DCS system autonomous controllable upgrading and system fault prediction method and system

Abstract

The invention provides a DCS system autonomous controllable upgrading and system fault prediction method and system, and relates to the technical field of fault prediction, and the method comprises the steps: obtaining historical operation data and real-time operation data of a DCS system, carrying out the preprocessing, and generating an autonomous controllable upgrading scheme through a deep reinforcement learning algorithm; the method comprises the following steps of: setting specified time for upgrading time, carrying out consistent backup on the state of the DCS, executing software updating, hardware configuration adjustment and system parameter optimization according to upgrading steps, analyzing indexes of the DCS in real time in the process of executing the autonomous controllable upgrading scheme, and recovering the DCS to the state before upgrading when an abnormal condition is detected. And verifying and optimizing the real-time fault prediction result based on a knowledge graph, and generating an explanatory report including a fault prediction and upgrade decision process according to the optimized fault prediction result and the autonomous controllable upgrade scheme.

Description

Technical Field

[0001] The present invention relates to fault prediction technology, and in particular to a DCS system autonomous controllable upgrade and system fault prediction method and system. Background Art

[0002] Distributed control system (DCS) is the core system of industrial automation control, and its safe and stable operation is crucial to the production process. With the expansion of industrial production scale and the improvement of intelligent requirements, DCS systems need to be continuously upgraded to adapt to new application requirements. However, traditional DCS system upgrade methods rely on manual experience decisions, lack scientific evaluation and decision-making mechanisms, and are prone to improper upgrades leading to system failures; existing fault prediction methods are mainly based on single features and simple models, which are difficult to effectively handle the complex failure modes of DCS systems; the upgrade process lacks an effective safety guarantee mechanism, and once an abnormal situation occurs, it may cause the system to paralyze; traditional prediction models consume large computing resources and are difficult to meet real-time prediction needs; prediction results lack interpretability and are difficult to effectively combine with expert knowledge.

[0003] To solve the above problems, the present invention proposes a method for autonomous and controllable upgrade and fault prediction of a DCS system, which realizes intelligent generation of upgrade plans, accurate prediction of faults and safe and controllable system upgrade. Summary of the invention

[0004] The embodiments of the present invention provide a method and system for autonomous and controllable upgrading of a DCS system and prediction of system faults, which can solve the problems in the prior art.

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

[0006] Provided is a method for autonomous and controllable DCS system upgrade and system fault prediction, including:

[0007] The historical operation data and real-time operation data of the DCS system are obtained and preprocessed, and the time series features, frequency domain features and topological features are extracted from the preprocessed historical operation data and real-time operation data, and the extracted features are weightedly fused to obtain fused features. Based on a hybrid algorithm of a long short-term memory network and a graph convolutional neural network, the fused features are used to perform multi-scale fault prediction to obtain fault prediction results. According to the fault prediction results and the current state of the DCS system, an autonomous and controllable upgrade plan is generated through a deep reinforcement learning algorithm, and the autonomous and controllable upgrade plan includes an upgrade schedule, upgrade content and upgrade steps;

[0008] At the time specified in the upgrade schedule, the DCS system status is backed up consistently, and software updates, hardware configuration adjustments, and system parameter optimizations are performed according to the upgrade steps. In the process of executing the autonomous and controllable upgrade plan, the DCS system indicators are analyzed in real time. When an abnormality is detected, the DCS system is restored to the state before the upgrade. After the upgrade is completed, the DCS system is functionally tested, performance evaluated, and security audited to obtain the operating state of the upgraded DCS system;

[0009] Acquire multi-source heterogeneous data under the operating state of the upgraded DCS system and perform adaptive sampling, use the sampled data to collaboratively train the fault prediction model to obtain a pre-trained model, migrate the pre-trained model knowledge to a lightweight online model, perform real-time fault prediction based on the lightweight online model to obtain a real-time fault prediction result, and simultaneously construct a knowledge graph including historical fault cases, expert experience and equipment manual information, verify and optimize the real-time fault prediction result based on the knowledge graph, and generate an explanatory report including the fault prediction and upgrade decision-making process according to the optimized fault prediction result and the autonomous and controllable upgrade plan.

[0010] In an optional embodiment,

[0011] Based on the hybrid algorithm of long short-term memory network and graph convolutional neural network, the fusion features are used to perform multi-scale fault prediction to obtain fault prediction results. According to the fault prediction results and the current state of the DCS system, an autonomous and controllable upgrade plan is generated through a deep reinforcement learning algorithm, including:

[0012] An initial graph is constructed based on the physical topology of the DCS system, and graph structures of different granularities are generated through a graph coarsening algorithm. A graph attention network is used to extract features on the graph structures of different granularities to obtain multi-scale graph features.

[0013] An improved spatiotemporal attention mechanism is constructed to divide the pre-acquired fusion features into fixed-size blocks, a full attention mechanism is used within the blocks, and sparse connections are used between blocks to obtain spatiotemporal attention features. An adaptive fusion module is designed to dynamically weight the multi-scale graph features and the spatiotemporal attention features using a gating mechanism to obtain the final feature representation.

[0014] Based on the final feature representation, a multi-task learning model is constructed, including a fault prediction task, a system state prediction task, and an anomaly detection task;

[0015] Training the multi-task learning model to obtain a trained fault prediction model, and using the trained fault prediction model to perform multi-scale fault prediction on the DCS system to obtain a fault prediction result;

[0016] The upgrade process of the DCS system is modeled as a Markov decision process, and a state space, an action space and a reward function are defined, wherein the state space includes the version number, configuration parameters, performance indicators and the fault prediction results of each component of the system, and the action space includes executable upgrade operations;

[0017] Construct a deep reinforcement learning model, perform hierarchical design on the action space of the deep reinforcement learning model, divide the actions into high-level strategies and low-level strategies, wherein the high-level strategy determines the overall direction of the upgrade, and the low-level strategy determines the specific upgrade operation, and at the same time construct a priority experience replay mechanism based on the temporal difference error, calculate the temporal difference error of each experience sample, determine the sampling probability according to the error size, give priority to replaying experience samples with a large temporal difference error, train the environmental dynamics model, predict the short-term impact of the upgrade operation, input the prediction result into the deep reinforcement learning model as additional state information, and add safety constraints in the decision-making process of the deep reinforcement learning model to obtain an improved deep reinforcement learning model;

[0018] By using the improved deep reinforcement learning model, based on the fault prediction results and the current state of the DCS system, an autonomous and controllable upgrade plan including upgrade schedule, upgrade content and upgrade steps is generated.

[0019] In an optional embodiment,

[0020] Using the improved deep reinforcement learning model, based on the fault prediction results and the current state of the DCS system, an autonomous and controllable upgrade plan including upgrade schedule, upgrade content and upgrade steps is generated, including:

[0021] Input the current system status, fault prediction results, and system evolution trajectory into the optimized deep reinforcement learning model, and output high-level strategies through the optimized deep reinforcement learning model to determine the overall direction of the upgrade, including the upgrade path and resource allocation plan;

[0022] Based on the output high-level strategy, the optimized deep reinforcement learning model further outputs the low-level strategy to determine the specific upgrade operations, including upgrade sequencing and risk avoidance strategies;

[0023] Combine the output high-level and low-level strategies to form a complete upgrade plan, including upgrade schedule, upgrade content and upgrade steps, and generate component-level and system-level upgrade schedules, version selection matrices and configuration change lists;

[0024] Use a pre-trained multi-scale environmental dynamics model to simulate and validate the generated upgrade scenarios, and evaluate the short-term and long-term impacts of the current upgrade scenarios at different time scales and system levels;

[0025] Design a comprehensive evaluation function to evaluate the upgrade solution in multiple dimensions by combining simulation results, preset performance and safety standards;

[0026] Determine whether the evaluation result meets the requirements. If so, output the final upgrade plan. If not, start the adaptive feedback mechanism, adjust the reward function weight and exploration strategy of the deep reinforcement learning model, and return to the step of inputting the current system status and fault prediction results into the optimized deep reinforcement learning model to regenerate the upgrade plan.

[0027] In an optional embodiment,

[0028] In the process of executing the autonomous and controllable upgrade solution, the DCS system indicators are analyzed in real time. When an abnormality is detected, the DCS system is restored to the state before the upgrade, including:

[0029] Build a digital twin model of the DCS system, which includes a physical model and a data-driven model, and synchronize data between the physical system and the digital twin model through edge computing nodes;

[0030] Based on the digital twin model, a large amount of upgrade scenario data is generated by using the Monte Carlo method, including normal operation, abnormal conditions and extreme conditions. The upgrade scenario data is used to train the upgrade strategy agent to obtain an initial upgrade strategy.

[0031] The initial upgrade strategy is input into the multi-objective optimization model to construct a multi-objective function including upgrade efficiency, system stability and resource consumption. The adaptive weight adjustment mechanism is used to dynamically adjust the importance of each goal according to the current state of the system. By solving the multi-objective optimization model, the optimized upgrade strategy is finally output.

[0032] Based on the optimization upgrade strategy, an autonomous and controllable upgrade plan is generated. In the process of executing the autonomous and controllable upgrade plan, a multimodal sensor network and a software-defined network are deployed to collect the operation status data of the DCS system in real time, and the operation status data is processed in real time and adaptively sampled to generate system operation status information. Based on the system operation status information, a multi-level anomaly detection system is constructed to analyze the indicators of the DCS system in real time and output anomaly detection results;

[0033] When the abnormality detection result indicates that an abnormality exists, a multi-level rollback mechanism is triggered to restore the DCS system to the state before the upgrade. If no abnormality is detected, the autonomous and controllable upgrade plan continues to be executed until completion.

