Dcs early warning method and system based on intelligent ai visual recognition

Abstract

The application provides a DCS early warning method and system based on intelligent AI visual identification, relates to the technical field of DCS early warning, and comprises the following steps: acquiring real-time operation parameters of a DCS system to obtain real-time operation parameter characteristics in a unified format; inputting the real-time operation parameter characteristics into a pre-constructed intelligent AI visual identification model to analyze the operation state of the DCS system, identify abnormal modes at the system level, and compare the abnormal modes with normal operation modes in a preset normal operation mode library; adopting a digital twin simulation technology based on a physical mechanism to predict fault modes, fault causes and occurrence probabilities of the DCS system, and automatically generating early warning information of corresponding levels according to the prediction results and a preset multi-level early warning decision matrix, so that the early warning information is pushed to an intelligent operation and maintenance platform of the DCS system in real time, system risks are identified in advance by an operator, and specific preventive maintenance measure suggestions are given.

Description

Technical Field

[0001] The present invention relates to DCS early warning technology, and in particular to a DCS early warning method and system based on intelligent AI visual recognition. Background Art

[0002] Distributed control systems (DCSs) are widely used in industries such as petroleum, chemical engineering, and electric power, enabling real-time monitoring and control of production processes. However, DCS systems are large in scale, complex in structure, operate in harsh environments, and present a high risk of failure. Failures can cause significant economic losses and safety incidents. Traditional DCS early warning methods rely primarily on manual experience and simple threshold judgments, making them inadequate for increasingly complex industrial scenarios and unable to guarantee the accuracy and real-time nature of early warnings. In recent years, the development of artificial intelligence (AI) technology has provided new avenues for intelligent early warning in DCS systems. Some research has attempted to apply machine learning algorithms, such as support vector machines and neural networks, to DCS fault diagnosis and prediction, achieving some success. However, these methods primarily focus on single devices or subsystems, lacking a macroscopic analysis of the entire DCS system, and thus struggle to detect system-level anomalies. Furthermore, the complexity and variability of industrial big data pose challenges to data processing and feature extraction for intelligent early warning.

[0003] Therefore, there is an urgent need for a new DCS system early warning method that can make full use of industrial big data and artificial intelligence technology to identify abnormal patterns at the system level, and combine digital twin simulation to predict failure risks, generate intelligent early warning information, and assist operation and maintenance personnel to make preventive maintenance decisions in a timely manner, thereby improving the safety and reliability of the DCS system. Summary of the Invention

[0004] The embodiments of the present invention provide a DCS early warning method and system based on intelligent AI visual recognition, which can solve the problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention,

[0006] Provides a DCS early warning method based on intelligent AI visual recognition, including:

[0007] Acquiring real-time operating parameters of the DCS system, including operating status data, alarm information, operation logs, and video monitoring data of each subsystem of the DCS system, wherein the operating status data includes real-time process parameters and equipment operating parameters of each subsystem, and preprocessing the real-time operating parameters to obtain real-time operating parameter features in a unified format;

[0008] Input the real-time operating parameter features into a pre-built intelligent AI visual recognition model, analyze the operating status of the DCS system, identify abnormal patterns at the system level, compare the abnormal patterns with normal operating patterns in a preset normal operating pattern library, and calculate the similarity between the abnormal patterns and the normal operating patterns;

[0009] If the similarity is lower than the preset abnormality threshold, it is determined that there is an abnormal risk in the current operation of the DCS system. The digital twin simulation technology based on physical mechanisms is used to simulate the real-time operation scenario of the DCS system to predict the failure mode, failure cause and occurrence probability of the DCS system. Based on the prediction results and the preset multi-level early warning decision matrix, the corresponding level of early warning information is automatically generated and pushed to the intelligent operation and maintenance platform of the DCS system in real time to assist operators in identifying system risks in advance and provide targeted preventive maintenance measures.

[0010] In an optional embodiment,

[0011] Inputting the real-time operating parameter features into a pre-built intelligent AI visual recognition model to analyze the operating status of the DCS system and identify abnormal patterns at the system level include:

[0012] A multimodal feature fusion network is designed in the intelligent AI visual recognition model. Text mining is performed on alarm information in real-time operation parameter features to obtain alarm information features. Sequential pattern mining is performed on operation logs to obtain operation log features. Target detection, target tracking, and behavior recognition are performed on video surveillance data to obtain video surveillance features.

[0013] The extracted alarm information features, operation log features, and video surveillance features are input into a multimodal feature fusion network, and each feature is mapped to a latent space of the same dimension through a feature mapping method.

[0014] The attention mechanism is used to calculate the similarity between features in the latent space, assign weights to each feature, and use the hybrid expert network to fuse the features in the latent space to obtain fused features. The fused features are then compressed using the autoencoder to obtain the final multimodal fused features.

[0015] A heterogeneous graph is constructed based on the multimodal fusion features. The nodes in the heterogeneous graph include devices, parameters, and alarms, and the edges include physical connections, logical associations, and temporal dependencies. A graph neural network is applied to the heterogeneous graph to iteratively update the hidden states of the nodes through a message passing mechanism. In each propagation step, the node first aggregates messages from neighboring nodes and then updates the node hidden state based on its own features and neighbor information.

[0016] For the local anomaly detection task, a graph decoder is used to map the node hidden state into anomaly metrics, and the graph neural network corresponding to the local anomaly detection task is trained by optimizing the cross-entropy loss function between the anomaly metrics and the true anomaly labels of the nodes. For the global anomaly association task, a graph pooling method based on anomaly metrics is used to gradually aggregate node information to generate an abnormal region hierarchy, and a graph decoder is used to map the abnormal region features into a global anomaly metric. The graph neural network corresponding to the global anomaly association task is trained by optimizing the cross-entropy loss function between the global anomaly metric and the true anomaly labels at the system level.

[0017] The two subtasks of local anomaly detection and global anomaly association share the same graph encoder, and achieve joint optimization through alternating training and parameter sharing. The outputs of the two subtasks are combined to finally output the system-level anomaly pattern recognition results.

[0018] In an optional embodiment,

[0019] The two subtasks, local anomaly detection and global anomaly association, share the same graph encoder. Joint optimization is achieved through alternating training and parameter sharing. The outputs of the two subtasks are combined to produce the following system-level anomaly pattern recognition results:

[0020] A joint optimization loss function is constructed based on the loss function of the local anomaly detection subtask and the loss function of the global anomaly association subtask. The training steps of the local anomaly detection and global anomaly association subtasks are repeated until the joint optimization loss function converges and the system-level anomaly pattern recognition result is obtained. The formula of the joint optimization loss function is as follows:

[0021]

[0022] Among them, L represents the joint optimization loss function, λ1 represents the weight of the local anomaly detection task, V represents the number of nodes, α i represents the weight factor of node i, represents the predicted abnormality metric of node i, y i represents the true anomaly label of node i, γ represents the exponential parameter of anomaly measurement error, λ2 represents the weight of the global anomaly association task, C represents the number of operation log features, α c represents the weight factor of neighbor node c, y c represents the true abnormal label of the neighbor node c, represents the predicted abnormality metric of neighbor node c, λ3 represents the weight of the regularization term, C i represents the set of neighbor nodes of node i, and p represents the exponential parameter of the regularization term.

[0023] In an optional embodiment,

[0024] Using digital twin simulation technology based on physical mechanisms, we simulate the real-time operation scenarios of the DCS system and predict the failure modes, causes, and occurrence probabilities of the DCS system, including:

[0025] Build a multidisciplinary, high-fidelity digital twin model of the DCS system, which includes a production process model, equipment model, and control logic model;

[0026] Extract key feature parameters of abnormal patterns, use feature dimensionality reduction methods to reduce the dimensions of key feature parameters, and use semantic mapping technology to map the reduced key feature parameters to the corresponding positions and object attributes of the digital twin model, triggering the digital twin model to perform simulation predictions under abnormal working conditions;

[0027] The simulation prediction starts from the anomaly point, adopts a multi-scenario parallel simulation strategy, sets different initial conditions, boundary conditions and random disturbances, generates multiple parallel simulation instances, and adopts model simplification and multi-scale modeling methods to selectively simplify the digital twin model at different spatiotemporal scales, performs grid encryption and computational intensification for key areas and key processes, and introduces uncertainty quantification technology to adaptively adjust the spatiotemporal resolution and computational accuracy of the simulation;

[0028] Based on the simulation of multi-scenario abnormal working conditions, with the abnormal pattern as the top event, deductive reasoning and causal reverse tracing methods are used to automatically construct a fault tree, identify the key links and influencing factors of abnormal propagation, and use the Monte Carlo simulation method to randomly sample the probability of occurrence of basic events in the fault tree and evaluate the abnormal pattern.

[0029] According to the results of fault tree analysis and Monte Carlo simulation, a case-based reasoning fault diagnosis method is adopted to measure the similarity between abnormal patterns and historical fault cases, match fault patterns from the historical fault library, and finally output the predicted fault mode, fault cause and occurrence probability of the DCS system.

