Abnormal diagnosis and optimization method and system for petrochemical production process integrating knowledge graph

Abstract

The present invention provides a petrochemical production process abnormality diagnosis and optimization method and system integrating knowledge graph, which relates to the field of petrochemical production technology, including: collecting process parameters and equipment status data, using adaptive wavelet threshold denoising and deep variational mode decomposition for preprocessing and feature extraction respectively, constructing a heterogeneous graph neural network and combining a neural causal discovery network to construct a causal relationship graph, optimizing the causal relationship graph through knowledge distillation and completing missing relationships to obtain a knowledge graph, comparing real-time features with normal feature distributions of a pre-trained contrast diffusion model, and after triggering anomaly detection, using a causal reasoning engine in combination with a knowledge graph to analyze the anomaly propagation path, determining the root cause of the anomaly and predicting the trend based on multi-agent reinforcement learning, constructing a multi-objective optimization problem according to the knowledge graph, generating an optimization strategy and dynamically adjusting it, and evaluating the optimization results in combination with a hybrid model of knowledge distillation, and finally achieving progressive optimization and updating the knowledge graph.

Description

Technical Field

[0001] The present invention relates to the field of petrochemical production technology, and in particular to a petrochemical production process abnormality diagnosis and optimization method and system integrating knowledge graph. Background Art

[0002] The petrochemical production process is a complex system involving numerous process parameters, equipment units and complex physical and chemical reactions. It is crucial to ensure the safe and stable operation of the petrochemical production process and to improve production efficiency and product quality.

[0003] With the rapid development of automation and information technology, a large amount of production process data is collected and stored, which provides a basis for abnormal diagnosis and optimization using data-driven methods. Traditional petrochemical production process abnormality diagnosis methods mainly rely on expert experience and rule-based systems, which are usually difficult to handle complex nonlinear relationships and massive data, and are difficult to adapt to the dynamic changes of the production process.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a petrochemical production process abnormality diagnosis and optimization method and system integrating knowledge graph, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a petrochemical production process abnormality diagnosis and optimization method integrating a knowledge graph, comprising:

[0007] Collect process parameters and equipment status monitoring data in the petrochemical production process, pre-process the process parameters through adaptive wavelet threshold denoising, select the optimal decomposition scale and threshold in combination with the attention mechanism to obtain denoised process parameters, extract multi-scale features of the equipment status monitoring data through deep variational mode decomposition, and adaptively weight them in combination with the multi-head self-attention mechanism to obtain the equipment status feature vector, build a heterogeneous graph neural network based on the denoised process parameters and the equipment status feature vector, extract node representations through graph contrast learning and calculate the causal strength between different nodes in combination with the neural causal discovery network, build a causal relationship graph, convert expert experience into graph structure constraints and node attribute constraints through knowledge distillation method and optimize the causal relationship graph, and complete the missing relationship in combination with graph structure reasoning to obtain the knowledge graph corresponding to the petrochemical production process;

[0008] Collect real-time monitoring data and perform feature extraction to obtain real-time features and add them to the pre-trained contrast diffusion model. Combine the contrast learning algorithm to determine the normal feature distribution corresponding to the normal working condition. Model the deviation between the real-time feature and the normal feature distribution through the graph diffusion network. If the deviation exceeds the preset deviation threshold, trigger anomaly detection, activate the causal reasoning engine according to the node where the anomaly is detected, combine the causal relationship in the knowledge graph and the neural structural equation model to analyze the propagation path of the anomaly in the graph structure, combine the multi-agent reinforcement learning method to determine the root cause of the anomaly, and combine the pre-set hierarchical anomaly prediction model to predict the trend of the current anomaly, and obtain the credibility evaluation corresponding to the current anomaly;

[0009] The constraints and optimization objectives related to the current anomaly are extracted from the knowledge graph to construct a multi-objective optimization problem, an optimization strategy generator based on the meta-learning period is constructed, and the weights of the optimization objectives are adjusted according to the abnormal scenarios, the local optimization strategy and the global optimization strategy are solved respectively by a hybrid optimization method based on deep reinforcement learning and evolutionary algorithm, the initial optimization strategy is obtained by combination and executed, a predictive controller based on Neural Ordinary Differential Equations is constructed, and the optimization process is dynamically constrained and dynamically adjusted, the initial optimization result is obtained and added to the hybrid model based on knowledge distillation, the reliability of the initial optimization result is evaluated according to the pre-acquired mechanism knowledge and historical data, a progressive optimization execution strategy is constructed according to the evaluation results, the effect data when executing the progressive optimization execution strategy is collected and it is determined whether the optimization goal is achieved, and if completed, the knowledge graph is updated according to the progressive optimization execution strategy.