[0034] In an optional embodiment,

[0035] Based on the system operation status information, a multi-level anomaly detection system is constructed to analyze the indicators of the DCS system in real time, and the output anomaly detection results include:

[0036] Construct the first-layer anomaly detection model, take the acquired system operation status information as input, calculate the statistical characteristics of system component status and indicators through statistical analysis methods, and obtain the baseline model of normal system operation;

[0037] Using the constructed first-layer anomaly detection model, the current system operation status information is compared with the benchmark model in real time to calculate the deviation value. When the deviation value exceeds the preset threshold, the second-layer anomaly detection is triggered;

[0038] Build a second-layer anomaly detection model that takes the status and indicators of system components that trigger anomalies as input, trains a classifier through a machine learning algorithm, and identifies the type and severity of anomalies;

[0039] Based on the output results of the second-layer anomaly detection model, a third-layer anomaly detection model is constructed. The third-layer anomaly detection model adopts complex network theory to map abnormal components and related indicators into network nodes, establishes connections between nodes according to the probability of abnormal propagation, performs topological analysis on the complex network constructed by the third-layer anomaly detection model, calculates node centrality and community structure, and identifies target abnormal nodes and risk areas;

[0040] According to the analysis results of the third-layer anomaly detection model, an anomaly propagation prediction model is established to simulate the diffusion process of anomalies in the system, predict other components and indicators that may be affected, and finally output the anomaly detection results.

[0041] In an optional embodiment,

[0042] Construct a knowledge graph containing historical fault cases, expert experience, and equipment manual information, verify and optimize the real-time fault prediction results based on the knowledge graph, and generate an explanatory report containing the fault prediction and upgrade decision-making process based on the optimized fault prediction results and the autonomous and controllable upgrade plan, including:

[0043] Obtain historical fault cases, expert experience rules and equipment manual information to build a multimodal knowledge graph, which includes physical topological relationships, equipment operating parameters, fault evolution paths and maintenance decision information;

[0044] A conditional independence test is performed based on historical failure cases in the multimodal knowledge graph, an initial causal structure between variables is constructed using a kernel method, and physical connection constraints, expert rule constraints, and temporal evolution constraints are extracted from the multimodal knowledge graph, and the constraints are fused with the initial causal structure to generate a causal adjacency matrix;

[0045] A nonlinear structural equation model is constructed using a neural network, and the parameters of the quantitative relationship between variables are learned based on the causal adjacency matrix, and a regularization term is introduced for optimization to obtain a structural equation model that describes the evolution process of equipment failure.

[0046] Mapping the obtained real-time fault prediction results to the corresponding nodes of the structural equation model, calculating the intervention effect and obtaining the prediction deviation through counterfactual reasoning;

[0047] The correction amount is calculated according to the prediction deviation, and the real-time fault prediction result is optimized and adjusted to obtain a corrected fault prediction result. Based on the expert experience rules and equipment manual information in the multimodal knowledge graph, an explanatory report including the fault prediction process and decision-making basis is generated in combination with the corrected fault prediction result.

[0048] In an optional embodiment,

[0049] The obtained real-time fault prediction results are mapped to the corresponding nodes of the structural equation model, the intervention effect is calculated, and the prediction deviation is obtained through counterfactual reasoning, including:

[0050] Establishing a correspondence matrix between the prediction output and the structural equation model nodes, wherein the correspondence matrix includes a direct mapping relationship, an indirect mapping relationship, and a non-mapping relationship between the prediction model output vector and the structural equation model node set;

[0051] According to the corresponding relationship matrix, the output vector mapping of the prediction model is converted into the node value of the structural equation model, and for each target intervention node in the structural equation model, based on the node value, the intervention time window of the target intervention node is determined through time series correlation analysis, and the causal effect value of the target output node corresponding to the target intervention node is calculated using the causal intervention operation rule within the intervention time window;

[0052] Performing time series accumulation on the causal effect value to obtain a corresponding relationship curve between the dynamic intervention intensity of the target intervention node and the causal effect value, and identifying a key intervention threshold based on the inflection point of the curve;

[0053] Constructing a multi-level counterfactual reasoning model, the multi-level counterfactual reasoning model includes node-level counterfactual analysis and path-level counterfactual analysis, and inputting the node values ​​into the multi-level counterfactual reasoning model based on the structural equation model;

[0054] In the node-level counterfactual analysis, the counterfactual prediction value caused by the change of a single node is calculated, and in the path-level counterfactual analysis, the counterfactual prediction value caused by the change of a combination of nodes is calculated;

[0055] The counterfactual predicted value is compared with the actual fault value and the output vector of the prediction model in multiple dimensions, and the final prediction deviation is obtained by weighting based on the node importance.

[0056] According to a second aspect of the embodiments of the present invention,

[0057] Provide a DCS system autonomous controllable upgrade and system fault prediction system, including:

[0058] The first unit is used to obtain and preprocess the historical operation data and real-time operation data of the DCS system, extract time series features, frequency domain features and topological features from the preprocessed historical operation data and real-time operation data, perform weighted fusion on the extracted features to obtain fusion features, perform multi-scale fault prediction using the fusion features based on a hybrid algorithm of a long short-term memory network and a graph convolutional neural network, and obtain a fault prediction result, and generate an autonomous and controllable upgrade plan based on the fault prediction result and the current state of the DCS system through a deep reinforcement learning algorithm, wherein the autonomous and controllable upgrade plan includes an upgrade schedule, upgrade content and upgrade steps;

[0059] The second unit is used to perform a consistency backup of the DCS system state at the time specified by the upgrade schedule, and perform software updates, hardware configuration adjustments and system parameter optimization according to the upgrade steps. In the process of executing the autonomous and controllable upgrade plan, the DCS system indicators are analyzed in real time. When an abnormal situation is detected, the DCS system is restored to the state before the upgrade. After the upgrade is completed, the DCS system is functionally tested, performance evaluated and securely audited to obtain the operating state of the upgraded DCS system;

[0060] The third unit is used to obtain multi-source heterogeneous data under the operating state of the upgraded DCS system and perform adaptive sampling, use the sampled data to collaboratively train the fault prediction model to obtain a pre-trained model, migrate the pre-trained model knowledge to the lightweight online model, perform real-time fault prediction based on the lightweight online model, obtain real-time fault prediction results, and simultaneously construct a knowledge graph including historical fault cases, expert experience and equipment manual information, verify and optimize the real-time fault prediction results based on the knowledge graph, and generate an explanatory report including fault prediction and upgrade decision-making process according to the optimized fault prediction results and the autonomous and controllable upgrade plan.

[0061] According to a third aspect of the embodiments of the present invention,

[0062] An electronic device is provided, comprising:

[0063] processor;

[0064] a memory for storing processor-executable instructions;

[0065] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0066] A fourth aspect of the embodiments of the present invention is:

[0067] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0068] In this embodiment, fault prediction is performed through multi-feature extraction and fusion combined with a hybrid deep learning algorithm, thereby improving prediction accuracy and reliability; deep reinforcement learning is used to automatically generate upgrade plans, reduce reliance on manual decision-making, and improve the scientificity and feasibility of upgrade plans; through state backup and real-time monitoring mechanisms, the security and rollback of the system upgrade process are ensured, and upgrade risks are reduced; real-time prediction is achieved based on a lightweight model based on adaptive sampling and knowledge migration, reducing computing resource consumption and improving prediction efficiency; combined with the verification and optimization mechanism of the knowledge graph, the knowledge of domain experts is integrated to improve the reliability and interpretability of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 A schematic diagram of a flow chart of a method for autonomous and controllable DCS system upgrade and system fault prediction according to an embodiment of the present invention;

[0070] Figure 2 It is a structural diagram of a DCS system autonomous controllable upgrade and system fault prediction system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0073] Figure 1 FIG. 1 is a flow chart of a method for autonomous and controllable DCS system upgrade and system fault prediction according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0074] S101. Obtain historical operation data and real-time operation data of the DCS system and perform preprocessing, extract time series features, frequency domain features and topological features from the preprocessed historical operation data and real-time operation data, perform weighted fusion on the extracted features to obtain fused features, and use the fused features to perform multi-scale fault prediction based on a hybrid algorithm of long short-term memory network and graph convolutional neural network to obtain fault prediction results, and generate an autonomous and controllable upgrade plan based on the fault prediction results and the current state of the DCS system through a deep reinforcement learning algorithm, wherein the autonomous and controllable upgrade plan includes upgrade schedule, upgrade content and upgrade steps.

[0075] First, obtain the historical operation data and real-time operation data of the DCS system. The historical operation data includes the operation logs, alarm information, performance indicators, etc. of each module of the system in the past period of time (such as 6 months). The real-time operation data is the real-time status information of each module of the current system. The acquired data is preprocessed, including data cleaning, missing value processing, outlier detection, etc. For example, for temperature data, the moving average method can be used to fill in missing values, and the 3σ principle can be used to eliminate outliers.

[0076] Next, features are extracted from the preprocessed data. Time series features include statistical features such as the mean, variance, and periodicity of each indicator. Frequency domain features are obtained by Fourier transforming the time series data, including spectrum and power spectrum. Topological features are obtained by constructing a correlation graph between system modules, including node degree and centrality. The extracted features are weighted and fused, with the weights of time series, frequency domain, and topological features being 0.5, 0.3, and 0.2, respectively.

