and divided into normal clusters and abnormal clusters according to the preset anomaly score threshold. The controller state anomaly metric is determined based on the total number of sample points in the abnormal cluster and each anomaly score.

4. The method according to claim 3, characterized in that The minimum distance factor is formulated as follows: Among them, δ i represents the minimum distance factor of sample point i, i represents sample point i, j represents another sample point j, ρ i represents the local density of sample point i, ρ j Represents the local density of sample point j, max c ρ c represents the local density maximum, d ij Represents the distance between sample point i and sample point j; The anomaly metric is formulated as follows: Among them, a represents the sample point in the abnormal cluster, X a represents the abnormality measure of sample point a, N represents the total number of all sample points, N A represents the total number of abnormal cluster sample points, A represents the abnormal cluster, s a Indicates the abnormal score of sample point a, max b∈A s b represents the maximum anomaly score in the anomaly cluster, q represents the anomaly score power index, and ρ a Represents the local density of sample point a.

5. The method according to claim 4, characterized in that Based on the controller state abnormality measurement, the preset membership function is used to map it to a language value feature, and the evaluation cloud is constructed by calculating the mathematical statistical features. The comprehensive evaluation cloud is determined by weighted fusion, and the similarity is matched with the preset trustworthy state level standard cloud to determine the current trustworthy state of the DCS controller, including: Mapping the controller state abnormality metric to a preset language value feature set through a pre-constructed Gaussian membership function, calculating the membership of the abnormality metric to each language value feature, and determining a membership vector; For each language value feature, based on the membership vector, the membership-weighted anomaly metric mean is calculated, the expectation is determined, the membership-weighted anomaly metric standard deviation is calculated, the entropy is determined, based on a preset proportionality coefficient, the product of the entropy and the proportionality coefficient is calculated, the hyperentropy is determined, and based on the expectation, the entropy and the hyperentropy, an evaluation cloud is constructed; The weighted average cloud operation rule is adopted to weight the expectation, entropy and super entropy of each evaluation cloud to generate a comprehensive evaluation cloud of the overall trustworthy state of the controller; Based on a set of pre-constructed credible status grade standard clouds, the similarity between the comprehensive evaluation cloud and each grade standard cloud is calculated, the adaptation grade standard cloud corresponding to the maximum similarity is extracted, and based on the credible status corresponding to the adaptation grade standard cloud, the current credible status of the DCS controller is determined.

6. The method according to claim 5, characterized in that Based on the expectation, the entropy and the super entropy, constructing an evaluation cloud comprises: The expectation is as follows: Among them, B o represents the expectation of the linguistic value feature o, ω oa Indicates the weight corresponding to the sample point a mapping the language value feature o, u oa Indicates the membership of the abnormal measure of sample point a to the language value feature o; The entropy is formulated as follows: Among them, T o The entropy of the linguistic value feature o; The super entropy is formulated as follows: H o =η1.(T o ) T +η2·log(T o +1); Among them, H o It represents the super entropy of the language value feature o, η1 represents the entropy influence degree control coefficient, r represents the entropy amplification degree index, and η2 represents the logarithmic term influence degree coefficient.

7. A DCS controller trusted state perception system, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain multimodal heterogeneous data in multiple time windows of the DCS controller, use a multi-view subspace learning method to extract features from the multimodal heterogeneous data in each time window, map different modal data to a common subspace, and obtain corresponding multimodal fusion features; The second unit is used to use the multimodal fusion feature as input, adopt an adaptive anomaly measurement algorithm, adaptively select cluster centers based on the product of local density and distance factor, determine normal clusters and abnormal clusters through clustering and calculating anomaly scores, and obtain controller state abnormality measurement; The third unit is used to map the controller state abnormality measurement into language value features through a preset membership function, construct an evaluation cloud by calculating mathematical statistical features, determine the comprehensive evaluation cloud through weighted fusion, and perform similarity matching with the preset trustworthy state level standard cloud to determine the current trustworthy state of the DCS controller.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images