[0030] In an optional embodiment,

[0031] Using a multi-scenario parallel simulation strategy, setting different initial conditions, boundary conditions, and random perturbations, multiple parallel simulation instances are generated, including:

[0032] Obtain historical anomaly data from the DCS system for preprocessing, extract multi-scale spatiotemporal features of the historical anomaly data, including anomaly features at different time and spatial scales, and construct a multi-scale spatiotemporal feature representation of the anomaly scene;

[0033] Inputting the multi-scale spatiotemporal feature representation of the abnormal scene into a multi-scale conditional generative adversarial network, wherein the multi-scale conditional generative adversarial network includes a generator and a discriminator, wherein the generator receives random noise and conditional variables as input, and the discriminator receives the real abnormal scene features and the abnormal scene features generated by the generator as input;

[0034] In the generator, multi-scale feature extraction is performed on random noise and conditional variables to obtain noise features and conditional features of different scales. The attention mechanism is used to adaptively fuse the noise features and conditional features of different scales to generate abnormal scene features that meet the conditional constraints. In the discriminator, multi-scale feature extraction is performed on the input real abnormal scene features and the generated abnormal scene features to obtain real features and generated features of different scales. The attention mechanism is used to adaptively fuse the real features and generated features of different scales.

[0035] Constructing the objective function of a multi-scale conditional generative adversarial network, which includes the logarithmic probability of the discriminator's true abnormal scene features and the false probability of the abnormal scene features generated by the generator. Optimization training is performed by maximizing the discriminator's objective function and minimizing the generator's objective function, ultimately obtaining a trained abnormal scene distribution model.

[0036] Setting an abnormality type and severity condition for an abnormal scenario, encoding the abnormality type and severity condition into a conditional vector, concatenating the conditional vector with a random noise vector to form a conditional noise vector, inputting the conditional noise vector into a generator of an abnormal scenario distribution model, and generating an abnormal scenario that meets the abnormality type and severity conditions through the generator;

[0037] Taking the diversity, coverage and abnormality degree of the abnormal scenarios as the optimization goals, a fitness function is constructed, an adaptive sampling strategy is adopted to optimize the sampling of abnormal scenarios, and the sampling parameters are adaptively adjusted to obtain an optimized set of abnormal scenarios. The optimized set of abnormal scenarios is then input into the digital twin model for multi-scenario parallel simulation.

[0038] In an optional embodiment,

[0039] The calculation formula for the objective function of constructing a multi-scale conditional generative adversarial network is as follows:

[0040]

[0041] in, represents the objective function of the multi-scale conditional generative adversarial network, R(D, G) represents the value function between the generator G and the discriminator D, and E f [·] represents the expectation of the characteristics of the real abnormal scene, E z[·] represents the expectation of the generated synthetic abnormal scene features, J represents the time scale level, K represents the spatial scale level, and a (D) represents the attention weight in the discriminator, f (j，k) Multi-scale representation of real abnormal scene features, s (j，k) represents the multi-scale representation of the conditional variable, a (G) represents the attention weight in the generator, z (j，k) Representing multi-scale representations of random noise.

[0042] In an optional embodiment,

[0043] According to the prediction results and the preset multi-level warning decision matrix, the corresponding level of warning information is automatically generated, including:

[0044] Construct an abnormal risk warning domain ontology, define the relationship mapping between abnormality types, influencing factors, warning levels and response measures, and form an abnormal risk warning ontology model;

[0045] Integrate structured data, semi-structured data, and unstructured data, realize the association of multi-source heterogeneous data through data integration, and use the knowledge extraction model of deep learning to extract entities, attributes, and relationships from unstructured data to form an abnormal risk warning knowledge graph;

[0046] Taking the anomaly type and influencing factors output by the anomaly risk prediction as known conditions, matching the corresponding entity nodes in the consistent anomaly risk warning knowledge graph, and applying the anomaly risk warning ontology reasoning rules and graph reasoning algorithms to infer the concepts, attributes, and relationships associated with the known conditions, generating an anomaly risk warning reasoning explanation;

[0047] Using the knowledge subgraph obtained from the abnormal risk warning reasoning interpretation, an abnormal risk warning question and answer template is constructed to form a structured question description. The question description is input into the Seq2Seq-based natural language generation model to automatically generate a text of recommended abnormal risk response measures;

[0048] Based on the confidence, coverage and novelty indicators of abnormal risk response measures knowledge, a knowledge sorting algorithm is used to prioritize the suggested texts of abnormal risk response measures. The abnormal risk level is calculated according to the abnormal prediction results, and the abnormal risk response measures of the corresponding level are selected to automatically generate early warning information of the corresponding abnormal risk level.

[0049] According to a second aspect of the embodiments of the present invention,

[0050] Provides a DCS early warning system based on intelligent AI visual recognition, including:

[0051] The first unit is used to obtain real-time operating parameters of the DCS system, including operating status data, alarm information, operation logs, and video monitoring data of each subsystem of the DCS system. The operating status data includes real-time process parameters and equipment operating parameters of each subsystem, and pre-process the real-time operating parameters to obtain real-time operating parameter features in a unified format;

[0052] The second unit is used to input the real-time operating parameter features into a pre-built intelligent AI visual recognition model, analyze the operating status of the DCS system, identify abnormal modes at the system level, compare the abnormal modes with normal operating modes in a preset normal operating mode library, and calculate the similarity between the abnormal modes and the normal operating modes;

[0053] The third unit is used to determine that there is an abnormal risk in the current operation of the DCS system if the similarity is lower than the preset abnormal threshold. It uses digital twin simulation technology based on physical mechanisms to simulate the real-time operation scenario of the DCS system, predict the failure mode, cause and probability of the DCS system, and automatically generate warning information of the corresponding level based on the prediction results combined with the preset multi-level warning decision matrix. The warning information is pushed to the intelligent operation and maintenance platform of the DCS system in real time to assist operators in identifying system risks in advance and provide targeted preventive maintenance measures.

[0054] According to a third aspect of the embodiments of the present invention,

[0055] An electronic device is provided, comprising:

[0056] processor;

[0057] a memory for storing processor-executable instructions;

[0058] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

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

[0060] 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 method described above is implemented.

[0061] In this embodiment, an intelligent AI visual recognition model identifies abnormal patterns at the system level, avoiding the limitations of traditional methods based on manual experience and threshold judgment, and achieving comprehensive perception and analysis of the DCS system's operating status. Digital twin simulation technology, combined with physical mechanisms, predicts and verifies identified abnormal risks, providing failure modes, causes, and probability of occurrence. This provides a reliable basis for early warning decision-making and improves the interpretability and credibility of early warning information. A multi-level early warning decision matrix is ​​introduced to automatically generate corresponding levels of early warning information based on the severity and urgency of the abnormal risk. This enables intelligent grading and dynamic adjustment of early warning information, avoiding insufficient or excessive early warnings and improving the relevance and effectiveness of early warnings. Through real-time integration with the intelligent operation and maintenance platform, early warning information is pushed to operators and preventive maintenance measures are recommended. This ensures timely transmission and resolution of early warning information, shortening the duration of DCS system abnormalities and reducing failure risks and losses. The overall solution integrates emerging technologies such as industrial big data, artificial intelligence, and digital twins to establish a complete intelligent early warning closed loop for DCS systems. This highly innovative and practical solution is of great significance for improving the operational safety and maintenance efficiency of industrial control systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 Schematic diagram of the flow of a DCS early warning method based on intelligent AI visual recognition according to an embodiment of the present invention;

[0063] Figure 2 This is a structural diagram of a DCS early warning system based on intelligent AI visual recognition according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying 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 making creative efforts shall fall within the scope of protection of the present invention.

[0065] The following specific embodiments are used to describe the technical solution of the present invention in detail. 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.

[0066] Figure 1 FIG. 1 is a flow chart of a DCS early warning method based on intelligent AI visual recognition according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0067] S101. Acquire real-time operating parameters of the DCS system, including operating status data, alarm information, operation logs, and video monitoring data of each subsystem of the DCS system. The operating status data includes real-time process parameters and equipment operating parameters of each subsystem. Preprocess the real-time operating parameters to obtain real-time operating parameter features in a unified format.

[0068] In addition to real-time operating parameters, DCS systems also include various types of heterogeneous data, including alarm information, operation logs, and video surveillance. To perform global abnormal pattern recognition, these heterogeneous data must first be preprocessed and feature extracted to form a unified feature representation.

[0069] Specifically, for alarm information, text mining techniques such as TF-IDF and Word2Vec can be used to extract keywords and semantic features. Specifically, the alarm information is segmented into a corpus, and the frequency and inverse document frequency of each word are counted to generate a TF-IDF weight matrix. This matrix is ​​then reduced in dimension using methods such as SVD to obtain a low-dimensional semantic representation of the alarm information. Pre-trained word embedding models such as BERT and GPT can also be used to map the alarm information into a high-dimensional semantic space, capturing the contextual associations between words.