[0077] Then, based on the hybrid algorithm of long short-term memory network and graph convolutional neural network, multi-scale fault prediction is performed using fusion features. Long short-term memory network is used to capture time series features, and graph convolutional neural network is used to process topological features. The prediction results include fault type, occurrence probability and expected occurrence time.

[0078] Based on the prediction results and the current state of the system, a self-controlled upgrade plan is generated through a deep reinforcement learning algorithm. The upgrade schedule takes into account factors such as system load and maintenance window. The upgrade content includes software version updates, hardware configuration adjustments, system parameter optimization, etc. The upgrade steps list the specific operations of each action in detail.

[0079] In an optional embodiment,

[0080] Based on the hybrid algorithm of long short-term memory network and graph convolutional neural network, the fusion features are used to perform multi-scale fault prediction to obtain fault prediction results. According to the fault prediction results and the current state of the DCS system, an autonomous and controllable upgrade plan is generated through a deep reinforcement learning algorithm, including:

[0081] An initial graph is constructed based on the physical topology of the DCS system, and graph structures of different granularities are generated through a graph coarsening algorithm. A graph attention network is used to extract features on the graph structures of different granularities to obtain multi-scale graph features.

[0082] An improved spatiotemporal attention mechanism is constructed to divide the pre-acquired fusion features into fixed-size blocks, a full attention mechanism is used within the blocks, and sparse connections are used between blocks to obtain spatiotemporal attention features. An adaptive fusion module is designed to dynamically weight the multi-scale graph features and the spatiotemporal attention features using a gating mechanism to obtain the final feature representation.

[0083] Based on the final feature representation, a multi-task learning model is constructed, including a fault prediction task, a system state prediction task, and an anomaly detection task;

[0084] Training the multi-task learning model to obtain a trained fault prediction model, and using the trained fault prediction model to perform multi-scale fault prediction on the DCS system to obtain a fault prediction result;

[0085] The upgrade process of the DCS system is modeled as a Markov decision process, and a state space, an action space and a reward function are defined, wherein the state space includes the version number, configuration parameters, performance indicators and the fault prediction results of each component of the system, and the action space includes executable upgrade operations;

[0086] Construct a deep reinforcement learning model, perform hierarchical design on the action space of the deep reinforcement learning model, divide the actions into high-level strategies and low-level strategies, wherein the high-level strategy determines the overall direction of the upgrade, and the low-level strategy determines the specific upgrade operation, and at the same time construct a priority experience replay mechanism based on the temporal difference error, calculate the temporal difference error of each experience sample, determine the sampling probability according to the error size, give priority to replaying experience samples with a large temporal difference error, train the environmental dynamics model, predict the short-term impact of the upgrade operation, input the prediction result into the deep reinforcement learning model as additional state information, and add safety constraints in the decision-making process of the deep reinforcement learning model to obtain an improved deep reinforcement learning model;

[0087] By using the improved deep reinforcement learning model, based on the fault prediction results and the current state of the DCS system, an autonomous and controllable upgrade plan including upgrade schedule, upgrade content and upgrade steps is generated.

[0088] In this implementation, a hybrid algorithm based on long short-term memory network and graph convolutional neural network is used to perform multi-scale fault prediction using fusion features to obtain fault prediction results. Then, based on the fault prediction results and the current state of the DCS system, an autonomous and controllable upgrade plan is generated through a deep reinforcement learning algorithm. The specific implementation process is as follows:

[0089] First, an initial graph is constructed based on the physical topology of the DCS system. The various components in the DCS system (such as controllers, I / O modules, workstations, etc.) are used as nodes in the graph, and the physical connection relationships between components are used as edges. For example, for a DCS system containing 10 controllers, 50 I / O modules, and 5 workstations, an initial graph containing 65 nodes can be constructed.

[0090] Next, the graph coarsening algorithm is used to generate graph structures of different granularities. A multi-layer graph coarsening strategy is adopted, and each layer aggregates adjacent nodes to generate a coarser-grained graph. For example, three granularity levels of graph structures can be generated: the original graph (65 nodes), the medium-grained graph (about 30 nodes), and the coarse-grained graph (about 10 nodes).

[0091] Then, the graph attention network is used to extract features on graph structures of different granularities to obtain multi-scale graph features. For each granularity level of the graph, a neural network containing multiple graph attention layers is applied to extract local and global features of the nodes. For example, for the original graph, a 3-layer graph attention network can be used; for the medium-grained graph, 2 layers are used; and for the coarse-grained graph, 1 layer is used. Finally, 3 scale graph feature vectors are obtained, each with a dimension of 128.

[0092] Next, an improved spatiotemporal attention mechanism is constructed. The pre-acquired fusion features (including time series data and static features) are divided into fixed-size blocks, such as one block for every 32 time steps. The full attention mechanism is used within the block to calculate the attention weight of each time step and other time steps in the block. Sparse connections are used between blocks, and only the attention between adjacent blocks is calculated. In this way, spatiotemporal attention features that take into account long-term and short-term dependencies are obtained.

[0093] Then, an adaptive fusion module is designed to dynamically weight the multi-scale graph features and spatiotemporal attention features using a gating mechanism. Specifically, a multi-layer perceptron network is used to calculate the importance weight of each feature, and the weighted sum of different features is performed according to the weight to obtain the final feature representation. This feature representation integrates the topological structure information and temporal dynamic information of the DCS system.

[0094] Based on the final feature representation, a multi-task learning model is constructed. The model includes three tasks: fault prediction task, system state prediction task and anomaly detection task. The fault prediction task aims to predict the type and probability of possible future faults; the system state prediction task predicts the future trend of key performance indicators; and the anomaly detection task identifies abnormal states in the current system. These three tasks share the underlying feature extraction network, but each has an independent output layer.

[0095] The multi-task learning model is trained using historical operation data as the training set, including system status data, fault records, and performance indicators. The cross-validation method is used to evaluate the model performance and select the optimal hyperparameter configuration. After the training is completed, a model capable of multi-scale fault prediction is obtained.

[0096] The trained fault prediction model is used to perform multi-scale fault prediction on the DCS system. The current system status and historical data are input, and the model outputs fault prediction results at different time scales in the future (such as 1 hour, 1 day, 1 week), including the probability of occurrence of various types of faults, the expected time of occurrence, etc. For example, the model may predict that a controller has an 80% probability of a communication failure within the next 24 hours.

[0097] Next, the upgrade process of the DCS system is modeled as a Markov decision process. Define the state space, action space and reward function. The state space includes the version numbers of each component of the system (such as controller firmware version, workstation operating system version, etc.), configuration parameters (such as communication protocol settings, control algorithm parameters, etc.), performance indicators (such as CPU utilization, network latency, etc.) and the fault prediction results obtained previously. The action space includes executable upgrade operations, such as updating firmware, adjusting configuration parameters, replacing hardware, etc. The reward function is designed based on the degree of performance improvement, security improvement and upgrade cost of the system after the upgrade operation.

[0098] A deep reinforcement learning model is constructed, using a hierarchical reinforcement learning architecture. The action space is divided into high-level strategies and low-level strategies. The high-level strategy determines the overall direction of the upgrade, such as "improving system reliability" or "optimizing network performance". The low-level strategy determines the specific upgrade operations based on the high-level strategy, such as "updating the firmware of controller A to version v2.1" or "adjusting the QoS parameters of network switch B".

[0099] In order to improve learning efficiency, a priority experience replay mechanism based on temporal difference error is constructed. During the training process, the temporal difference error of each experience sample is calculated, that is, the difference between the actual reward and the expected reward. The sampling probability is determined according to the error size, and experience samples with larger temporal difference errors are replayed preferentially. This allows the model to learn key experiences faster.

[0100] At the same time, an environmental dynamics model is trained to predict the short-term impact of the upgrade operation. The model receives the current system state and the proposed upgrade operation as input, and predicts the state change of the system in the short term (such as 1 hour) after the operation is executed. The prediction results are input into the deep reinforcement learning model as additional state information to help the model make more intelligent decisions.

[0101] Add safety constraints to the decision-making process of the deep reinforcement learning model. Define a series of safety rules, such as "no more than 30% of controllers can be upgraded at the same time", "at least 60% of redundant systems must be online", etc. When selecting actions, first filter out actions that violate safety rules, and select the best action from the remaining safe actions.

[0102] Through the above steps, an improved deep reinforcement learning model is obtained. This model can generate intelligent upgrade solutions based on the current state of the system, fault prediction results and safety constraints.

[0103] Finally, the improved deep reinforcement learning model is used to generate an autonomous and controllable upgrade plan based on the fault prediction results and the current status of the DCS system. The plan includes the following contents:

[0104] 1) Upgrade schedule: Specify the execution time of each upgrade step, taking into account factors such as system load, maintenance windows, etc. For example, the upgrade of key components may be scheduled during the low-load period on weekends.

[0105] 2) Upgrade content: List in detail the components, version numbers, and parameter adjustments that need to be upgraded, such as "upgrade the firmware of controllers A, B, and C from v1.5 to v2.0", "adjust the buffer size of switch D to 512KB", etc.

[0106] 3) Upgrade steps: Specific operation steps sorted by priority and dependency. For example, "Step 1: Back up system configuration; Step 2: Upgrade controller A; Step 3: Verify controller A function; Step 4: Upgrade controller B..." etc.