[0010] In an optional embodiment,

[0011] Collect process parameters and equipment status monitoring data in the petrochemical production process, pre-process the process parameters through adaptive wavelet threshold denoising, select the optimal decomposition scale and threshold in combination with the attention mechanism, obtain denoised process parameters, extract multi-scale features of the equipment status monitoring data through deep variational mode decomposition, and perform adaptive weighting in combination with the multi-head self-attention mechanism to obtain the equipment status feature vector including:

[0012] Collecting process parameters and equipment status monitoring data in the petrochemical production process through a distributed data acquisition network, wherein the process parameters include temperature parameters, pressure parameters and flow parameters, and the equipment status monitoring data includes equipment vibration data and corresponding timestamps;

[0013] The process parameters are preprocessed by an adaptive wavelet threshold denoising network based on Transformer. The adaptive wavelet threshold denoising network includes an encoder and a decoder. The encoder analyzes the time-frequency characteristics of the signal through a multi-layer self-attention module. The decoder adaptively determines the soft threshold parameters according to the noise characteristics of different frequency bands. The self-attention module in the encoder retains the signal timing information through position encoding and analyzes the correlation between time domain and frequency domain characteristics. The optimal wavelet decomposition parameters are adaptively selected according to the characteristics of different types of process parameters. The temperature parameter is decomposed by the db4 wavelet basis function, the pressure parameter is decomposed by the sym8 wavelet basis function, and the flow parameter is decomposed by the coif5 wavelet basis function. The denoising thresholds of different frequency bands are adaptively calculated by residual connection and layer normalization algorithm combined with Bayesian optimization. The signal energy, variance and entropy value are used as optimization targets to obtain the denoising process parameters.

[0014] A deep variational mode decomposition network is used to extract features from the equipment status monitoring data. The deep variational mode decomposition network performs preliminary feature extraction through multi-layer one-dimensional convolution, decomposes the equipment vibration data into multiple characteristic mode functions, each characteristic mode function characterizes the equipment characteristics of a specific frequency band, the number of channels of the one-dimensional convolution increases successively, and a multi-head self-attention mechanism is used to perform weighted fusion on the characteristic mode functions. The multi-head self-attention mechanism calculates the attention scores of different frequency bands through dot product operations, adaptively allocates weights according to the importance of frequency band features, and generates equipment status feature vectors.

[0015] In an optional embodiment,

[0016] A heterogeneous graph neural network is constructed based on the denoised process parameters and the equipment status feature vector. Node representations are extracted through graph contrast learning and the causal strength between different nodes is calculated in combination with a neural causal discovery network to construct a causal relationship graph. Expert experience is converted into graph structure constraints and node attribute constraints through a knowledge distillation method and the causal relationship graph is optimized. The missing relationships are completed in combination with graph structure reasoning to obtain a knowledge graph corresponding to the petrochemical production process:

[0017] Constructing a heterogeneous graph neural network based on the denoising process parameters and the equipment state feature vector, wherein the heterogeneous graph neural network includes process parameter nodes and equipment state nodes, and performing feature conversion on different types of nodes through a multi-layer graph attention layer, wherein the feature conversion matrix of the graph attention layer is implemented by a fully connected layer;

[0018] A graph contrast learning method with dynamic negative sampling is used to extract node representation. Positive sample pairs are obtained by sampling from the temporal neighborhood under the same working condition, and negative sample pairs are obtained by random sampling from different working conditions. The node feature similarity is calculated based on cosine similarity, and the node representation is optimized through the contrast loss function.

[0019] The causal relationship between nodes is calculated through a neural causal discovery network based on a dual-stream architecture. The neural causal discovery network includes a temporal feature extraction stream and a causal inference stream. The temporal feature extraction stream uses a deep residual network to extract temporal features, and the causal inference stream uses a graph attention network to perform causal inference. The features of the two streams are fused through a gating mechanism to calculate the causal scores between node pairs and establish a causal relationship graph.

[0020] Through the knowledge distillation method of the teacher-student network structure, the expert experience is converted into graph structure constraints and node attribute constraints, and a graph neural reasoning network is used to complete the missing relationships in the causal graph. The graph neural reasoning network includes a feature propagation layer and a relationship reasoning layer. The feature propagation layer updates the node representation in the causal graph through gated graph convolution, and the relationship reasoning layer calculates the confidence of the edge through the attention mechanism. The reasoning results are filtered according to the pre-set confidence threshold to obtain the knowledge graph corresponding to the petrochemical production process.