[0070] For operation logs, sequential pattern mining techniques, such as N-grams and sequence association rules, can be used to extract operational behavior features. Specifically, operation logs are sorted by timestamp and a fixed-length time window is used to generate operation sequence fragments. The frequency and confidence of various operation sequences are calculated to discover frequent operation patterns and association rules. These patterns are then converted into real-valued vectors using one-hot encoding or embedding methods.

[0071] For video surveillance, computer vision technologies such as object detection, object tracking, and behavior recognition can be used to extract device operational status features. Specifically, pre-trained object detection models such as YOLO and Faster R-CNN can be used to detect the device's location and category within video frames. Multi-object tracking algorithms such as DeepSORT and TrackletNet can then be used to track the device's trajectory. Based on the trajectory, motion features such as velocity, acceleration, and rotation angle, as well as behavioral features such as trajectory shape and movement frequency, can be extracted. Furthermore, optical flow and skeleton information can be used to analyze the device's operational status and posture changes.

[0072] Finally, the preprocessed different modal data are converted into the form of feature vectors, and they are spliced ​​or fused to obtain a unified format of real-time operation parameter feature representation.

[0073] In this embodiment, by acquiring and preprocessing multi-source heterogeneous data from the DCS system, a solid data foundation is laid for efficient and comprehensive system status monitoring. This enables comprehensive monitoring of the DCS system's real-time operating status, encompassing multiple dimensions such as process parameters, equipment status, alarm information, operation records, and video surveillance, providing a panoramic view of system performance. This overcomes the limitation of traditional methods that can only process a single data source, significantly improving the coverage and accuracy of condition monitoring. This provides high-quality data support for subsequent advanced applications such as fault diagnosis, predictive maintenance, and optimized decision-making, thereby enhancing the operational reliability and efficiency of the DCS system.

[0074] S102. Input the real-time operating parameter features into a pre-built intelligent AI visual recognition model, analyze the operating status of the DCS system, identify abnormal modes at the system level, compare the abnormal modes with normal operating modes in a preset normal operating mode library, and calculate the similarity between the abnormal modes and the normal operating modes.

[0075] Optionally, the intelligent AI visual recognition model can adopt a deep learning algorithm, through the combined structure of convolutional neural network (CNN) and long short-term memory network (LSTM), to automatically extract the multi-dimensional and multi-scale key features contained in the feature vector of real-time operation parameters, explore the spatiotemporal correlation between parameters, and realize the accurate identification of abnormal patterns.

[0076] The model construction steps include collecting a large amount of historical DCS system operating data, including normal operating conditions and various abnormal conditions, and preprocessing it to obtain a sample set of parameter feature vectors. Based on the characteristics of the DCS system and the complexity of the abnormal patterns, a combined CNN + LSTM architecture was designed. The CNN extracts local features of the parameter feature vectors, while the LSTM mines the temporal correlations between features. A supervised learning approach is used to train the model using sample data. The model parameters are updated through a backpropagation algorithm to continuously improve the model's feature extraction and anomaly detection capabilities. Optimization techniques such as attention mechanisms and regularization are introduced to enhance the model's generalization performance. Methods such as holdout and cross-validation are used to evaluate the model's anomaly detection accuracy, precision, and recall, verifying the model's effectiveness and reliability.

[0077] During operation, the real-time operating parameter feature vectors are input into a trained intelligent AI visual recognition model. The model automatically and comprehensively analyzes the operating status of each subsystem and identifies potential abnormal patterns. To determine the severity of the abnormal pattern, the identified abnormal pattern is compared with the patterns in the preset normal operating pattern library and the similarity between the abnormal and normal operating patterns is calculated. The lower the similarity, the further the current operating status deviates from normal conditions, and the greater the abnormal risk. When the similarity falls below the preset abnormal threshold, the DCS system is judged to have an abnormal risk, and further early warning measures are required.

[0078] In an optional embodiment,

[0079] Inputting the real-time operating parameter features into a pre-built intelligent AI visual recognition model to analyze the operating status of the DCS system and identify abnormal patterns at the system level include:

[0080] A multimodal feature fusion network is designed in the intelligent AI visual recognition model. Text mining is performed on alarm information in real-time operation parameter features to obtain alarm information features. Sequential pattern mining is performed on operation logs to obtain operation log features. Target detection, target tracking, and behavior recognition are performed on video surveillance data to obtain video surveillance features.

[0081] The extracted alarm information features, operation log features, and video surveillance features are input into a multimodal feature fusion network, and each feature is mapped to a latent space of the same dimension through a feature mapping method.

[0082] The attention mechanism is used to calculate the similarity between features in the latent space, assign weights to each feature, and use the hybrid expert network to fuse the features in the latent space to obtain fused features. The fused features are then compressed using the autoencoder to obtain the final multimodal fused features.

[0083] A heterogeneous graph is constructed based on the multimodal fusion features. The nodes in the heterogeneous graph include devices, parameters, and alarms, and the edges include physical connections, logical associations, and temporal dependencies. A graph neural network is applied to the heterogeneous graph to iteratively update the hidden states of the nodes through a message passing mechanism. In each propagation step, the node first aggregates messages from neighboring nodes and then updates the node hidden state based on its own features and neighbor information.

[0084] For the local anomaly detection task, a graph decoder is used to map the node hidden state into anomaly metrics, and the graph neural network corresponding to the local anomaly detection task is trained by optimizing the cross-entropy loss function between the anomaly metrics and the true anomaly labels of the nodes. For the global anomaly association task, a graph pooling method based on anomaly metrics is used to gradually aggregate node information to generate an abnormal region hierarchy, and a graph decoder is used to map the abnormal region features into a global anomaly metric. The graph neural network corresponding to the global anomaly association task is trained by optimizing the cross-entropy loss function between the global anomaly metric and the true anomaly labels at the system level.

[0085] The two subtasks of local anomaly detection and global anomaly association share the same graph encoder, and achieve joint optimization through alternating training and parameter sharing. The outputs of the two subtasks are combined to finally output the system-level anomaly pattern recognition results.

[0086] As an example, a multimodal feature fusion network is first designed. This network consists of three modules: feature mapping, feature interaction, and feature fusion. The feature mapping module uses a deep neural network to map raw features from different modalities into a latent space of the same dimensionality, achieving feature alignment and unified representation. The feature interaction module utilizes attention and gating mechanisms to model the interactions and dependencies between features from different modalities. The attention mechanism assigns weights to each feature by calculating their similarity, highlighting key features and suppressing noisy features. The gating mechanism adaptively controls the flow and fusion of information from different modalities by learning gating functions, dynamically adjusting the contribution of each modality. The feature fusion module utilizes a hybrid expert network and autoencoder to achieve deep fusion and high-level abstraction of different interaction features. The hybrid expert network comprises multiple subnetworks, each corresponding to a different interaction mode. The weights of each subnetwork are adaptively adjusted based on task requirements and sample distribution, integrating the decision results of different interaction modes. The autoencoder compresses the fused features into a lower-dimensional latent space by reconstructing the task, optimizing the compactness and discriminability of the feature representation.

[0087] After obtaining multimodal fusion features, a graph neural network is further used to model the topological structure of the system to which the heterogeneous data belongs, mining the complex associations and interaction patterns between entities such as devices, parameters, and alarms. Specifically, a heterogeneous graph is constructed with devices, parameters, and alarms as nodes and physical connections, logical associations, and temporal dependencies as edges. The graph neural network consists of two parts: a graph encoder and a graph decoder. The graph encoder uses a message passing mechanism to iteratively update the node's hidden state. In each propagation step, a node first aggregates messages from neighboring nodes and then updates its hidden state by combining its own features with neighbor information. Through multiple rounds of message passing, nodes can capture multi-hop relationships and long-range dependencies, learning more global, higher-order feature representations. The graph decoder uses a readout function to generate an anomaly metric based on the node's hidden state, indicating the likelihood that the node belongs to an anomaly pattern. By optimizing the loss function between the anomaly metric and the true label, the graph neural network can be trained end-to-end, adaptively mining the node interactions and propagation paths most relevant to anomaly patterns. Graph regularization is also introduced during training. Graph regularization models the similarities and differences between nodes, guiding the graph neural network to learn more generalized and robust representations of anomaly patterns. On the one hand, similar nodes are encouraged to learn similar representations to avoid overfitting to noise patterns. On the other hand, different nodes are penalized for learning the same representation, which improves the discriminability and diversity of features.

[0088] Building on graph neural networks (GNNs), we further designed a multi-task learning framework for anomaly pattern recognition. This framework comprises two subtasks: local anomaly detection and global anomaly association. Through joint modeling and training, it fully leverages the multi-scale and multi-granular nature of anomaly patterns, improving the accuracy and generalization of anomaly pattern recognition. First, for the local anomaly detection task, the GNN node hidden states are used to generate anomaly metrics. Specifically, by aggregating neighbor information and updating its own state through a graph encoder, the GNN learns a high-level feature representation for each node, characterizing its role and importance in the graph. Then, the graph decoder maps the node hidden state into an anomaly metric, representing the likelihood that the node belongs to an anomaly pattern. The anomaly metric can be generated using a simple fully connected layer or a more complex attention mechanism or gating mechanism to adaptively adjust the contribution of different nodes. The anomaly metric aims to be as consistent as possible with the node's true anomaly label. Therefore, the GNN for the local anomaly detection task is trained by minimizing the cross-entropy loss between the anomaly metric and the true label.