[0107] In this embodiment, the deep features of the DCS system are extracted through a multi-scale graph structure and an improved spatiotemporal attention mechanism, which enables accurate identification and prediction of fault modes. A multi-task learning framework is used to simultaneously optimize fault prediction, state prediction, and anomaly detection tasks, thereby improving the generalization ability and prediction accuracy of the model. An upgrade strategy is formulated through a hierarchically designed deep reinforcement learning model, combined with priority experience replay and environmental dynamics prediction, to ensure the reliability and security of the upgrade solution. This solution not only improves the accuracy and timeliness of fault prediction, but also realizes intelligent decision-making and risk control in the upgrade process, making the maintenance and upgrade of the DCS system more efficient and controllable. At the same time, the feature extraction method based on the graph structure improves the model's adaptability to changes in system topology and enhances the robustness of prediction and decision-making.

[0108] In an optional embodiment,

[0109] Using the improved deep reinforcement learning model, based on the fault prediction results and the current state of the DCS system, an autonomous and controllable upgrade plan including upgrade schedule, upgrade content and upgrade steps is generated, including:

[0110] Input the current system status, fault prediction results, and system evolution trajectory into the optimized deep reinforcement learning model, and output high-level strategies through the optimized deep reinforcement learning model to determine the overall direction of the upgrade, including the upgrade path and resource allocation plan;

[0111] Based on the output high-level strategy, the optimized deep reinforcement learning model further outputs the low-level strategy to determine the specific upgrade operations, including upgrade sequencing and risk avoidance strategies;

[0112] Combine the output high-level and low-level strategies to form a complete upgrade plan, including upgrade schedule, upgrade content and upgrade steps, and generate component-level and system-level upgrade schedules, version selection matrices and configuration change lists;

[0113] Use a pre-trained multi-scale environmental dynamics model to simulate and validate the generated upgrade scenarios, and evaluate the short-term and long-term impacts of the current upgrade scenarios at different time scales and system levels;

[0114] Design a comprehensive evaluation function to evaluate the upgrade solution in multiple dimensions by combining simulation results, preset performance and safety standards;

[0115] Determine whether the evaluation result meets the requirements. If so, output the final upgrade plan. If not, start the adaptive feedback mechanism, adjust the reward function weight and exploration strategy of the deep reinforcement learning model, and return to the step of inputting the current system status and fault prediction results into the optimized deep reinforcement learning model to regenerate the upgrade plan.

[0116] Exemplarily, first, the current system state, fault prediction results and system evolution trajectory are used as input data and sent to the optimized deep reinforcement learning model. The model adopts a multi-layer neural network structure, including an input layer, multiple hidden layers and an output layer. The input layer receives multi-dimensional features such as system state, fault prediction and evolution trajectory, the hidden layer extracts and combines features through nonlinear activation functions, and the output layer generates high-level strategies. Specifically, the input data includes the operating parameters, failure probability, remaining life prediction values ​​of each component of the system, and load forecasts for a period of time in the future. The high-level strategy output by the model mainly includes two parts: upgrade path and resource allocation plan. The upgrade path defines the upgrade order and priority of each component of the system, and the resource allocation plan gives allocation suggestions for various resources (such as manpower, equipment, time, etc.) during the upgrade process.

[0117] For example, for a DCS system with 100 nodes, the input data may include parameters such as CPU usage, memory usage, network traffic, failure rate, etc. of each node, as well as the business load forecast for the next 7 days. The high-level strategy output by the model may recommend upgrading the core switch first, then the key business server, and finally the ordinary terminal equipment, while giving a specific manpower and time allocation plan.

[0118] Next, based on the output high-level strategy, the deep reinforcement learning model further generates low-level strategies to determine the specific upgrade operation details. The low-level strategy mainly includes two aspects: upgrade sorting and risk avoidance. The upgrade sorting gives the upgrade order accurate to the level of a single component, and the risk avoidance strategy formulates countermeasures for various risks that may arise during the upgrade process. In this step, the model fully considers factors such as the dependencies between components, business continuity requirements, and backup and recovery mechanisms.

[0119] Taking the DCS system above as an example, the low-level policy may give the precise upgrade order of 100 nodes and formulate specific plans on how to deal with risks such as network interruption and data loss that may occur during the upgrade process. For example, for a critical business server, the policy may recommend a full backup first, then upgrade during the business downturn, and prepare a rollback plan.

[0120] Combine high-level and low-level strategies to form a complete upgrade plan. The plan includes a detailed upgrade schedule, a list of upgrade contents, and specific upgrade step instructions. At the same time, component-level and system-level upgrade schedules, version selection matrices, and configuration change lists are also generated. The upgrade schedule gives the specific upgrade time window for each component, the version selection matrix indicates the specific version that each component should be upgraded to, and the configuration change list lists the configuration items that need to be modified and their new values ​​in detail.

[0121] For the example system, the final upgrade plan may include a detailed 14-day upgrade plan, which lists the specific devices that need to be upgraded each day, the upgrade content, the time and manpower required, etc. The version selection matrix may indicate that the core switch needs to be upgraded to the latest v5.2 version firmware, while some terminal devices only need to be upgraded to v4.8. The configuration change list may contain hundreds of configuration items that need to be modified, such as network addresses, security policies, load balancing parameters, etc.

[0122] In order to verify the feasibility and effectiveness of the generated upgrade plan, a pre-trained multi-scale environmental dynamics model is used for simulation verification. This model can simulate the behavior of the system at different time scales (from seconds to months) and different system levels (from single components to the overall system). Through simulation, the short-term and long-term impact of the current upgrade plan is evaluated, including system performance, availability, security and other aspects.

[0123] Simulation verification may find that in the example system, the upgrade of a critical business server may cause a 15-minute service interruption, which exceeds the maximum allowed interruption time. At the same time, in the long run, the new version of the firmware may increase the overall system throughput by 20%, but also increase energy consumption by 10%.

[0124] Based on the simulation results, a comprehensive evaluation function is designed to evaluate the upgrade solution in multiple dimensions. The evaluation function takes into account multiple factors, including performance improvement after the upgrade, security improvement, upgrade cost, risk level, business interruption time, etc. Each factor is assigned a weight to reflect its relative importance. The evaluation function combines these factors to obtain an overall score.

[0125] For example, for the example system, the evaluation function may give a 30% weight to performance improvement, a 25% weight to security improvement, a 20% weight to upgrade cost, a 15% weight to risk level, and a 10% weight to business interruption time. Based on this weight distribution, the overall score of the upgrade solution is calculated.

[0126] Finally, determine whether the evaluation results meet the preset requirements. If they do, output the final upgrade plan. If not, start the adaptive feedback mechanism to adjust the reward function weight and exploration strategy of the deep reinforcement learning model. After the adjustment, re-execute the entire process, starting from inputting the current system status and fault prediction results into the optimized deep reinforcement learning model, and regenerate the upgrade plan.

[0127] In the example system, if the overall score is found to be lower than the preset threshold, such as only 75 points (out of 100 points), the system will automatically adjust the reward function, which may increase the emphasis on business continuity, and adjust the exploration strategy to find a better upgrade path. Then regenerate the upgrade plan until a plan that meets the requirements is obtained.

[0128] In this embodiment, accurate planning of the upgrade plan is achieved through hierarchical deep reinforcement learning strategy design. The high-level strategy ensures the strategic nature of the upgrade direction, and the low-level strategy ensures the executability of specific operations. A multi-scale environmental dynamics model is introduced for simulation verification to effectively evaluate the impact of each level of the upgrade plan. The designed comprehensive evaluation function and adaptive feedback mechanism ensure the dynamic optimization and security of the upgrade plan. The scheme can automatically generate a fine-grained upgrade schedule and configuration change list, which improves the accuracy and operability of the upgrade plan. Through multiple rounds of evaluation and optimization, it is ensured that the upgrade plan meets performance and safety standards, reducing the upgrade risk. At the same time, the adaptive feedback mechanism enables the model to continuously adjust and improve the upgrade strategy according to the evaluation results, improving the adaptability and reliability of the scheme.

[0129] S102. At the time specified in the upgrade schedule, the DCS system status is backed up consistently, and software updates, hardware configuration adjustments and system parameter optimizations are performed according to the upgrade steps. During the execution of the autonomous and controllable upgrade plan, the DCS system indicators are analyzed in real time. When an abnormality is detected, the DCS system is restored to the state before the upgrade. After the upgrade is completed, the DCS system is subjected to functional testing, performance evaluation and security audit to obtain the operating state of the upgraded DCS system.

[0130] In an optional embodiment,

[0131] In the process of executing the autonomous and controllable upgrade solution, the DCS system indicators are analyzed in real time. When an abnormality is detected, the DCS system is restored to the state before the upgrade, including:

[0132] Build a digital twin model of the DCS system, which includes a physical model and a data-driven model, and synchronize data between the physical system and the digital twin model through edge computing nodes;

[0133] Based on the digital twin model, a large amount of upgrade scenario data is generated by using the Monte Carlo method, including normal operation, abnormal conditions and extreme conditions. The upgrade scenario data is used to train the upgrade strategy agent to obtain an initial upgrade strategy.

[0134] The initial upgrade strategy is input into the multi-objective optimization model to construct a multi-objective function including upgrade efficiency, system stability and resource consumption. The adaptive weight adjustment mechanism is used to dynamically adjust the importance of each goal according to the current state of the system. By solving the multi-objective optimization model, the optimized upgrade strategy is finally output.