[0021] In an optional embodiment,

[0022] Collect real-time monitoring data and perform feature extraction to obtain real-time features and add them to the pre-trained contrast diffusion model. Combine the contrast learning algorithm to determine the normal feature distribution corresponding to the normal working condition. Model the deviation between the real-time feature and the normal feature distribution through the graph diffusion network. If the deviation exceeds the preset deviation threshold, trigger anomaly detection. Activate the causal reasoning engine according to the node where the anomaly is detected. Combine the causal relationship in the knowledge graph and the neural structural equation model to analyze the propagation path of the anomaly in the graph structure. Combine the multi-agent reinforcement learning method to determine the root cause of the anomaly and combine the preset hierarchical anomaly prediction model to predict the trend of the current anomaly. Obtain the corresponding credibility assessment of the current anomaly, including:

[0023] Collecting real-time monitoring data, the real-time monitoring data includes real-time process parameters and real-time equipment status monitoring data, the real-time process parameters extract time domain features through a deep temporal convolutional network, the deep temporal convolutional network consists of an entry control layer, a feature extraction layer and a fusion layer, the entry control layer uses a gated linear unit to select effective information flow, the feature extraction layer uses a multi-scale hole convolution to extract features, and the fusion layer integrates features based on an attention mechanism, the real-time equipment status monitoring data extracts frequency domain features through a wavelet packet decomposition algorithm, and adaptively selects the optimal basis function to obtain real-time features;

[0024] The real-time features are input into a pre-trained contrast diffusion model, wherein the contrast diffusion model is constructed by a multi-stage pre-training method. In the first stage, unlabeled data is used for self-supervised training, a feature encoder adopts a residual network structure, and a contrast predictor implements feature projection based on a multi-layer perceptron. In the second stage, labeled data is introduced for supervised fine-tuning, and a normal feature distribution corresponding to a normal operating condition is determined in combination with a contrast learning algorithm.

[0025] Modeling the deviation between the real-time feature and the normal feature distribution based on a heterogeneous graph diffusion network, wherein the heterogeneous graph diffusion network includes process parameter nodes and equipment status nodes, the connection relationship between the nodes is determined according to the physical connection relationship and the process flow, the diffusion process is controlled by a heat kernel matrix, and feature propagation is performed based on multi-layer diffusion convolution and attention mechanism, and anomaly detection is triggered when the node deviation exceeds a preset deviation threshold;

[0026] The causal inference engine is activated according to the detected abnormal nodes, and a causal probability graph is constructed based on the conditional random field model. The nodes represent variables and the edges represent conditional dependencies. The variational inference algorithm is used to calculate the edge probability distribution, and the neural structural equation model is combined to fit the nonlinear relationship between variables and analyze the propagation path of the anomaly in the graph structure.

[0027] A hierarchical multi-agent reinforcement learning system is used to determine the root cause of the anomaly. The hierarchical multi-agent reinforcement learning system includes a strategy agent, an expert agent, and an execution agent. The strategy agent is responsible for global task allocation, the expert agent handles specific types of anomalies, and the execution agent completes specific diagnosis. The agents transmit messages and make collaborative decisions through a graph attention network.

[0028] Based on the hierarchical anomaly prediction model, the trend of the current anomaly is predicted. The hierarchical anomaly prediction model includes a short-term prediction module, a medium-term prediction module and a long-term prediction module. The short-term prediction module uses a bidirectional long short-term memory network to model local features, the medium-term prediction module uses a causal convolutional network to expand the receptive field, and the long-term prediction module uses a multi-head self-attention mechanism to capture long-range dependencies to obtain anomaly diagnosis results;

[0029] A credibility evaluation system is used to evaluate the abnormal diagnosis results. The credibility evaluation system includes an abnormal scoring module, a confidence calculation module and a decision optimization module. The abnormal scoring module integrates multiple basic detectors for multi-dimensional evaluation. The confidence calculation module fuses multiple evaluation results based on an integrated learning framework. The decision optimization module optimizes the evaluation results using a Bayesian decision method to obtain a credibility evaluation result corresponding to the current abnormality.

[0030] In an optional embodiment,

[0031] Based on the hierarchical anomaly prediction model, the trend of the current anomaly is predicted. The hierarchical anomaly prediction model includes a short-term prediction module, a medium-term prediction module and a long-term prediction module. The short-term prediction module uses a bidirectional long short-term memory network to model local features, the medium-term prediction module uses a causal convolutional network to expand the receptive field, and the long-term prediction module uses a multi-head self-attention mechanism to capture long-range dependencies. The abnormal diagnosis results include:

[0032] Obtain an abnormal state data sequence corresponding to the current abnormality, perform adaptive segmentation based on the data change rate, and dynamically window divide the abnormal state data sequence to obtain a segmented data sequence;