[0089] Secondly, for the global anomaly association task, an anomaly propagation mechanism based on graph pooling was designed. Graph pooling generates subgraph-level representations by aggregating node features, gradually compressing the graph into higher-level abstractions. At each pooling layer, nodes are sorted according to the anomaly metric, and several nodes with the highest anomaly metrics are selected as anomaly centers. Their neighboring nodes are aggregated to the anomaly centers to form an anomaly subgraph. The features of the anomaly centers are aggregated using an aggregation function to summarize the information within the anomaly subgraph and pass it to the next layer. Through multiple iterations, anomaly information can be propagated and diffused over the graph, forming hierarchical anomaly regions. The feature representation of the anomaly region is used to generate a global anomaly metric through a graph decoder, which characterizes the anomaly risk of the entire system. The learning goal of the global anomaly metric is to be consistent with the true anomaly labels at the system level. Therefore, the graph neural network corresponding to the global anomaly association task is trained by minimizing the cross-entropy loss function between the global anomaly metric and the true labels.

[0090] Finally, to effectively integrate local anomaly detection and global anomaly association, a multi-task learning paradigm is adopted to jointly optimize the two subtasks. Multi-task learning enables information exchange and knowledge transfer between different tasks by sharing some network parameters, improving the generalization and robustness of the model. Specifically, the two subtasks share the same graph encoder and independently use task-specific graph decoders. The graph encoder is responsible for learning universal node feature representations to capture the commonalities of anomaly patterns. The graph decoder is responsible for adapting to the needs of different tasks and generating task-specific anomaly metrics. During joint training, the loss functions of the two subtasks are optimized sequentially, and the shared graph encoder parameters are updated during backpropagation. Through alternating training and parameter sharing, multi-task learning can explicitly balance local anomaly detection and global anomaly association, exploring the performance of anomaly patterns at different scales and granularities, thereby obtaining more accurate and comprehensive anomaly pattern recognition results.

[0091] During the inference phase, for a given test sample, the trained graph encoder is first used to generate node hidden states. Then, the graph decoders for local anomaly detection and global anomaly association are used to generate node-level and system-level anomaly metrics, respectively. Finally, the predictions from the two anomaly metrics are combined, and the final anomaly pattern recognition result is generated through a weighted average or voting mechanism. The weights of the two anomaly metrics can be flexibly adjusted based on task requirements and domain knowledge to balance the impact of local and global anomalies.

[0092] In an optional embodiment,

[0093] The two subtasks, local anomaly detection and global anomaly association, share the same graph encoder. Joint optimization is achieved through alternating training and parameter sharing. The outputs of the two subtasks are combined to produce the following system-level anomaly pattern recognition results:

[0094] A joint optimization loss function is constructed based on the loss function of the local anomaly detection subtask and the loss function of the global anomaly association subtask. The training steps of the local anomaly detection and global anomaly association subtasks are repeated until the joint optimization loss function converges and the system-level anomaly pattern recognition result is obtained. The formula of the joint optimization loss function is as follows:

[0095]

[0096] Among them, L represents the joint optimization loss function, λ1 represents the weight of the local anomaly detection task, V represents the number of nodes, α i represents the weight factor of node i, represents the predicted abnormality metric of node i, y i represents the true anomaly label of node i, γ represents the exponential parameter of anomaly measurement error, λ2 represents the weight of the global anomaly association task, C represents the number of operation log features, α c represents the weight factor of neighbor node c, y c represents the true abnormal label of the neighbor node c, represents the predicted abnormality metric of neighbor node c, λ3 represents the weight of the regularization term, C i represents the set of neighbor nodes of node i, and p represents the exponential parameter of the regularization term.

[0097] In this embodiment, a multimodal feature fusion network is designed to map different modal data, such as text alarm information, time-series operation logs, and video surveillance data, into the same semantic space for fusion. Using an attention mechanism and a hybrid expert network, weights of different modal features are adaptively assigned, fully exploiting the correlations between them. The fused multimodal features are constructed into a heterogeneous graph, which can depict the complex topological structures and relationships between devices, parameters, and alarms. The message propagation mechanism of a graph neural network efficiently aggregates local and global information on the heterogeneous graph, enabling learning and reasoning about complex anomaly patterns. For local anomaly detection tasks, the graph neural network accurately detects anomalies for each parameter or device based on the corresponding anomaly metrics output by the node hidden state. Furthermore, through graph pooling operations based on anomaly metrics, the graph neural network can also learn global anomaly patterns, enabling system-level anomaly correlation analysis. The local and global anomaly subtasks share the same graph encoder. Through alternating training and parameter sharing, joint optimization of the two tasks is achieved. This end-to-end optimization approach avoids the error accumulation of traditional step-by-step processing and improves the overall performance of the model.

[0098] S103. If the similarity is lower than the preset abnormality threshold, it is determined that there is an abnormal risk in the current operation of the DCS system. The real-time operation scenario of the DCS system is simulated using digital twin simulation technology based on physical mechanisms to predict the failure mode, cause and probability of the DCS system. Based on the prediction results and the preset multi-level early warning decision matrix, early warning information of the corresponding level is automatically generated and pushed to the intelligent operation and maintenance platform of the DCS system in real time to assist operators in identifying system risks in advance and provide targeted preventive maintenance measures.

[0099] Based on the results of the abnormality risk prediction, the system automatically matches a pre-set multi-level early warning decision matrix to generate warning information of the corresponding level. Optionally, the early warning decision matrix can be pre-set based on historical experience and expert knowledge, linking the abnormality risk level with the warning level, warning method, and recommended response measures. The generated warning information is pushed to DCS system operators in real time via the intelligent operation and maintenance platform, including the warning level, the cause of the abnormality, possible consequences, and recommended response measures. This helps operators take targeted preventive maintenance measures immediately to curb the further deterioration of the abnormality and avoid accidents.

[0100] The intelligent operation and maintenance platform utilizes a B / S architecture, supports web and mobile access, and features a user-friendly visual interface, allowing operators to monitor system status and receive anomaly warnings anytime, anywhere. Key functional modules of the intelligent operation and maintenance platform include: Real-time Data Display: Displays operating parameters and KPIs for each DCS subsystem in real time using graphs, bar charts, and dashboards, providing a visual overview of the overall system's operational status. Abnormal Alert Push: Displays anomaly warning information prominently, including the warning level, cause of the anomaly, possible consequences, and recommended countermeasures. This information is delivered promptly via various channels, including audible and visual alarms, SMS, and email. Abnormal Analysis and Diagnosis: Visual analysis tools for abnormality patterns, such as fault tree diagrams and cause-and-effect diagrams, are provided to help operators quickly locate the root cause of anomalies and develop targeted response plans. Emergency Plan Knowledge Base: Establishes an abnormality warning and emergency plan knowledge base, summarizing standard handling procedures and emergency response plans for various abnormal situations, providing operators with a reference for decision-making. Mobile App: Develops an intelligent operation and maintenance mobile app to enable real-time push notifications of abnormality warnings, allowing operators to monitor system dynamics and respond promptly.

[0101] In an optional embodiment,

[0102] Using digital twin simulation technology based on physical mechanisms, we simulate the real-time operation scenarios of the DCS system and predict the failure modes, causes, and occurrence probabilities of the DCS system, including:

[0103] Build a multidisciplinary, high-fidelity digital twin model of the DCS system, which includes a production process model, equipment model, and control logic model;

[0104] Extract key feature parameters of abnormal patterns, use feature dimensionality reduction methods to reduce the dimensions of key feature parameters, and use semantic mapping technology to map the reduced key feature parameters to the corresponding positions and object attributes of the digital twin model, triggering the digital twin model to perform simulation predictions under abnormal working conditions;

[0105] The simulation prediction starts from the anomaly point, adopts a multi-scenario parallel simulation strategy, sets different initial conditions, boundary conditions and random disturbances, generates multiple parallel simulation instances, and adopts model simplification and multi-scale modeling methods to selectively simplify the digital twin model at different spatiotemporal scales, performs grid encryption and computational intensification for key areas and key processes, and introduces uncertainty quantification technology to adaptively adjust the spatiotemporal resolution and computational accuracy of the simulation;

[0106] Based on the simulation of multi-scenario abnormal working conditions, with the abnormal pattern as the top event, deductive reasoning and causal reverse tracing methods are used to automatically construct a fault tree, identify the key links and influencing factors of abnormal propagation, and use the Monte Carlo simulation method to randomly sample the probability of occurrence of basic events in the fault tree and evaluate the abnormal pattern.