[0135] Based on the optimization upgrade strategy, an autonomous and controllable upgrade plan is generated. In the process of executing the autonomous and controllable upgrade plan, a multimodal sensor network and a software-defined network are deployed to collect the operation status data of the DCS system in real time, and the operation status data is processed in real time and adaptively sampled to generate system operation status information. Based on the system operation status information, a multi-level anomaly detection system is constructed to analyze the indicators of the DCS system in real time and output anomaly detection results;

[0136] When the abnormality detection result indicates that an abnormality exists, a multi-level rollback mechanism is triggered to restore the DCS system to the state before the upgrade. If no abnormality is detected, the autonomous and controllable upgrade plan continues to be executed until completion.

[0137] Exemplarily, a digital twin model of the DCS system is first constructed. The model consists of two parts: a physical model and a data-driven model. The physical model is modeled based on the physical characteristics and process flow of the system, including equipment models, pipeline models, valve models, etc. The data-driven model is based on historical operating data and uses machine learning methods such as long short-term memory networks (LSTM) to build a system dynamic characteristic model. A lightweight digital twin engine is deployed on the edge computing node to achieve real-time data synchronization between the physical system and the digital twin model. For example, temperature, pressure, flow and other data are collected from the site every 100 milliseconds, and the data is transmitted to the digital twin model in real time for status update.

[0138] Next, based on the constructed digital twin model, the Monte Carlo method is used to generate a large amount of upgrade scenario data. Specifically, by randomly injecting faults into the digital twin model and setting extreme working conditions, upgrade scenario data including normal operation, abnormal conditions and extreme conditions can be simulated and generated. For example, 100,000 sets of different upgrade scenario data can be generated, including 90% normal operation data, 9% abnormal condition data and 1% extreme condition data. Using these upgrade scenario data, deep reinforcement learning methods such as the Proximal Policy Optimization (PPO) algorithm are used to train the upgrade strategy agent to obtain the initial upgrade strategy. During the training process, the upgrade completion time, system stability indicators, etc. are used as reward functions to guide the agent to learn the optimal upgrade strategy.

[0139] Then, the obtained initial upgrade strategy is input into the multi-objective optimization model for further optimization. A multi-objective function including upgrade efficiency, system stability and resource consumption is constructed. The upgrade efficiency can be measured by the upgrade completion time, the system stability can be characterized by the fluctuation amplitude of key process variables, and the resource consumption can be measured by indicators such as CPU usage and memory usage. An adaptive weight adjustment mechanism is adopted to dynamically adjust the importance of each goal according to the current state of the system. For example, when the system is under high load, the weight of system stability can be appropriately increased. By solving the multi-objective optimization model, an optimized upgrade strategy that balances multiple goals is finally output.

[0140] Based on the optimized upgrade strategy, a detailed autonomous and controllable upgrade plan is generated, including the upgrade sequence, upgrade time window, rollback strategy, etc. In the process of executing the upgrade plan, a multimodal sensor network and a software-defined network are deployed to collect the operating status data of the DCS system in real time. The multimodal sensor network includes temperature sensors, pressure sensors, vibration sensors, etc., and the software-defined network is used to flexibly schedule network resources to ensure the real-time and reliability of data transmission. The collected operating status data is processed in real time and adaptively sampled to generate system operating status information. Real-time stream processing uses stream computing engines such as Apache Flink to achieve millisecond-level data processing delays. Adaptive sampling dynamically adjusts the sampling frequency according to the data change rate to reduce system load while ensuring data quality.

[0141] Based on the generated system operation status information, a multi-level anomaly detection system is constructed to analyze the indicators of the DCS system in real time and output the anomaly detection results. The multi-level anomaly detection system includes three levels: the first level uses simple statistical methods such as the 3σ principle for rapid anomaly detection; the second level uses machine learning methods such as the isolation forest algorithm for fine anomaly detection; the third level uses an expert system for knowledge-driven anomaly diagnosis. The three levels work together to ensure detection speed and improve detection accuracy.

[0142] When the anomaly detection results indicate an abnormal situation, the multi-level rollback mechanism is triggered to restore the DCS system to the state before the upgrade. The multi-level rollback mechanism includes three levels: software rollback, data rollback, and hardware rollback. Software rollback is achieved by rolling back the software version; data rollback is achieved by restoring data snapshots; and hardware rollback is achieved by switching to backup hardware. Different levels of rollback mechanisms are triggered step by step according to the severity of the anomaly to maximize the availability of the system. For example, when an anomaly is detected in a software module, software rollback is tried first; if the software rollback fails, data rollback is performed; if data rollback still cannot solve the problem, hardware rollback is performed last.

[0143] If no abnormal situation is detected, the autonomous and controllable upgrade plan will continue to be executed until it is completed. During the entire upgrade process, the system continuously monitors various indicators and dynamically adjusts the upgrade strategy according to the actual situation to ensure the safety and controllability of the upgrade. For example, when a high system load is detected, the upgrade speed can be automatically reduced to avoid affecting production. In this way, a safe, reliable, and autonomous upgrade of the DCS system is achieved.

[0144] In this embodiment, real-time synchronization between the physical system and the digital model is achieved through digital twin technology and edge computing, which improves the accuracy of system state perception. The Monte Carlo method is used to generate rich upgrade scenario data, which enhances the robustness of the upgrade strategy. Multi-objective optimization and adaptive weight adjustment ensure the balance of the upgrade strategy in multiple dimensions such as efficiency, stability and resource consumption. The deployment of multimodal sensor networks and software-defined networks realizes all-round monitoring of the system status, and the multi-level anomaly detection system provides timely and reliable anomaly identification capabilities. When an anomaly occurs, the multi-level rollback mechanism can quickly restore the system to a safe state, effectively reducing the upgrade risk and improving the reliability and stability of the system. The overall solution realizes intelligent management and security of the upgrade process.

[0145] In an optional embodiment,

[0146] Based on the system operation status information, a multi-level anomaly detection system is constructed to analyze the indicators of the DCS system in real time, and the output anomaly detection results include:

[0147] Construct the first-layer anomaly detection model, take the acquired system operation status information as input, calculate the statistical characteristics of system component status and indicators through statistical analysis methods, and obtain the baseline model of normal system operation;

[0148] Using the constructed first-layer anomaly detection model, the current system operation status information is compared with the benchmark model in real time to calculate the deviation value. When the deviation value exceeds the preset threshold, the second-layer anomaly detection is triggered;

[0149] Build a second-layer anomaly detection model that takes the status and indicators of system components that trigger anomalies as input, trains a classifier through a machine learning algorithm, and identifies the type and severity of anomalies;

[0150] Based on the output results of the second-layer anomaly detection model, a third-layer anomaly detection model is constructed. The third-layer anomaly detection model adopts complex network theory to map abnormal components and related indicators into network nodes, establishes connections between nodes according to the probability of abnormal propagation, performs topological analysis on the complex network constructed by the third-layer anomaly detection model, calculates node centrality and community structure, and identifies target abnormal nodes and risk areas;

[0151] According to the analysis results of the third-layer anomaly detection model, an anomaly propagation prediction model is established to simulate the diffusion process of anomalies in the system, predict other components and indicators that may be affected, and finally output the anomaly detection results.

[0152] Exemplarily, first, a first-layer anomaly detection model is constructed. The model takes the acquired system operation status information as input, and calculates the statistical characteristics of the status of each component of the system and key indicators through statistical analysis methods, including mean, variance, skewness, kurtosis, etc. For example, for the CPU usage indicator, its average value, standard deviation, maximum value, minimum value and other characteristics within 24 hours can be calculated. Based on these statistical characteristics, a baseline model for normal system operation is established.

[0153] Next, the first-layer anomaly detection model is used to compare the current system operation status with the baseline model in real time. Specifically, the deviation between the current value and the baseline model is calculated for each monitoring indicator, such as (current value - mean) / standard deviation. When the deviation value of an indicator exceeds the preset threshold (such as 3 times the standard deviation), the second-layer anomaly detection is triggered. For example, if the current value of the CPU utilization is 95%, and the mean in the baseline model is 60% and the standard deviation is 10%, then the deviation value is 3.5, which exceeds the 3 times standard deviation threshold, triggering further detection.

[0154] Then, the second-layer anomaly detection model is built. This model uses the status and indicators of the system components that trigger the anomaly as input features, and trains the classifier through machine learning algorithms (such as support vector machines, random forests, etc.) to identify the type and severity of the anomaly. The training data of the classifier comes from historical abnormal event records, including various known types of abnormal samples. The model outputs the specific type of anomaly (such as resource exhaustion, network congestion, etc.) and severity (such as minor, moderate, severe, etc.).

[0155] Based on the output of the second-layer model, the third-layer anomaly detection model is further constructed. This model uses complex network theory to map the detected abnormal components and related indicators into network nodes. Connections between nodes are established based on the dependencies between components and the correlations between indicators, as well as the possible propagation paths of anomalies. For example, if an abnormal CPU usage of a Web server is detected, it will be connected to related application servers, database servers and other nodes.

[0156] Perform topological analysis on the constructed complex network and calculate the centrality index of each node (such as degree centrality, betweenness centrality, etc.) and the community structure of the network. Nodes with high centrality are often key anomaly sources or potential risk points. Through the community discovery algorithm, highly associated node groups, i.e. risk areas, can be identified. For example, it may be found that the web server, application server, and load balancer form a community, indicating that this is a potential fault propagation area.

[0157] Finally, based on the analysis results of the third-layer model, an anomaly propagation prediction model is established. This model simulates the diffusion process of anomalies in the system and predicts other components and indicators that may be affected. Specifically, starting from the anomaly source node, a random walk is performed based on the connection weights between nodes (representing the probability of anomaly propagation) to generate multiple possible propagation paths. The frequency of each node being visited is counted to obtain the possibility of them being affected.