[0033] Inputting the segmented data sequence into an enhanced bidirectional long short-term memory network for short-term prediction, wherein the enhanced bidirectional long short-term memory network comprises four bidirectional long short-term memory layers, each layer comprises a residual connection mechanism and an adaptive Dropout mechanism, and weights are assigned to features of different time steps through a temporal attention mechanism to obtain a short-term prediction result;

[0034] Input the segmented data sequence into a multi-branch causal convolutional network for medium-term prediction, wherein the multi-branch causal convolutional network comprises a time domain branch, a frequency domain branch and a trend branch, wherein the time domain branch extracts time series features through progressive dilated convolution, the frequency domain branch extracts spectrum features through wavelet transform, and the trend branch extracts main trend features through an adaptive smoothing filter, and the time series features, spectrum features and main trend features are dynamically fused through an attention mechanism to obtain a medium-term prediction result;

[0035] Input the segmented data sequence, the short-term prediction result and the mid-term prediction result into the improved Transformer network for long-term prediction, store historical abnormal pattern information through multi-scale receptive field mechanism and learning position embedding, combine memory enhancement module and time-aware mask mechanism, and fuse the historical abnormal pattern information through multi-head attention mechanism to obtain long-term prediction result;

[0036] A dynamic confidence assessment network is established, and basic confidence is assigned to the short-term prediction results, the medium-term prediction results, and the long-term prediction results based on the current abnormal characteristics and the historical prediction accuracy. The prediction results are fused through a multi-layer attention network to obtain a fused prediction result. The uncertainty of the fused prediction result is estimated according to the Bayesian neural network, and a probability distribution model of the prediction result is established to obtain the abnormal diagnosis result.

[0037] In an optional embodiment,

[0038] Extracting constraints and optimization objectives related to the current anomaly from the knowledge graph to construct a multi-objective optimization problem, constructing an optimization strategy generator based on the meta-learning period and adjusting the weight of the optimization objective according to the abnormal scenario, solving the local optimization strategy and the global optimization strategy respectively through a hybrid optimization method based on deep reinforcement learning and evolutionary algorithm, combining and executing the initial optimization strategy, constructing a predictive controller based on Neural Ordinary Differential Equations and dynamically constraining and dynamically adjusting the optimization process, obtaining the initial optimization result and adding it to the hybrid model based on knowledge distillation, performing reliability evaluation on the initial optimization result according to the pre-acquired mechanism knowledge and historical data, constructing a progressive optimization execution strategy according to the evaluation result, collecting effect data when executing the progressive optimization execution strategy and judging whether the optimization objective is achieved, and if achieved, updating the knowledge graph according to the progressive optimization execution strategy, including:

[0039] Extract the constraints and optimization goals related to the current anomaly from the knowledge graph, and obtain multi-channel features through the feature extraction layer, wherein the feature extraction layer includes a process parameter channel, an equipment status channel, and an operating condition channel, and integrate the features extracted from each channel to obtain multi-channel feature data;

[0040] A meta-learning optimization strategy generator is constructed based on the multi-channel feature data, wherein the meta-learning optimization strategy generator includes a strategy generation layer, wherein the strategy generation layer sets an input gate and a forget gate, wherein the input gate controls the feature inflow intensity, and the forget gate manages the retention ratio of historical information, and dynamically allocates optimization target weights according to abnormal scene features through a soft attention mechanism;

[0041] A hybrid optimization method is used to solve the optimization strategy. The hybrid optimization method includes local optimization and global optimization. The local optimization adopts a hierarchical reinforcement learning structure, including a high-level strategy network and a dual value evaluation mechanism. The high-level strategy network includes four fully connected layers and uses a LeakyReLU activation function. The dual value evaluation mechanism includes an immediate value network and a long-term value network. The immediate value network evaluates the single-step optimization effect. The long-term value network uses a long-short-term memory network to predict long-term benefits. The optimized samples are stored in a hierarchical experience pool according to priority to obtain a local optimization strategy.

[0042] The global optimization adopts a multi-scale evolutionary framework, divides the population into multiple sub-populations, performs population mutation through Gaussian mutation operator and Cauchy mutation operator, wherein the mutation intensity is negatively correlated with the individual fitness, adopts adaptive factorial crossover to perform population crossover, determines the crossover site through a probabilistic graphical model, maintains population diversity based on crowding sorting, obtains a global optimization strategy and combines it with the local optimization strategy, obtains the initial optimization strategy and executes it;

[0043] In the process of executing the initial optimization strategy, a prediction controller based on a neural ordinary differential equation is constructed and a discrete observation sequence is obtained. The prediction controller includes a state encoding network, a dynamic approximation network and a decoding network. The state encoding network maps the discrete observation sequence into a continuous state vector. The control sequence is restored through the residual block structure in the dynamic approximation network and the deconvolution operation in the decoding network. The process limit is converted into a penalty term in combination with a constraint processing module to obtain an initial optimization result.