[0107] According to the results of fault tree analysis and Monte Carlo simulation, a case-based reasoning fault diagnosis method is adopted to measure the similarity between abnormal patterns and historical fault cases, match fault patterns from the historical fault library, and finally output the predicted fault mode, fault cause and occurrence probability of the DCS system.

[0108] As an example, a multidisciplinary, high-fidelity digital twin model of a discrete control system was first constructed, comprising a production process model, an equipment model, and a control logic model. The production process model employs an object-oriented modeling approach, abstracting the production process into an object network consisting of equipment, pipelines, and valves, and defining the connections and interaction interfaces between these components. The equipment model employs numerical simulation methods, combining the equipment's geometric parameters, material properties, and operating conditions to simulate the physical and chemical processes within the equipment, such as flow, heat transfer, mass transfer, and chemical reactions. This model also considers the equipment's mechanical characteristics and failure modes, improving the integrity and accuracy of the equipment model. The control logic model converts control strategies and logical relationships into an executable code model, dynamically coupling it with the production process model and equipment model to achieve closed-loop control of the system. The operating parameters of the discrete control system are collected in real time through IoT devices and sensor networks and mapped into the digital twin model, enabling synchronized updates between the physical and virtual worlds.

[0109] Based on the construction of a digital twin model, an intelligent visual recognition model is used to detect abnormal operating conditions in discrete control systems. Key characteristic parameters of abnormal patterns are extracted, and feature dimensionality reduction methods such as principal component analysis and independent component analysis are used to reduce the dimensionality of these key characteristic parameters. Principal component analysis transforms original features into linearly independent new features, called principal components, through an orthogonal transformation. These features are sorted by variance, and the first few principal components with the largest variance are selected as the reduced features. Independent component analysis assumes that the observed data is a linear combination of several statistically independent signals and optimizes the objective function to find these independent signals, thereby achieving feature dimensionality reduction. Following feature dimensionality reduction, this paper uses semantic mapping technology to map the reduced key characteristic parameters to the corresponding locations and object attributes of the digital twin model, recreating the abnormal conditions in a realistic virtual environment. Semantic mapping utilizes knowledge representation methods such as ontologies and semantic networks to establish semantic associations between abnormal features and the digital twin model, enabling seamless transfer of abnormal information between physical and virtual space.

[0110] After reproducing the abnormal operating conditions, the digital twin model is triggered to simulate and predict the abnormal operating conditions. A multi-scenario parallel simulation strategy is employed, setting different initial conditions, boundary conditions, and random perturbations to generate multiple parallel simulation instances. Multi-scenario parallel simulation comprehensively explores the evolution path of abnormal operating conditions, improving the coverage and reliability of simulation results. During the parallel simulation process, methods such as model simplification and multi-scale modeling are employed to selectively simplify the digital twin model at different spatiotemporal scales. Model simplification involves removing minor factors that have little impact on the simulation while ensuring key characteristics, thereby reducing model complexity. Multi-scale modeling involves selecting an appropriate time step and spatial grid based on the characteristics of the simulation object to describe system behavior at different scales. Grid refinement and computational intensification are implemented in key areas and processes involved in anomaly propagation and diffusion, improving local simulation accuracy. Furthermore, to quantify uncertainty in the simulation process, uncertainty quantification techniques are introduced to assess the uncertainty introduced by model simplification and numerical discretization, and to adaptively adjust the spatiotemporal resolution and computational accuracy of the simulation, achieving a trade-off between simulation efficiency and reliability.

[0111] Based on multi-scenario abnormal operating condition simulations, with the abnormal pattern as the top event, deductive reasoning and causal backtracking are used to automatically construct a fault tree of possible causes of the abnormality. Deductive reasoning starts from general knowledge and derives conclusions for individual cases based on logical rules, while causal backtracking starts from the abnormal result and searches backwards for possible influencing factors and fault chains. By combining these two methods, the key links and influencing factors of the abnormality propagation are extracted from the digital twin model, forming a complete fault tree. After constructing the fault tree, Monte Carlo simulation is used to randomly sample the probability of occurrence of basic events in the fault tree. Monte Carlo simulation simulates a probability distribution using random numbers and estimates the characteristics or numerical solutions of the probability distribution function through a large number of random experiments. Monte Carlo simulation can assess the probability of the abnormal pattern leading to system failure. For low-probability events in complex systems, improved Monte Carlo methods such as importance sampling and subset simulation are used to improve the efficiency and accuracy of probability estimation.

[0112] Finally, based on the results of fault tree analysis and Monte Carlo simulation, this paper adopts a case-based reasoning fault diagnosis method. By measuring the similarity between abnormal patterns and historical fault cases, this method matches specific fault patterns that may be caused by abnormalities from the historical fault library. Case-based reasoning is a common analogical reasoning method that extracts case features, calculates the similarity between cases, finds the case most similar to the target problem, and then uses the solution of the case to solve the target problem. When matching fault patterns, it not only considers the characteristic features of the abnormal pattern but also the contextual information of the fault occurrence, such as equipment type, operating conditions, and environmental conditions. Ultimately, it outputs the predicted DCS system fault mode, fault cause, and occurrence probability, improving the accuracy of fault matching.

[0113] In this example, a digital twin-based abnormal risk prediction and early warning method was developed, integrating multidisciplinary simulation, intelligent feature extraction, fault tree analysis, Monte Carlo simulation, and case-based reasoning to establish a comprehensive abnormality diagnosis and risk assessment process. By building a high-fidelity digital twin model, real-time monitoring, prediction, and early warning of abnormal operating conditions were achieved, providing strong support for the safe operation of the DCS system.

[0114] In an optional embodiment,

[0115] Using a multi-scenario parallel simulation strategy, setting different initial conditions, boundary conditions, and random perturbations, multiple parallel simulation instances are generated, including:

[0116] Obtain historical anomaly data from the DCS system for preprocessing, extract multi-scale spatiotemporal features of the historical anomaly data, including anomaly features at different time and spatial scales, and construct a multi-scale spatiotemporal feature representation of the anomaly scene;

[0117] Inputting the multi-scale spatiotemporal feature representation of the abnormal scene into a multi-scale conditional generative adversarial network, wherein the multi-scale conditional generative adversarial network includes a generator and a discriminator, wherein the generator receives random noise and conditional variables as input, and the discriminator receives the real abnormal scene features and the abnormal scene features generated by the generator as input;

[0118] In the generator, multi-scale feature extraction is performed on random noise and conditional variables to obtain noise features and conditional features of different scales. The attention mechanism is used to adaptively fuse the noise features and conditional features of different scales to generate abnormal scene features that meet the conditional constraints. In the discriminator, multi-scale feature extraction is performed on the input real abnormal scene features and the generated abnormal scene features to obtain real features and generated features of different scales. The attention mechanism is used to adaptively fuse the real features and generated features of different scales.

[0119] Constructing the objective function of a multi-scale conditional generative adversarial network, which includes the logarithmic probability of the discriminator's true abnormal scene features and the false probability of the abnormal scene features generated by the generator. Optimization training is performed by maximizing the discriminator's objective function and minimizing the generator's objective function, ultimately obtaining a trained abnormal scene distribution model.

[0120] Setting an abnormality type and severity condition for an abnormal scenario, encoding the abnormality type and severity condition into a conditional vector, concatenating the conditional vector with a random noise vector to form a conditional noise vector, inputting the conditional noise vector into a generator of an abnormal scenario distribution model, and generating an abnormal scenario that meets the abnormality type and severity conditions through the generator;

[0121] Taking the diversity, coverage and abnormality degree of the abnormal scenarios as the optimization goals, a fitness function is constructed, an adaptive sampling strategy is adopted to optimize the sampling of abnormal scenarios, and the sampling parameters are adaptively adjusted to obtain an optimized set of abnormal scenarios. The optimized set of abnormal scenarios is then input into the digital twin model for multi-scenario parallel simulation.

[0122] For example, the historical anomaly data of the DCS system is first preprocessed through cleaning, denoising, and normalization to improve data quality and consistency. Multi-scale spatiotemporal features are then extracted from the preprocessed historical anomaly data. Temporal scale features reflect the evolution of anomaly scenarios in the temporal dimension, including anomaly duration, anomaly frequency, and anomaly change trends. Spatial scale features reflect the distribution of anomaly scenarios in the spatial dimension, including anomaly impact range, anomaly propagation path, and anomaly correlation relationships. By extracting anomaly features at different temporal and spatial scales, a multi-scale spatiotemporal feature representation of anomaly scenarios is constructed, providing an information basis for subsequent anomaly scenario generation and optimization.

[0123] The generation of abnormal scenarios requires careful consideration of constraints such as anomaly type and severity, while also ensuring their authenticity and diversity. To this end, we introduce a Multi-Scale Conditional Generative Adversarial Network (MS-CGAN) to model the distribution of abnormal scenarios. The MS-CGAN consists of two parts: a generator and a discriminator. The generator is responsible for generating realistic abnormal scenarios based on the conditions, while the discriminator is responsible for determining the differences between the generated abnormal scenarios and real-world abnormal scenarios.