[0158] The analysis results of the above multi-level models are combined to finally output an anomaly detection report. The report content includes: the type and severity of the detected anomaly, the source component of the anomaly, the key indicators affected, the potential risk area, the expected propagation path, and the recommended disposal measures. For example, the report may point out: "A serious CPU overload anomaly has been detected in the Web server cluster, which has affected the application response time and is expected to spread to the database server within 10 minutes. It is recommended to immediately load balance and increase server resources."

[0159] This multi-level anomaly detection method can comprehensively analyze the system status, accurately identify anomalies, and predict their scope of impact, providing strong support for timely handling of system failures.

[0160] In this embodiment, accurate identification and location of system anomalies are achieved through a multi-level anomaly detection architecture. The first layer establishes a benchmark model based on statistical analysis to quickly discover system deviations; the second layer uses machine learning to classify anomalies and accurately determine the type and degree of anomalies; the third layer analyzes the propagation characteristics of anomalies based on complex network theory to effectively identify key risk nodes and areas. The progressive detection mechanism improves the accuracy and reliability of anomaly identification and avoids false positives and false negatives. The anomaly propagation prediction model is used to simulate the spread of anomalies, achieve early warning of system risks, and provide a basis for timely prevention and control measures. The overall solution enhances the system's anomaly monitoring capabilities and improves the accuracy and foresight of risk management.

[0161] S103. Acquire multi-source heterogeneous data under the operating state of the upgraded DCS system and perform adaptive sampling, use the sampled data to collaboratively train the fault prediction model to obtain a pre-trained model, migrate the pre-trained model knowledge to a lightweight online model, perform real-time fault prediction based on the lightweight online model, and obtain real-time fault prediction results. At the same time, construct a knowledge graph containing historical fault cases, expert experience, and equipment manual information, verify and optimize the real-time fault prediction results based on the knowledge graph, and generate an explanatory report containing fault prediction and upgrade decision-making processes according to the optimized fault prediction results and the autonomous and controllable upgrade plan.

[0162] Obtain multi-source heterogeneous data of the upgraded system, including device status, network traffic, application logs, etc. Adopt adaptive sampling methods to increase the sampling frequency of important indicators. Use sampling data to collaboratively train the fault prediction model to improve model accuracy. Migrate pre-trained model knowledge to lightweight online models to achieve real-time fault prediction.

[0163] At the same time, a knowledge graph is constructed, which includes historical failure cases, expert experience, and equipment manual information. Based on the knowledge graph, the real-time failure prediction results are verified and optimized to improve the prediction accuracy. Finally, an explanatory report containing the failure prediction and upgrade decision-making process is generated, detailing the prediction basis and upgrade plan formulation logic.

[0164] In an optional embodiment,

[0165] Construct a knowledge graph containing historical fault cases, expert experience, and equipment manual information, verify and optimize the real-time fault prediction results based on the knowledge graph, and generate an explanatory report containing the fault prediction and upgrade decision-making process based on the optimized fault prediction results and the autonomous and controllable upgrade plan, including:

[0166] Obtain historical fault cases, expert experience rules and equipment manual information to build a multimodal knowledge graph, which includes physical topological relationships, equipment operating parameters, fault evolution paths and maintenance decision information;

[0167] A conditional independence test is performed based on historical failure cases in the multimodal knowledge graph, an initial causal structure between variables is constructed using a kernel method, and physical connection constraints, expert rule constraints, and temporal evolution constraints are extracted from the multimodal knowledge graph, and the constraints are fused with the initial causal structure to generate a causal adjacency matrix;

[0168] A nonlinear structural equation model is constructed using a neural network, and the parameters of the quantitative relationship between variables are learned based on the causal adjacency matrix, and a regularization term is introduced for optimization to obtain a structural equation model that describes the evolution process of equipment failure.

[0169] Mapping the obtained real-time fault prediction results to the corresponding nodes of the structural equation model, calculating the intervention effect and obtaining the prediction deviation through counterfactual reasoning;

[0170] The correction amount is calculated according to the prediction deviation, and the real-time fault prediction result is optimized and adjusted to obtain a corrected fault prediction result. Based on the expert experience rules and equipment manual information in the multimodal knowledge graph, an explanatory report including the fault prediction process and decision-making basis is generated in combination with the corrected fault prediction result.

[0171] Exemplarily, first, historical fault cases, expert experience rules and equipment manual information are obtained to construct a multimodal knowledge graph. The knowledge graph contains physical topological relationships, equipment operating parameters, fault evolution paths and maintenance decision information. Specifically, a graph database such as Neo4j can be used to store knowledge graphs, using nodes to represent entities such as equipment, components, and fault types, and edges to represent relationships between entities. For example, the nodes "compressor" and "bearing" can be used to represent equipment components, and the edge "contains" can be used to connect them; the node "excessive vibration" can be used to represent the fault type, and the edge "may lead to" can be used to connect "bearing" and "excessive vibration". At the same time, attribute information such as equipment parameters and fault characteristics are attached to the nodes.

[0172] Next, based on the historical fault case data in the constructed multimodal knowledge graph, a conditional independence test is performed, and the kernel method is used to construct the initial causal structure between variables. Specifically, the kernel independence test method, such as the Hilbert-Schmidt Independence Criterion (HSIC), can be used to calculate the independence statistic between variable pairs. If the statistic is less than the preset threshold, the two variables are considered conditionally independent. Based on the test results, the initial variable causal relationship diagram is drawn.

[0173] Then, physical connection constraints, expert rule constraints, and temporal evolution constraints are extracted from the multimodal knowledge graph. Physical connection constraints can be obtained from the equipment topology, such as bearing vibration will inevitably affect compressor vibration. Expert rule constraints can be extracted from empirical rules, such as "if the bearing temperature rises abnormally, a bearing failure may occur." Temporal evolution constraints can be extracted from historical fault evolution paths, such as "bearing vibration abnormalities may cause bearing damage after a period of time." These constraints are fused with the initial causal structure to generate a causal adjacency matrix. During the fusion process, if there is a conflict between the constraints and the initial structure, the initial structure is modified based on the constraints.

[0174] A nonlinear structural equation model is constructed using a neural network, and the parameters of the quantitative relationship between variables are learned based on the causal adjacency matrix. Specifically, a multilayer perceptron (MLP) can be used as the neural network structure, with the input layer nodes corresponding to the parent node variables in the causal graph, and the output layer nodes corresponding to the child node variables. The network parameters are learned by minimizing the prediction error. At the same time, the L1 regularization term is introduced to sparse the network weights to avoid overfitting. After optimization, a structural equation model describing the evolution process of equipment failure is obtained.

[0175] The fault prediction results obtained in real time are mapped to the corresponding nodes of the structural equation model. For example, if the prediction result is "excessive bearing vibration", the information is assigned to the "bearing vibration" node in the model. Then the intervention effect is calculated, that is, the change in the model output after the node value is fixed. Through counterfactual reasoning, the output of the model is compared with the actual output when the prediction result is not established to obtain the prediction deviation.

[0176] The initial prediction results are corrected based on the calculated prediction deviation. For example, if the deviation shows that the actual failure probability is lower than the predicted value, the predicted failure probability is appropriately reduced. Each prediction result is adjusted separately to obtain a corrected failure prediction result.

[0177] Finally, based on the expert experience rules and equipment manual information stored in the multimodal knowledge graph, combined with the corrected fault prediction results, an explanatory report containing the fault prediction process and decision-making basis is generated. The report details the historical data based on the prediction, the algorithm model used, the expert rules applied, and the confidence of the final prediction results. At the same time, according to the maintenance guidance in the equipment manual, corresponding processing suggestions are given for the predicted potential faults.

[0178] Through the above steps, the fault prediction results based on the knowledge graph are verified and optimized, and explainable prediction and decision reports are generated, providing strong decision-making support for equipment maintenance personnel.

[0179] In this embodiment, historical cases, expert experience and equipment document information are integrated through multimodal knowledge graphs to establish a complete knowledge system. The causal structure constructed based on conditional independence test and kernel method, combined with physical constraints and expert rules, improves the accuracy of causal relationship identification. The nonlinear structural equation model realizes the precise modeling of complex relationships between variables, and counterfactual reasoning and intervention effect analysis enhance the reliability of prediction results. The prediction results are verified and optimized through knowledge graphs, which improves the accuracy and interpretability of fault prediction. The final explanatory report integrates the prediction process and decision-making basis, making the prediction results highly understandable and traceable, and providing reliable knowledge support for system maintenance and decision-making.

[0180] In an optional embodiment,

[0181] The obtained real-time fault prediction results are mapped to the corresponding nodes of the structural equation model, the intervention effect is calculated, and the prediction deviation is obtained through counterfactual reasoning, including:

[0182] Establishing a correspondence matrix between the prediction output and the structural equation model nodes, wherein the correspondence matrix includes a direct mapping relationship, an indirect mapping relationship, and a non-mapping relationship between the prediction model output vector and the structural equation model node set;

[0183] According to the corresponding relationship matrix, the output vector mapping of the prediction model is converted into the node value of the structural equation model, and for each target intervention node in the structural equation model, based on the node value, the intervention time window of the target intervention node is determined through time series correlation analysis, and the causal effect value of the target output node corresponding to the target intervention node is calculated using the causal intervention operation rule within the intervention time window;

[0184] Performing time series accumulation on the causal effect value to obtain a corresponding relationship curve between the dynamic intervention intensity of the target intervention node and the causal effect value, and identifying a key intervention threshold based on the inflection point of the curve;

[0185] Constructing a multi-level counterfactual reasoning model, the multi-level counterfactual reasoning model includes node-level counterfactual analysis and path-level counterfactual analysis, and inputting the node values ​​into the multi-level counterfactual reasoning model based on the structural equation model;

[0186] In the node-level counterfactual analysis, the counterfactual prediction value caused by the change of a single node is calculated, and in the path-level counterfactual analysis, the counterfactual prediction value caused by the change of a combination of nodes is calculated;

[0187] The counterfactual predicted value is compared with the actual fault value and the output vector of the prediction model in multiple dimensions, and the final prediction deviation is obtained by weighting based on the node importance.