[0044] Input the initial optimization result into a hybrid model based on knowledge distillation, wherein the hybrid model includes a teacher network and a student network, wherein the teacher network integrates a process mechanism model, and the student network adopts a deep residual network structure, and performs knowledge migration through an attention migration mechanism, and constructs a three-layer evaluation system to perform reliability evaluation on the optimization scheme, wherein the three-layer evaluation system includes a static evaluation layer, a dynamic evaluation layer, and a comprehensive evaluation layer, wherein the static evaluation layer evaluates the feasibility of the initial optimization strategy based on fuzzy rule reasoning, the dynamic evaluation layer performs random perturbation simulation through a Monte Carlo method, and the comprehensive evaluation layer uses a deep belief network to output a reliability score and outputs the reliability score as an evaluation result;

[0045] A progressive optimization execution strategy is constructed based on the evaluation results, execution effect data is collected and it is determined whether the optimization goal is achieved. When the optimization goal is achieved, the progressive optimization execution strategy is updated to the knowledge graph.

[0046] In an optional embodiment,

[0047] The global optimization adopts a multi-scale evolutionary framework, divides the population into multiple sub-populations, and performs population mutation through Gaussian mutation operator and Cauchy mutation operator, wherein the mutation intensity is negatively correlated with the individual fitness, adopts adaptive factorial crossover to perform population crossover, determines the crossover site through a probabilistic graph model, maintains population diversity based on crowding sorting, and obtains the global optimization strategy including:

[0048] Divide the process parameter space to obtain multiple search areas of different scales, establish the corresponding relationship between each search area and a sub-population, allocate the search space to each sub-population according to the value range and accuracy requirements of the process parameters, and construct the initial population;

[0049] Performing a mutation operation on each individual in the initial population to obtain a mutant individual, using a Gaussian mutation operator to perform a local fine search on the individual, using a Cauchy mutation operator to perform a global wide-area search on the individual, detecting the relationship between the fitness value of the individual and the average fitness value of the initial population, and when it is detected that the fitness value of the individual is lower than the average fitness value of the initial population, increasing the mutation strength to a preset first mutation strength, and when it is detected that the fitness value of the individual is higher than the average fitness value of the initial population, reducing the mutation strength to a preset second mutation strength;

[0050] Constructing a parameter correlation network and calculating the correlation strength between process parameters, identifying parameter pairs whose correlation strength is greater than a preset threshold to form a strongly correlated parameter group, performing an adaptive factorial crossover operation on the parameters in the strongly correlated parameter group, detecting the fitness value of the parent individual and adjusting the crossover probability according to the detection result;

[0051] Count the evolutionary generations and trigger population migration when the preset cycle is reached, calculate the fitness values ​​of all individuals in the population and sort them, randomly select individuals from the top ten percent of the individuals in fitness as high-quality individuals and migrate the high-quality individuals to the sub-population of the adjacent scale, replace the individuals with the lowest fitness in the target sub-population with the high-quality individuals, count the number of migrations of the high-quality individuals, and remove the high-quality individuals whose migration times exceed the preset number;

[0052] Calculate the distance between each individual in the population and other individuals in the parameter space and the target space, perform normalization and weighted summation to obtain the crowding degree of each individual, select and retain individuals based on the crowding degree, detect the population diversity index, and increase the mutation intensity and generate random individuals when the population diversity index is detected to be lower than a preset threshold;

[0053] The mutation operation, crossover operation, population migration and diversity maintenance are performed cyclically until the preset number of iterations is reached to obtain the global optimization strategy.

[0054] A second aspect of an embodiment of the present invention provides a petrochemical production process abnormality diagnosis and optimization system integrating a knowledge graph, comprising:

[0055] The first unit is used to collect process parameters and equipment status monitoring data in the petrochemical production process, pre-process the process parameters through adaptive wavelet threshold denoising, select the optimal decomposition scale and threshold in combination with the attention mechanism to obtain denoised process parameters, extract multi-scale features of the equipment status monitoring data through deep variational mode decomposition, and adaptively weight them in combination with the multi-head self-attention mechanism to obtain equipment status feature vectors, construct a heterogeneous graph neural network based on the denoised process parameters and the equipment status feature vectors, extract node representations through graph contrast learning and calculate the causal strength between different nodes in combination with the neural causal discovery network, construct a causal relationship graph, convert expert experience into graph structure constraints and node attribute constraints through the knowledge distillation method and optimize the causal relationship graph, and complete the missing relationships in combination with graph structure reasoning to obtain the knowledge graph corresponding to the petrochemical production process;