[0124] In the generator of the MS-CGAN, random noise and conditional variables are input and subjected to multi-scale feature extraction, respectively, to generate noise features and conditional features at different scales. The noise features introduce randomness, increasing the diversity of the generated scenes; the conditional features introduce constraint information, controlling the properties of the generated scenes. The attention mechanism then adaptively fuses the noise and conditional features at different scales. The attention mechanism dynamically adjusts the weights of the noise features based on the conditional information, ensuring that the generated anomaly scenes better conform to the conditional constraints. The fused features are then subjected to upsampling operations such as deconvolution to gradually restore them to the original anomaly scene scale, ultimately generating anomaly scene features that conform to the conditional constraints.

[0125] In the MS-CGAN discriminator, real abnormal scene features and abnormal scene features generated by the generator are used as input. They undergo multi-scale feature extraction to obtain real features and generated features at different scales. The attention mechanism then adaptively fuses these real and generated features at different scales. The attention mechanism dynamically adjusts the weights of features at different scales based on the differences between real and generated features, improving the discriminator's discriminative ability. The fused features are passed through a fully connected layer, outputting the probability values ​​of real and generated abnormal scenes.

[0126] The training process of MS-CGAN is a minimax optimization problem. The discriminator's goal is to maximize the difference between the log probability of true anomaly features and the false probability of generated anomaly features—that is, its ability to distinguish true anomalies from generated ones. The generator's goal is to minimize the false probability of generated anomaly features, thereby deceiving the discriminator into misclassifying generated anomalies as true. By alternately training the generator and discriminator and continuously updating their parameters, a generator capable of generating realistic anomalies, or an anomaly distribution model, is ultimately obtained.

[0127] After the abnormal scenario distribution model is established, the conditions for the abnormal scenarios, such as abnormality type and severity, can be set as needed. The abnormality type and severity conditions are encoded into a fixed-length conditional vector and concatenated with a randomly sampled noise vector to form a conditional noise vector. The conditional noise vector is input into the generator of the abnormal scenario distribution model. After multi-scale feature extraction, attention fusion, and upsampling, abnormal scenario features that meet the conditional constraints are generated. The abnormality type condition controls the abnormality category of the generated scenario, such as equipment failure, parameter out-of-limit, and environmental anomaly; the severity condition controls the abnormality degree of the generated scenario, such as minor abnormality, moderate abnormality, and severe abnormality. By setting different condition combinations, a variety of abnormal scenarios can be generated, covering different abnormality types and severities.

[0128] Although the abnormal scenarios generated by the abnormal scenario distribution model meet the conditional constraints, they may contain duplication, redundancy, or insufficient abnormality. Therefore, the generated abnormal scenarios require optimized sampling. A fitness function is constructed with the diversity, coverage, and abnormality level of abnormal scenarios as optimization objectives. The diversity objective measures the differences between abnormal scenarios to avoid generating duplicate or similar scenarios; the coverage objective measures the coverage of abnormal scenarios for different abnormal types and severities to ensure that the generated scenarios are representative; and the abnormality level objective measures the abnormality level of abnormal scenarios, prioritizing scenarios with higher abnormality levels.

[0129] An adaptive sampling strategy is employed during the sampling of abnormal scenarios. Initially, sampling is performed uniformly across the abnormal scenario space to ensure randomness and comprehensiveness. As sampling progresses, the quality of the sampled scenarios is evaluated based on a fitness function, and the sampling distribution is adaptively adjusted. The sampling probability is increased for abnormal scenario regions with high fitness, while it is decreased for regions with low fitness. Furthermore, sampling parameters such as the sampling step size and acceptance probability are dynamically adjusted to balance exploration and exploitation. This adaptive sampling strategy continuously optimizes the abnormal scenario collection, improving both quality and efficiency.

[0130] To assess the impact and consequences of abnormal scenarios, the optimized set of abnormal scenarios needs to be input into a digital twin model for simulation. A digital twin model is a virtual representation of a physical system, capable of simulating its behavior and response. By injecting abnormal scenarios into the digital twin model, the system's operating status and fault evolution under abnormal conditions can be observed. Given the large size of the abnormal scenario set and the low efficiency of serial simulation, a multi-scenario parallel simulation approach is adopted. The abnormal scenario set is divided into multiple subsets, each assigned to a different compute node, and simulated simultaneously in multiple replicas of the digital twin model. Parallel simulation fully utilizes computing resources and accelerates the evaluation and analysis of abnormal scenarios. After the simulation results are summarized, a comprehensive assessment of the impact, severity, and evolution trends of different abnormal scenarios can be made, providing support for abnormal warning and decision-making.

[0131] In an optional embodiment,

[0132] The calculation formula for the objective function of constructing a multi-scale conditional generative adversarial network is as follows:

[0133]

[0134] in, represents the objective function of the multi-scale conditional generative adversarial network, R(D, G) represents the value function between the generator G and the discriminator D, and E f [·] represents the expectation of the characteristics of the real abnormal scene, E z [·] represents the expectation of the generated synthetic abnormal scene features, J represents the time scale level, K represents the spatial scale level, and a (D) represents the attention weight in the discriminator, f (j，k) Multi-scale representation of real abnormal scene features, s (j，k) represents the multi-scale representation of the conditional variable, a (G) represents the attention weight in the generator, z (j，k) Representing multi-scale representations of random noise.

[0135] In this embodiment, historical anomaly data from the DCS system is fully utilized to extract multi-scale spatiotemporal features and construct an information foundation for anomaly scenarios. A multi-scale conditional generative adversarial network is introduced to establish an anomaly scenario distribution model, generating realistic, diverse, and conditionally constrained anomaly scenarios. An adaptive sampling strategy is employed to optimize the generated anomaly scenarios, improving their quality and efficiency. Parallel simulation of multiple scenarios accelerates the evaluation and analysis of anomaly scenarios, providing support for anomaly warning and decision-making. This integrated approach enables intelligent generation and optimization of anomaly scenarios, improving anomaly warning and location capabilities, and reducing the cost of manually designing anomaly scenarios, demonstrating strong practicality and operability.

[0136] In an optional embodiment,

[0137] According to the prediction results and the preset multi-level warning decision matrix, the corresponding level of warning information is automatically generated, including:

[0138] Construct an abnormal risk warning domain ontology, define the relationship mapping between abnormality types, influencing factors, warning levels and response measures, and form an abnormal risk warning ontology model;

[0139] Integrate structured data, semi-structured data, and unstructured data, realize the association of multi-source heterogeneous data through data integration, and use the knowledge extraction model of deep learning to extract entities, attributes, and relationships from unstructured data to form an abnormal risk warning knowledge graph;

[0140] Taking the anomaly type and influencing factors output by the anomaly risk prediction as known conditions, matching the corresponding entity nodes in the consistent anomaly risk warning knowledge graph, and applying the anomaly risk warning ontology reasoning rules and graph reasoning algorithms to infer the concepts, attributes, and relationships associated with the known conditions, generating an anomaly risk warning reasoning explanation;

[0141] Using the knowledge subgraph obtained from the abnormal risk warning reasoning interpretation, an abnormal risk warning question and answer template is constructed to form a structured question description. The question description is input into the Seq2Seq-based natural language generation model to automatically generate a text of recommended abnormal risk response measures;

[0142] Based on the confidence, coverage and novelty indicators of abnormal risk response measures knowledge, a knowledge sorting algorithm is used to prioritize the suggested texts of abnormal risk response measures. The abnormal risk level is calculated according to the abnormal prediction results, and the abnormal risk response measures of the corresponding level are selected to automatically generate early warning information of the corresponding abnormal risk level.

[0143] For example, first, through in-depth analysis of the abnormal risk warning field, core concepts such as abnormality type, influencing factors, warning level, and response measures are identified. Then, hierarchical and non-hierarchical relationships between concepts are defined, such as the causal relationship between abnormality type and influencing factors, and the correspondence between warning level and response measures. Next, attributes are added to the concepts, such as the severity attribute of the abnormality type and the weight attribute of the influencing factors. Finally, based on domain knowledge and expert experience, a series of ontological reasoning rules and constraints are constructed, such as that a certain combination of abnormality type and influencing factor will result in a specific warning level, and that a specific warning level requires corresponding response measures. Through these steps, an abnormal risk warning ontology model is formed, providing a semantic foundation for subsequent knowledge graph construction and reasoning.

[0144] Abnormal risk warning knowledge comes from various types of data, including structured, semi-structured, and unstructured data. Structured data includes exception record tables in relational databases, semi-structured data includes exception reports in XML format, and unstructured data includes fault repair logs and expert experience documents. To build a comprehensive and accurate abnormal risk warning knowledge graph, it is necessary to fuse and extract knowledge from multi-source heterogeneous data. First, data integration technologies such as ETL and data virtualization are used to clean, transform, and link data from different sources and formats to form a unified data view. Then, deep learning knowledge extraction models such as BERT and CNN-LSTM are used to identify and extract entities, attributes, and relationships related to abnormal risk warning from unstructured data. Deep learning models can automatically learn the semantic features of text, overcoming the limitations of traditional rule-based and template-based approaches. The extracted knowledge elements are then fused with structured data and mapped to corresponding concepts, attributes, and relationships based on the abnormal risk warning ontology model, ultimately forming a semantically rich abnormal risk warning knowledge graph.