[0188] Exemplarily, first, a correspondence matrix between the prediction output and the structural equation model nodes is established. The matrix contains direct mapping relationships, indirect mapping relationships, and no mapping relationships between the prediction model output vector and the set of structural equation model nodes. For example, for a certain equipment failure prediction system, the prediction model output includes 10 parameters such as temperature, pressure, and vibration, and the structural equation model contains 20 nodes. A 10×20 correspondence matrix can be constructed, where 1 represents a direct mapping relationship, 0.5 represents an indirect mapping relationship, and 0 represents no mapping relationship.

[0189] Next, according to the correspondence matrix, the output vector mapping of the prediction model is converted into the node value of the structural equation model. For each target intervention node in the structural equation model, the intervention time window of the target intervention node is determined by time series correlation analysis based on the converted node value. For example, for the temperature node, its change trend in the past 24 hours can be analyzed to determine the intervention time window as the last 6 hours. Within this intervention time window, the causal effect value of the target output node corresponding to the target intervention node is calculated using the causal intervention operation rule.

[0190] Then, the causal effect values ​​are accumulated in time series to obtain the corresponding relationship curve between the dynamic intervention intensity and the causal effect value of the target intervention node. The key intervention threshold is identified based on the inflection point of the curve. For example, the intervention intensity of the temperature node ranges from 0 to 100%, and the causal effect value is calculated every 10% increase. The corresponding relationship curve is drawn, and the point with the largest change in the slope of the curve is found as the key intervention threshold.

[0191] Next, a multi-level counterfactual reasoning model is constructed, including node-level counterfactual analysis and path-level counterfactual analysis. The node values ​​of the structural equation model are input into the multi-level counterfactual reasoning model. In the node-level counterfactual analysis, the counterfactual prediction value caused by the change of a single node is calculated. For example, the value of the temperature node is increased by 10%, and the other nodes are kept unchanged, and the counterfactual prediction result is calculated. In the path-level counterfactual analysis, the counterfactual prediction value caused by the change of the node combination is calculated. For example, the values ​​of the temperature and pressure nodes are changed at the same time, and the counterfactual prediction result is calculated.

[0192] Finally, the counterfactual predicted value is compared with the actual fault value and the output vector of the prediction model in multiple dimensions. Based on the pre-set node importance weights, the comparison results are weighted to obtain the final prediction deviation. For example, the importance of the temperature node can be set to 0.3, the pressure node to 0.2, the vibration node to 0.1, etc., and the weighted calculation is used to obtain the comprehensive prediction deviation.

[0193] Through the above steps, it is possible to map the real-time fault prediction results to the structural equation model, calculate the intervention effect, and obtain the prediction deviation through counterfactual reasoning, providing a basis for further optimizing the prediction model and the structural equation model. This method can effectively combine the real-time nature of the prediction model and the causal relationship of the structural equation model to improve the accuracy and interpretability of fault prediction.

[0194] Figure 2 FIG. 1 is a schematic diagram of the structure of the DCS system autonomous controllable upgrade and system fault prediction system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0195] The first unit is used to obtain and preprocess the historical operation data and real-time operation data of the DCS system, extract time series features, frequency domain features and topological features from the preprocessed historical operation data and real-time operation data, perform weighted fusion on the extracted features to obtain fusion features, perform multi-scale fault prediction using the fusion features based on a hybrid algorithm of a long short-term memory network and a graph convolutional neural network, and obtain a fault prediction result, and generate an autonomous and controllable upgrade plan based on the fault prediction result and the current state of the DCS system through a deep reinforcement learning algorithm, wherein the autonomous and controllable upgrade plan includes an upgrade schedule, upgrade content and upgrade steps;

[0196] The second unit is used to perform a consistency backup of the DCS system state at the time specified by the upgrade schedule, and perform software updates, hardware configuration adjustments and system parameter optimization according to the upgrade steps. In the process of executing the autonomous and controllable upgrade plan, the DCS system indicators are analyzed in real time. When an abnormal situation is detected, the DCS system is restored to the state before the upgrade. After the upgrade is completed, the DCS system is functionally tested, performance evaluated and securely audited to obtain the operating state of the upgraded DCS system;

[0197] The third unit is used to obtain multi-source heterogeneous data under the operating state of the upgraded DCS system and perform adaptive sampling, use the sampled data to collaboratively train the fault prediction model to obtain a pre-trained model, migrate the pre-trained model knowledge to the lightweight online model, perform real-time fault prediction based on the lightweight online model, obtain real-time fault prediction results, and simultaneously construct a knowledge graph including historical fault cases, expert experience and equipment manual information, verify and optimize the real-time fault prediction results based on the knowledge graph, and generate an explanatory report including fault prediction and upgrade decision-making process according to the optimized fault prediction results and the autonomous and controllable upgrade plan.

[0198] According to a third aspect of the embodiments of the present invention,

[0199] An electronic device is provided, comprising:

[0200] processor;

[0201] a memory for storing processor-executable instructions;

[0202] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0203] A fourth aspect of the embodiments of the present invention is:

[0204] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0205] 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.

[0206] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for autonomous and controllable DCS system upgrade and system fault prediction, characterized in that: include: The historical operation data and real-time operation data of the DCS system are obtained and preprocessed, and the time series features, frequency domain features and topological features are extracted from the preprocessed historical operation data and real-time operation data, and the extracted features are weightedly fused to obtain fused features. Based on a hybrid algorithm of a long short-term memory network and a graph convolutional neural network, the fused features are used to perform multi-scale fault prediction to obtain fault prediction results. According to the fault prediction results and the current state of the DCS system, an autonomous and controllable upgrade plan is generated through a deep reinforcement learning algorithm, and the autonomous and controllable upgrade plan includes an upgrade schedule, upgrade content and upgrade steps; At the time specified in the upgrade schedule, the DCS system status is backed up consistently, and software updates, hardware configuration adjustments, and system parameter optimizations are performed according to the upgrade steps. In the process of executing the autonomous and controllable upgrade plan, the DCS system indicators are analyzed in real time. When an abnormality is detected, the DCS system is restored to the state before the upgrade. After the upgrade is completed, the DCS system is functionally tested, performance evaluated, and security audited to obtain the operating state of the upgraded DCS system; Acquire multi-source heterogeneous data under the operating state of the upgraded DCS system and perform adaptive sampling, use the sampled data to collaboratively train the fault prediction model to obtain a pre-trained model, migrate the pre-trained model knowledge to a lightweight online model, perform real-time fault prediction based on the lightweight online model to obtain real-time fault prediction results, and simultaneously construct a knowledge graph including historical fault cases, expert experience and equipment manual information, verify and optimize the real-time fault prediction results based on the knowledge graph, and generate an explanatory report including the fault prediction and upgrade decision-making process according to the optimized fault prediction results and the autonomous and controllable upgrade plan.

2. The method according to claim 1, characterized in that Based on the hybrid algorithm of long short-term memory network and graph convolutional neural network, the fusion features are used to perform multi-scale fault prediction to obtain fault prediction results. According to the fault prediction results and the current state of the DCS system, an autonomous and controllable upgrade plan is generated through a deep reinforcement learning algorithm, including: An initial graph is constructed based on the physical topology of the DCS system, and graph structures of different granularities are generated through a graph coarsening algorithm. A graph attention network is used to extract features on the graph structures of different granularities to obtain multi-scale graph features. An improved spatiotemporal attention mechanism is constructed to divide the pre-acquired fusion features into fixed-size blocks, a full attention mechanism is used within the blocks, and sparse connections are used between blocks to obtain spatiotemporal attention features. An adaptive fusion module is designed to dynamically weight the multi-scale graph features and the spatiotemporal attention features using a gating mechanism to obtain the final feature representation. Based on the final feature representation, a multi-task learning model is constructed, including a fault prediction task, a system state prediction task, and an anomaly detection task; Training the multi-task learning model to obtain a trained fault prediction model, and using the trained fault prediction model to perform multi-scale fault prediction on the DCS system to obtain a fault prediction result; The upgrade process of the DCS system is modeled as a Markov decision process, and a state space, an action space and a reward function are defined, wherein the state space includes the version number, configuration parameters, performance indicators and the fault prediction results of each component of the system, and the action space includes executable upgrade operations; Construct a deep reinforcement learning model, perform hierarchical design on the action space of the deep reinforcement learning model, divide the actions into high-level strategies and low-level strategies, wherein the high-level strategy determines the overall direction of the upgrade, and the low-level strategy determines the specific upgrade operation, and at the same time construct a priority experience replay mechanism based on the temporal difference error, calculate the temporal difference error of each experience sample, determine the sampling probability according to the error size, give priority to replaying experience samples with a large temporal difference error, train the environmental dynamics model, predict the short-term impact of the upgrade operation, input the prediction result into the deep reinforcement learning model as additional state information, and add safety constraints in the decision-making process of the deep reinforcement learning model to obtain an improved deep reinforcement learning model; By using the improved deep reinforcement learning model, based on the fault prediction results and the current state of the DCS system, an autonomous and controllable upgrade plan including upgrade schedule, upgrade content and upgrade steps is generated.