[0056] The second unit is used to collect real-time monitoring data and perform feature extraction to obtain real-time features and add them to the pre-trained contrast diffusion model, determine the normal feature distribution corresponding to the normal working condition in combination with the contrast learning algorithm, and model the deviation between the real-time features and the normal feature distribution through the graph diffusion network. If the deviation exceeds the preset deviation threshold, anomaly detection is triggered, and the causal reasoning engine is activated according to the node where the abnormality is detected. The propagation path of the anomaly in the graph structure is analyzed in combination with the causal relationship in the knowledge graph and the neural structural equation model, and the root cause of the anomaly is determined in combination with the multi-agent reinforcement learning method, and the trend of the current anomaly is predicted in combination with the pre-set hierarchical anomaly prediction model, so as to obtain the credibility evaluation corresponding to the current anomaly;

[0057] The third unit is used to extract constraints and optimization objectives related to the current anomaly from the knowledge graph to construct a multi-objective optimization problem, construct an optimization strategy generator based on the meta-learning period and adjust the weights of the optimization objectives according to the abnormal scenario, solve the local optimization strategy and the global optimization strategy respectively through a hybrid optimization method based on deep reinforcement learning and evolutionary algorithm, combine and execute the initial optimization strategy, construct a predictive controller based on Neural Ordinary Differential Equations and dynamically constrain and dynamically adjust the optimization process, obtain the initial optimization result and add it to the hybrid model based on knowledge distillation, perform reliability evaluation on the initial optimization result according to the pre-acquired mechanism knowledge and historical data, construct a progressive optimization execution strategy according to the evaluation results, collect effect data when executing the progressive optimization execution strategy and judge whether the optimization goal is achieved, and if completed, update the knowledge graph according to the progressive optimization execution strategy.

[0058] According to a third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] a memory for storing processor-executable instructions;

[0062] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0063] According to a fourth aspect of the embodiments of the present invention,

[0064] 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.

[0065] In the present invention, advanced technologies such as deep learning, graph neural networks and causal reasoning are combined, which can more accurately identify the root causes and propagation paths of anomalies, avoid misjudgments and missed judgments, and improve the efficiency of anomaly diagnosis. By comparing diffusion models and graph diffusion networks, abnormal trends can be predicted, and credibility assessment can be performed in combination with knowledge graphs to provide operators with more reliable decision-making basis and take preventive measures in advance. The hybrid optimization method based on meta-learning, deep reinforcement learning and evolutionary algorithms can dynamically optimize different abnormal scenarios, and dynamically constrain and adjust through neural ordinary differential equations, ultimately achieving optimization of the production process, and updating the knowledge graph according to the optimization results to achieve knowledge accumulation and self-learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A schematic diagram of a process flow of a petrochemical production process abnormality diagnosis and optimization method integrating a knowledge graph according to an embodiment of the present invention;

[0067] Figure 2 This is a schematic diagram of the structure of a petrochemical production process abnormality diagnosis and optimization system that integrates a knowledge graph according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] 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.

[0069] 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.

[0070] Figure 1The flowchart of the petrochemical production process abnormality diagnosis and optimization method integrating knowledge graph according to the embodiment of the present invention is as follows: Figure 1 As shown, the method includes:

[0071] S1. Collect process parameters and equipment status monitoring data in the petrochemical production process, pre-process the process parameters through adaptive wavelet threshold denoising, select the optimal decomposition scale and threshold in combination with the attention mechanism to obtain denoised process parameters, perform multi-scale feature extraction on the equipment status monitoring data through deep variational mode decomposition, and adaptively weight them in combination with the multi-head self-attention mechanism to obtain the equipment status feature vector, construct a heterogeneous graph neural network based on the denoised process parameters and the equipment status feature vector, extract node representations through graph contrast learning and calculate the causal strength between different nodes in combination with the neural causal discovery network, construct a causal relationship graph, convert expert experience into graph structure constraints and node attribute constraints through knowledge distillation method and optimize the causal relationship graph, and complete the missing relationships in combination with graph structure reasoning to obtain the knowledge graph corresponding to the petrochemical production process.

[0072] The deep variational mode decomposition is a signal processing method that combines deep learning and variational inference, and is used to decompose complex time series data into multiple potential modes, thereby revealing the intrinsic structure of the data. The graph contrastive learning is a contrastive learning method based on graph data. By learning the feature representations of nodes and edges in the graph, graphs or nodes with similar structures are made closer in the embedding space. The neural causal discovery network is a deep learning framework that is specifically used to automatically discover causal relationships from data. The causal strength refers to the degree of influence of a causal factor on the outcome variable in the causal relationship. The causal relationship graph is a tool for graphically representing causal relationships, which represents variables and their causal dependencies through nodes and directed edges. The graph structure reasoning refers to performing reasoning and prediction tasks based on a given graph structure.