[0145] When the anomaly risk prediction module outputs anomaly types and influencing factors, it is necessary to infer the corresponding warning levels and response measures based on these known conditions. The anomaly risk warning knowledge graph provides comprehensive domain knowledge, including associations between anomaly types, influencing factors, warning levels, and response measures, to support this reasoning process. First, using anomaly types and influencing factors as query conditions, consistent entity nodes are matched in the anomaly risk warning knowledge graph. This matching process considers semantic similarity, including not only literal matching but also synonym matching, and hypernymy matching. Then, using the matched entity nodes as a starting point, the anomaly risk warning ontology reasoning rules and graph reasoning algorithms are applied to infer concepts, attributes, and relationships associated with the known conditions. Ontology reasoning rules are based on ontology axioms and domain constraints, utilizing formal methods such as description logic to deduce and reason about knowledge. Graph reasoning algorithms, such as the shortest path algorithm and random walk algorithm, analyze the topological structure and edge semantics of the knowledge graph to discover implicit associations. Finally, the inferred associated concepts, attributes, and relationships are organized into a knowledge subgraph to form an anomaly risk warning reasoning explanation, providing knowledge support for the subsequent generation of response measures.

[0146] Based on the knowledge subgraph derived from the abnormal risk warning reasoning, an abnormal risk warning question and answer template is constructed. The question and answer template defines a structured representation of the question, including elements such as abnormality type, influencing factors, warning level, and the logical relationship between these elements. The entities, attributes, and relationships in the knowledge subgraph are mapped to the question and answer template to form a structured problem description. Then, a natural language generation model based on Seq2Seq is used to convert the structured problem description into natural language text. The Seq2Seq model consists of an encoder and a decoder. The encoder encodes the problem description into a semantic vector, and the decoder generates natural language response measures based on the semantic vector. The model is trained on a corpus of abnormal risk response measures to learn the mapping relationship between problem descriptions and response measures. In this way, recommended text for abnormal risk response measures is automatically generated, providing highly readable and complete abnormal handling guidance.

[0147] The generated texts are evaluated based on the confidence, coverage, and novelty metrics of the abnormal risk response knowledge. Confidence reflects the credibility of the response knowledge, coverage reflects its completeness, and novelty reflects its innovativeness. Using knowledge ranking algorithms such as PageRank and HITS, the response suggestion texts are prioritized based on these metrics, with texts with high confidence, broad coverage, and strong novelty prioritized. After sorting, the response suggestion texts that match the abnormal risk prediction results are selected as a candidate set. Finally, based on the degree of abnormality in the abnormal risk prediction results, the abnormal risk level is calculated. The response suggestion texts corresponding to the abnormal risk level in the candidate set are selected and the corresponding warning information is automatically generated. The warning information includes the abnormal risk level, abnormality description, impact scope, and response measures, presented in a concise and clear manner to provide decision support for relevant personnel.

[0148] In this embodiment, an abnormal risk warning domain ontology is constructed to formally represent the semantic relationships between abnormality types, influencing factors, warning levels, and countermeasures, providing a unified conceptual framework for knowledge graph construction. By integrating multi-source heterogeneous data and leveraging a deep learning knowledge extraction model, abnormal risk warning knowledge is extracted from structured, semi-structured, and unstructured data to construct a comprehensive and accurate abnormal risk warning knowledge graph. Reasoning is performed based on the abnormal risk warning knowledge graph, utilizing ontological reasoning rules and graph reasoning algorithms to infer warning levels and countermeasures based on known abnormality types and influencing factors, generating reasoned explanations for abnormal risk warnings. A Seq2Seq natural language generation model is used to convert structured problem descriptions into recommended text for abnormal risk countermeasures, providing highly readable and comprehensive guidance on abnormal handling. Abnormal risk countermeasures are evaluated and optimized, ranked based on confidence, coverage, and novelty metrics, and selected to match the abnormal risk level, generating concise and clear warning information. This approach achieves an automated process from abnormal risk prediction to countermeasure generation, improving the timeliness and accuracy of abnormal risk warnings and providing intelligent support for abnormal risk management, with strong practical value.

[0149] Figure 2 FIG is a structural diagram of a DCS early warning system based on intelligent AI visual recognition according to an embodiment of the present invention. Figure 2 As shown, the system includes:

[0150] The first unit is used to obtain real-time operating parameters of the DCS system, including operating status data, alarm information, operation logs, and video monitoring data of each subsystem of the DCS system. The operating status data includes real-time process parameters and equipment operating parameters of each subsystem, and pre-process the real-time operating parameters to obtain real-time operating parameter features in a unified format;

[0151] The second unit is used to input the real-time operating parameter features into a pre-built intelligent AI visual recognition model, analyze the operating status of the DCS system, identify abnormal modes at the system level, compare the abnormal modes with normal operating modes in a preset normal operating mode library, and calculate the similarity between the abnormal modes and the normal operating modes;

[0152] The third unit is used to determine that there is an abnormal risk in the current operation of the DCS system if the similarity is lower than the preset abnormal threshold. It uses digital twin simulation technology based on physical mechanisms to simulate the real-time operation scenario of the DCS system, predict the failure mode, cause and probability of the DCS system, and automatically generate warning information of the corresponding level based on the prediction results combined with the preset multi-level warning decision matrix. The warning information is pushed to the intelligent operation and maintenance platform of the DCS system in real time to assist operators in identifying system risks in advance and provide targeted preventive maintenance measures.

[0153] According to a third aspect of the embodiments of the present invention,

[0154] An electronic device is provided, comprising:

[0155] processor;

[0156] a memory for storing processor-executable instructions;

[0157] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0158] According to a fourth aspect of the embodiments of the present invention,

[0159] 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 method described above is implemented.

[0160] 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.

[0161] 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 above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above 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 early warning method based on intelligent AI visual recognition, characterized in that: include: Acquiring real-time operating parameters of the DCS system, including operating status data, alarm information, operation logs, and video monitoring data of each subsystem of the DCS system, wherein the operating status data includes real-time process parameters and equipment operating parameters of each subsystem, and preprocessing the real-time operating parameters to obtain real-time operating parameter features in a unified format; Input the real-time operating parameter features into a pre-built intelligent AI visual recognition model, analyze the operating status of the DCS system, identify abnormal patterns at the system level, compare the abnormal patterns with normal operating patterns in a preset normal operating pattern library, and calculate the similarity between the abnormal patterns and the normal operating patterns; If the similarity is lower than the preset abnormality threshold, the DCS system is judged to have an abnormal risk in its current operation. The digital twin simulation technology based on physical mechanisms is used to simulate the real-time operation scenario of the DCS system to predict the failure mode, cause and probability of the DCS system. Based on the prediction results and the preset multi-level early warning decision matrix, the corresponding level of early warning information is automatically generated and pushed to the intelligent operation and maintenance platform of the DCS system in real time to assist operators in identifying system risks in advance and provide targeted preventive maintenance measures. Input the real-time operating parameter features into a pre-built intelligent AI visual recognition model to analyze the operating status of the DCS system and identify abnormal patterns at the system level, including: A multimodal feature fusion network is designed in the intelligent AI visual recognition model. Text mining is performed on alarm information in real-time operation parameter features to obtain alarm information features. Sequential pattern mining is performed on operation logs to obtain operation log features. Target detection, target tracking, and behavior recognition are performed on video surveillance data to obtain video surveillance features. The extracted alarm information features, operation log features, and video surveillance features are input into a multimodal feature fusion network, and each feature is mapped to a latent space of the same dimension through a feature mapping method. The attention mechanism is used to calculate the similarity between features in the latent space, assign weights to each feature, and use the hybrid expert network to fuse the features in the latent space to obtain fused features. The fused features are then compressed using the autoencoder to obtain the final multimodal fused features. A heterogeneous graph is constructed based on the multimodal fusion features. The nodes in the heterogeneous graph include devices, parameters, and alarms, and the edges include physical connections, logical associations, and temporal dependencies. A graph neural network is applied to the heterogeneous graph to iteratively update the hidden states of the nodes through a message passing mechanism. In each propagation step, the node first aggregates messages from neighboring nodes and then updates the node hidden state based on its own features and neighbor information. For the local anomaly detection task, a graph decoder is used to map the node hidden state into anomaly metrics, and the graph neural network corresponding to the local anomaly detection task is trained by optimizing the cross-entropy loss function between the anomaly metrics and the true anomaly labels of the nodes. For the global anomaly association task, a graph pooling method based on anomaly metrics is used to gradually aggregate node information to generate an abnormal region hierarchy, and a graph decoder is used to map the abnormal region features into a global anomaly metric. The graph neural network corresponding to the global anomaly association task is trained by optimizing the cross-entropy loss function between the global anomaly metric and the true anomaly labels at the system level. The two subtasks of local anomaly detection and global anomaly association share the same graph encoder, and achieve joint optimization through alternating training and parameter sharing. The outputs of the two subtasks are combined to finally output the system-level anomaly pattern recognition results.