3. The method according to claim 2, characterized in that Using the improved deep reinforcement learning model, based on the fault prediction results and the current state of the DCS system, an autonomous and controllable upgrade plan including upgrade schedule, upgrade content and upgrade steps is generated, including: Input the current system status, fault prediction results, and system evolution trajectory into the optimized deep reinforcement learning model, and output high-level strategies through the optimized deep reinforcement learning model to determine the overall direction of the upgrade, including the upgrade path and resource allocation plan; Based on the output high-level strategy, the optimized deep reinforcement learning model further outputs the low-level strategy to determine the specific upgrade operations, including upgrade sequencing and risk avoidance strategies; Combine the output high-level and low-level strategies to form a complete upgrade plan, including upgrade schedule, upgrade content and upgrade steps, and generate component-level and system-level upgrade schedules, version selection matrices and configuration change lists; Use a pre-trained multi-scale environmental dynamics model to simulate and validate the generated upgrade scenarios, and evaluate the short-term and long-term impacts of the current upgrade scenarios at different time scales and system levels; Design a comprehensive evaluation function to evaluate the upgrade solution in multiple dimensions by combining simulation results, preset performance and safety standards; Determine whether the evaluation result meets the requirements. If so, output the final upgrade plan. If not, start the adaptive feedback mechanism, adjust the reward function weight and exploration strategy of the deep reinforcement learning model, and return to the step of inputting the current system status and fault prediction results into the optimized deep reinforcement learning model to regenerate the upgrade plan.

4. The method according to claim 1, characterized in that In the process of executing the autonomous and controllable upgrade solution, the DCS system indicators are analyzed in real time. When an abnormality is detected, the DCS system is restored to the state before the upgrade, including: Build a digital twin model of the DCS system, which includes a physical model and a data-driven model, and synchronize data between the physical system and the digital twin model through edge computing nodes; Based on the digital twin model, a large amount of upgrade scenario data is generated by using the Monte Carlo method, including normal operation, abnormal conditions and extreme conditions. The upgrade scenario data is used to train the upgrade strategy agent to obtain an initial upgrade strategy. The initial upgrade strategy is input into the multi-objective optimization model to construct a multi-objective function including upgrade efficiency, system stability and resource consumption. The adaptive weight adjustment mechanism is used to dynamically adjust the importance of each goal according to the current state of the system. By solving the multi-objective optimization model, the optimized upgrade strategy is finally output. Based on the optimization upgrade strategy, an autonomous and controllable upgrade plan is generated. In the process of executing the autonomous and controllable upgrade plan, a multimodal sensor network and a software-defined network are deployed to collect the operation status data of the DCS system in real time, and the operation status data is processed in real time and adaptively sampled to generate system operation status information. Based on the system operation status information, a multi-level anomaly detection system is constructed to analyze the indicators of the DCS system in real time and output anomaly detection results; When the abnormality detection result indicates that an abnormality exists, a multi-level rollback mechanism is triggered to restore the DCS system to the state before the upgrade. If no abnormality is detected, the autonomous and controllable upgrade plan continues to be executed until completion.

5. The method according to claim 4, characterized in that Based on the system operation status information, a multi-level anomaly detection system is constructed to analyze the indicators of the DCS system in real time, and the output anomaly detection results include: Construct the first-layer anomaly detection model, take the acquired system operation status information as input, calculate the statistical characteristics of system component status and indicators through statistical analysis methods, and obtain the baseline model of normal system operation; Using the constructed first-layer anomaly detection model, the current system operation status information is compared with the benchmark model in real time to calculate the deviation value. When the deviation value exceeds the preset threshold, the second-layer anomaly detection is triggered; Build a second-layer anomaly detection model that takes the status and indicators of system components that trigger anomalies as input, trains a classifier through a machine learning algorithm, and identifies the type and severity of anomalies; Based on the output results of the second-layer anomaly detection model, a third-layer anomaly detection model is constructed. The third-layer anomaly detection model adopts complex network theory to map abnormal components and related indicators into network nodes, establishes connections between nodes according to the probability of abnormal propagation, performs topological analysis on the complex network constructed by the third-layer anomaly detection model, calculates node centrality and community structure, and identifies target abnormal nodes and risk areas; According to the analysis results of the third-layer anomaly detection model, an anomaly propagation prediction model is established to simulate the diffusion process of anomalies in the system, predict other components and indicators that may be affected, and finally output the anomaly detection results.

6. The method according to claim 1, characterized in that Construct a knowledge graph containing historical fault cases, expert experience, and equipment manual information, verify and optimize the real-time fault prediction results based on the knowledge graph, and generate an explanatory report containing the fault prediction and upgrade decision-making process based on the optimized fault prediction results and the autonomous and controllable upgrade plan, including: Obtain historical fault cases, expert experience rules and equipment manual information to build a multimodal knowledge graph, which includes physical topological relationships, equipment operating parameters, fault evolution paths and maintenance decision information; A conditional independence test is performed based on historical failure cases in the multimodal knowledge graph, an initial causal structure between variables is constructed using a kernel method, and physical connection constraints, expert rule constraints, and temporal evolution constraints are extracted from the multimodal knowledge graph, and the constraints are fused with the initial causal structure to generate a causal adjacency matrix; A nonlinear structural equation model is constructed using a neural network, and the parameters of the quantitative relationship between variables are learned based on the causal adjacency matrix, and a regularization term is introduced for optimization to obtain a structural equation model that describes the evolution process of equipment failure. Mapping the obtained real-time fault prediction results to the corresponding nodes of the structural equation model, calculating the intervention effect and obtaining the prediction deviation through counterfactual reasoning; The correction amount is calculated according to the prediction deviation, and the real-time fault prediction result is optimized and adjusted to obtain a corrected fault prediction result. Based on the expert experience rules and equipment manual information in the multimodal knowledge graph, an explanatory report including the fault prediction process and decision-making basis is generated in combination with the corrected fault prediction result.

7. The method according to claim 6, characterized in that The obtained real-time fault prediction results are mapped to the corresponding nodes of the structural equation model, the intervention effect is calculated, and the prediction deviation is obtained through counterfactual reasoning, including: Establishing a correspondence matrix between the prediction output and the structural equation model nodes, wherein the correspondence matrix includes a direct mapping relationship, an indirect mapping relationship, and a non-mapping relationship between the prediction model output vector and the structural equation model node set; According to the corresponding relationship matrix, the output vector mapping of the prediction model is converted into the node value of the structural equation model, and for each target intervention node in the structural equation model, based on the node value, the intervention time window of the target intervention node is determined through time series correlation analysis, and the causal effect value of the target output node corresponding to the target intervention node is calculated using the causal intervention operation rule within the intervention time window; Performing time series accumulation on the causal effect value to obtain a corresponding relationship curve between the dynamic intervention intensity of the target intervention node and the causal effect value, and identifying a key intervention threshold based on the inflection point of the curve; Constructing a multi-level counterfactual reasoning model, the multi-level counterfactual reasoning model includes node-level counterfactual analysis and path-level counterfactual analysis, and inputting the node values ​​into the multi-level counterfactual reasoning model based on the structural equation model; In the node-level counterfactual analysis, the counterfactual prediction value caused by the change of a single node is calculated, and in the path-level counterfactual analysis, the counterfactual prediction value caused by the change of a combination of nodes is calculated; The counterfactual predicted value is compared with the actual fault value and the output vector of the prediction model in multiple dimensions, and the final prediction deviation is obtained by weighting based on the node importance.

8. A DCS system autonomous controllable upgrade and system fault prediction system, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain and preprocess the historical operation data and real-time operation data of the DCS system, extract time series features, frequency domain features and topological features from the preprocessed historical operation data and real-time operation data, perform weighted fusion on the extracted features to obtain fusion features, perform multi-scale fault prediction using the fusion features based on a hybrid algorithm of a long short-term memory network and a graph convolutional neural network, and obtain a fault prediction result, and generate an autonomous and controllable upgrade plan based on the fault prediction result and the current state of the DCS system through a deep reinforcement learning algorithm, wherein the autonomous and controllable upgrade plan includes an upgrade schedule, upgrade content and upgrade steps; The second unit is used to perform a consistency backup of the DCS system state at the time specified by the upgrade schedule, and perform software updates, hardware configuration adjustments and system parameter optimization according to the upgrade steps. In the process of executing the autonomous and controllable upgrade plan, the DCS system indicators are analyzed in real time. When an abnormal situation is detected, the DCS system is restored to the state before the upgrade. After the upgrade is completed, the DCS system is functionally tested, performance evaluated and securely audited to obtain the operating state of the upgraded DCS system; The third unit is used to obtain multi-source heterogeneous data under the operating state of the upgraded DCS system and perform adaptive sampling, use the sampled data to collaboratively train the fault prediction model to obtain a pre-trained model, migrate the pre-trained model knowledge to a lightweight online model, perform real-time fault prediction based on the lightweight online model, and obtain real-time fault prediction results, while constructing a knowledge graph containing historical fault cases, expert experience, and equipment manual information, verifying and optimizing the real-time fault prediction results based on the knowledge graph, and generating an explanatory report containing fault prediction and upgrade decision-making processes according to the optimized fault prediction results and the autonomous and controllable upgrade plan.

9. 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 described in any one of claims 1 to 7.

10. 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 7 is implemented.

Images