[0073] In an optional embodiment,

[0074] Collect process parameters and equipment status monitoring data in the petrochemical production process, pre-process the process parameters through adaptive wavelet threshold denoising, select the optimal decomposition scale and threshold in combination with the attention mechanism, obtain denoised process parameters, extract multi-scale features of the equipment status monitoring data through deep variational mode decomposition, and perform adaptive weighting in combination with the multi-head self-attention mechanism to obtain the equipment status feature vector including:

[0075] Collecting process parameters and equipment status monitoring data in the petrochemical production process through a distributed data acquisition network, wherein the process parameters include temperature parameters, pressure parameters and flow parameters, and the equipment status monitoring data includes equipment vibration data and corresponding timestamps;

[0076] The process parameters are preprocessed by an adaptive wavelet threshold denoising network based on Transformer. The adaptive wavelet threshold denoising network includes an encoder and a decoder. The encoder analyzes the time-frequency characteristics of the signal through a multi-layer self-attention module. The decoder adaptively determines the soft threshold parameters according to the noise characteristics of different frequency bands. The self-attention module in the encoder retains the signal timing information through position encoding and analyzes the correlation between time domain and frequency domain characteristics. The optimal wavelet decomposition parameters are adaptively selected according to the characteristics of different types of process parameters. The temperature parameter is decomposed by the db4 wavelet basis function, the pressure parameter is decomposed by the sym8 wavelet basis function, and the flow parameter is decomposed by the coif5 wavelet basis function. The denoising thresholds of different frequency bands are adaptively calculated by residual connection and layer normalization algorithm combined with Bayesian optimization. The signal energy, variance and entropy value are used as optimization targets to obtain the denoising process parameters.

[0077] A deep variational mode decomposition network is used to extract features from the equipment status monitoring data. The deep variational mode decomposition network performs preliminary feature extraction through multi-layer one-dimensional convolution, decomposes the equipment vibration data into multiple characteristic mode functions, each characteristic mode function characterizes the equipment characteristics of a specific frequency band, the number of channels of the one-dimensional convolution increases successively, and a multi-head self-attention mechanism is used to perform weighted fusion on the characteristic mode functions. The multi-head self-attention mechanism calculates the attention scores of different frequency bands through dot product operations, adaptively allocates weights according to the importance of frequency band features, and generates equipment status feature vectors.

[0078] The one-dimensional convolution is a convolution operation in a convolutional neural network, which is specifically used to process one-dimensional data (such as time series or audio signals). The characteristic mode function is a function form used to represent the characteristics of different data modes (such as images, texts, time series, etc.).

[0079] Deploy a distributed data acquisition network to collect process parameters and equipment status monitoring data in the petrochemical production process in real time. Process parameters mainly include temperature, pressure and flow, and equipment status monitoring data mainly includes equipment vibration data and corresponding timestamps. For example, in the ethylene production unit of a petrochemical plant, the distributed data acquisition network collects temperature, pressure and flow data once per second, and equipment vibration data and corresponding timestamps once every millisecond.

[0080] The collected process parameters are preprocessed to remove noise interference. The Transformer-based adaptive wavelet threshold denoising network is used for denoising. The network consists of two parts: an encoder and a decoder. The encoder uses a multi-layer self-attention module to analyze the time-frequency characteristics of the signal. The self-attention module retains the signal timing information through position encoding and analyzes the correlation between time domain and frequency domain features. The decoder adaptively determines the soft threshold parameters according to the noise characteristics of different frequency bands. Different optimal wavelet decomposition parameters are selected for different types of process parameters. For example, the temperature parameter is decomposed using the db4 wavelet basis function, the pressure parameter is decomposed using the sym8 wavelet basis function, and the flow parameter is decomposed using the coif5 wavelet basis function. The denoising thresholds of different frequency bands are adaptively calculated by combining the residual connection and layer normalization algorithm with Bayesian optimization. The signal energy, variance and entropy value are used as optimization targets to obtain the denoised process parameters. For example, the temperature data collected by a temperature sensor is denoised. Assuming the original data is [100, 101, 102, 98, 103, 105, 104, 106], the data obtained after denoising is [100.2, 101.1, 101.9, 99.8, 103.1, 104.8, 104.2, 105.9].