2. The method according to claim 1, characterized in that The two subtasks, local anomaly detection and global anomaly association, share the same graph encoder. Joint optimization is achieved through alternating training and parameter sharing. The outputs of the two subtasks are combined to produce the following system-level anomaly pattern recognition results: A joint optimization loss function is constructed based on the loss function of the local anomaly detection subtask and the loss function of the global anomaly association subtask. The training steps of the local anomaly detection and global anomaly association subtasks are repeated until the joint optimization loss function converges and the system-level anomaly pattern recognition result is obtained. The formula of the joint optimization loss function is as follows: Among them, L represents the joint optimization loss function, λ1 represents the weight of the local anomaly detection task, V represents the number of nodes, α i represents the weight factor of node i, represents the predicted abnormality metric of node i, y i represents the true anomaly label of node i, γ represents the exponential parameter of anomaly measurement error, λ2 represents the weight of the global anomaly association task, C represents the number of operation log features, α c represents the weight factor of neighbor node c, y c represents the true abnormal label of the neighbor node c, represents the predicted abnormality metric of neighbor node c, λ3 represents the weight of the regularization term, C i represents the set of neighbor nodes of node i, and p represents the exponential parameter of the regularization term.

3. The method according to claim 1, characterized in that Using digital twin simulation technology based on physical mechanisms, we simulate the real-time operation scenarios of the DCS system and predict the failure modes, causes, and probability of the DCS system, including: Build a multidisciplinary, high-fidelity digital twin model of the DCS system, which includes a production process model, equipment model, and control logic model; Extract key feature parameters of abnormal patterns, use feature dimensionality reduction methods to reduce the dimensions of key feature parameters, and use semantic mapping technology to map the reduced key feature parameters to the corresponding positions and object attributes of the digital twin model, triggering the digital twin model to perform simulation predictions under abnormal working conditions; The simulation prediction starts from the anomaly point, adopts a multi-scenario parallel simulation strategy, sets different initial conditions, boundary conditions and random disturbances, generates multiple parallel simulation instances, and adopts model simplification and multi-scale modeling methods to selectively simplify the digital twin model at different spatiotemporal scales, performs grid encryption and computational intensification for key areas and key processes, and introduces uncertainty quantification technology to adaptively adjust the spatiotemporal resolution and computational accuracy of the simulation; Based on the simulation of multi-scenario abnormal working conditions, with the abnormal pattern as the top event, deductive reasoning and causal reverse tracing methods are used to automatically construct a fault tree, identify the key links and influencing factors of abnormal propagation, and use the Monte Carlo simulation method to randomly sample the probability of occurrence of basic events in the fault tree and evaluate the abnormal pattern. According to the results of fault tree analysis and Monte Carlo simulation, a case-based reasoning fault diagnosis method is adopted to measure the similarity between abnormal patterns and historical fault cases, match fault patterns from the historical fault library, and finally output the predicted fault mode, fault cause and occurrence probability of the DCS system.

4. The method according to claim 3, characterized in that Using a multi-scenario parallel simulation strategy, different initial conditions, boundary conditions, and random perturbations are set to generate multiple parallel simulation instances, including: Obtain historical anomaly data from the DCS system for preprocessing, extract multi-scale spatiotemporal features of the historical anomaly data, including anomaly features at different time and spatial scales, and construct a multi-scale spatiotemporal feature representation of the anomaly scene; Inputting the multi-scale spatiotemporal feature representation of the abnormal scene into a multi-scale conditional generative adversarial network, wherein the multi-scale conditional generative adversarial network includes a generator and a discriminator, wherein the generator receives random noise and conditional variables as input, and the discriminator receives the real abnormal scene features and the abnormal scene features generated by the generator as input; In the generator, multi-scale feature extraction is performed on random noise and conditional variables to obtain noise features and conditional features of different scales. The attention mechanism is used to adaptively fuse the noise features and conditional features of different scales to generate abnormal scene features that meet the conditional constraints. In the discriminator, multi-scale feature extraction is performed on the input real abnormal scene features and the generated abnormal scene features to obtain real features and generated features of different scales. The attention mechanism is used to adaptively fuse the real features and generated features of different scales. Constructing the objective function of a multi-scale conditional generative adversarial network, which includes the logarithmic probability of the discriminator's true abnormal scene features and the false probability of the abnormal scene features generated by the generator. Optimization training is performed by maximizing the discriminator's objective function and minimizing the generator's objective function, ultimately obtaining a trained abnormal scene distribution model. Setting an abnormality type and severity condition for an abnormal scenario, encoding the abnormality type and severity condition into a conditional vector, concatenating the conditional vector with a random noise vector to form a conditional noise vector, inputting the conditional noise vector into a generator of an abnormal scenario distribution model, and generating an abnormal scenario that meets the abnormality type and severity conditions through the generator; Taking the diversity, coverage and abnormality degree of the abnormal scenarios as the optimization goals, a fitness function is constructed, an adaptive sampling strategy is adopted to optimize the sampling of abnormal scenarios, and the sampling parameters are adaptively adjusted to obtain an optimized set of abnormal scenarios. The optimized set of abnormal scenarios is then input into the digital twin model for multi-scenario parallel simulation.

5. The method according to claim 4, characterized in that The calculation formula for the objective function of constructing a multi-scale conditional generative adversarial network is as follows: in, represents the objective function of the multi-scale conditional generative adversarial network, R(D, G) represents the value function between the generator G and the discriminator D, and E f [·] represents the expectation of the characteristics of the real abnormal scene, E z [·] represents the expectation of the generated synthetic abnormal scene features, J represents the time scale level, K represents the spatial scale level, and a (D) represents the attention weight in the discriminator, f (j，k) Multi-scale representation of real abnormal scene features, s (j，k) represents the multi-scale representation of the conditional variable, a (G) represents the attention weight in the generator, z (j，k) Representing multi-scale representations of random noise.

6. The method according to claim 1, characterized in that According to the prediction results and the preset multi-level warning decision matrix, the corresponding level of warning information is automatically generated, including: Construct an abnormal risk warning domain ontology, define the relationship mapping between abnormality types, influencing factors, warning levels and response measures, and form an abnormal risk warning ontology model; Integrate structured data, semi-structured data, and unstructured data, realize the association of multi-source heterogeneous data through data integration, and use the knowledge extraction model of deep learning to extract entities, attributes, and relationships from unstructured data to form an abnormal risk warning knowledge graph; Taking the anomaly type and influencing factors output by the anomaly risk prediction as known conditions, matching the corresponding entity nodes in the consistent anomaly risk warning knowledge graph, and applying the anomaly risk warning ontology reasoning rules and graph reasoning algorithms to infer the concepts, attributes, and relationships associated with the known conditions, generating an anomaly risk warning reasoning explanation; Using the knowledge subgraph obtained from the abnormal risk warning reasoning interpretation, an abnormal risk warning question and answer template is constructed to form a structured question description. The question description is input into the Seq2Seq-based natural language generation model to automatically generate a text of recommended abnormal risk response measures; Based on the confidence, coverage and novelty indicators of abnormal risk response measures knowledge, a knowledge sorting algorithm is used to prioritize the suggested texts of abnormal risk response measures. The abnormal risk level is calculated according to the abnormal prediction results, and the abnormal risk response measures of the corresponding level are selected to automatically generate early warning information of the corresponding abnormal risk level.

7. A DCS early warning system based on intelligent AI visual recognition, 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 real-time operating parameters of the DCS system, including operating status data, alarm information, operation logs, and video monitoring data of each subsystem of the DCS system. The operating status data includes real-time process parameters and equipment operating parameters of each subsystem, and pre-process the real-time operating parameters to obtain real-time operating parameter features in a unified format; The second unit is used to input the real-time operating parameter features into a pre-built intelligent AI visual recognition model, analyze the operating status of the DCS system, identify abnormal modes at the system level, compare the abnormal modes with normal operating modes in a preset normal operating mode library, and calculate the similarity between the abnormal modes and the normal operating modes; The third unit is used to determine that there is an abnormal risk in the current operation of the DCS system if the similarity is lower than the preset abnormal threshold. It uses digital twin simulation technology based on physical mechanisms to simulate the real-time operation scenario of the DCS system, predict the failure mode, cause and probability of the DCS system, and automatically generate warning information of the corresponding level based on the prediction results combined with the preset multi-level warning decision matrix. The warning information is pushed to the intelligent operation and maintenance platform of the DCS system in real time to assist operators in identifying system risks in advance and provide targeted preventive maintenance measures.

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