[0081] Feature extraction is performed on the equipment status monitoring data. A deep variational mode decomposition network is used for feature extraction. The network performs preliminary feature extraction through multiple layers of one-dimensional convolution, decomposing the equipment vibration data into multiple characteristic mode functions. Each characteristic mode function characterizes the equipment characteristics of a specific frequency band. The number of channels of the one-dimensional convolution increases successively. For example, the number of channels of the first layer of convolution is 16, the number of channels of the second layer of convolution is 32, and so on. A multi-head self-attention mechanism is used to perform weighted fusion of the characteristic mode functions. The multi-head self-attention mechanism calculates the attention scores of different frequency bands through dot product operations, adaptively assigns weights according to the importance of the frequency band features, and generates the equipment status feature vector. For example, when extracting features from the vibration data of a certain device, assuming that there are three extracted characteristic modal functions, namely [0.1, 0.2, 0.3], [0.4, 0.5, 0.6] and [0.7, 0.8, 0.9], after weighted fusion of the multi-head self-attention mechanism, the generated device state feature vector is [0.35, 0.45, 0.55].

[0082] In this embodiment, the process parameters are preprocessed through an adaptive wavelet threshold denoising network, which effectively removes noise interference, improves the accuracy of the process parameters, and provides a reliable data basis for subsequent control and optimization. The equipment status monitoring data is feature extracted through a deep variational mode decomposition network, which can extract more representative equipment status features, improve the accuracy of equipment status monitoring, and help to timely detect equipment abnormalities and avoid accidents. By accurately monitoring and analyzing process parameters and equipment status, the petrochemical production process can be better controlled and optimized, the stability and efficiency of the production process can be enhanced, energy consumption and material consumption can be reduced, and economic benefits can be improved.

[0083] In an optional embodiment,

[0084] A heterogeneous graph neural network is constructed based on the denoised process parameters and the equipment status feature vector. Node representations are extracted through graph contrast learning and the causal strength between different nodes is calculated in combination with a neural causal discovery network to construct a causal relationship graph. Expert experience is converted into graph structure constraints and node attribute constraints through a knowledge distillation method and the causal relationship graph is optimized. The missing relationships are completed in combination with graph structure reasoning to obtain a knowledge graph corresponding to the petrochemical production process:

[0085] Constructing a heterogeneous graph neural network based on the denoising process parameters and the equipment state feature vector, wherein the heterogeneous graph neural network includes process parameter nodes and equipment state nodes, and performing feature conversion on different types of nodes through a multi-layer graph attention layer, wherein the feature conversion matrix of the graph attention layer is implemented by a fully connected layer;

[0086] A graph contrast learning method with dynamic negative sampling is used to extract node representation. Positive sample pairs are obtained by sampling from the temporal neighborhood under the same working condition, and negative sample pairs are obtained by random sampling from different working conditions. The node feature similarity is calculated based on cosine similarity, and the node representation is optimized through the contrast loss function.

[0087] The causal relationship between nodes is calculated through a neural causal discovery network based on a dual-stream architecture. The neural causal discovery network includes a temporal feature extraction stream and a causal inference stream. The temporal feature extraction stream uses a deep residual network to extract temporal features, and the causal inference stream uses a graph attention network to perform causal inference. The features of the two streams are fused through a gating mechanism to calculate the causal scores between node pairs and establish a causal relationship graph.

[0088] Through the knowledge distillation method of the teacher-student network structure, the expert experience is converted into graph structure constraints and node attribute constraints, and a graph neural reasoning network is used to complete the missing relationships in the causal graph. The graph neural reasoning network includes a feature propagation layer and a relationship reasoning layer. The feature propagation layer updates the node representation in the causal graph through gated graph convolution, and the relationship reasoning layer calculates the confidence of the edge through the attention mechanism. The reasoning results are filtered according to the pre-set confidence threshold to obtain the knowledge graph corresponding to the petrochemical production process.

[0089] The dynamic negative sampling graph contrast learning method introduces a dynamic negative sample generation mechanism in graph contrast learning to enhance the learning effect of the model. The time series feature extraction flow is a network module specifically used to extract useful features from time series data. The causal inference flow refers to the part used to process and infer causal relationships in neural networks or other machine learning models.

[0090] Collect process parameters and equipment status data of petrochemical production process. Process parameters include temperature, pressure, flow, etc., and equipment status includes equipment operation status, fault information, etc. For example, collect process parameter data of ethylene production unit of a chemical plant, including cracking furnace outlet temperature, separation tower pressure, ethylene product flow, etc., as well as equipment status data, such as cracking furnace operation status, compressor vibration data, etc. Preprocess the collected data, including data cleaning, denoising, normalization and other operations. For example, use mean filtering method to remove noise in process parameters, and use minimum-maximum normalization method to normalize data to between 0-1.

[0091] A heterogeneous graph neural network is constructed, and the preprocessed process parameters and equipment status feature vectors are used as node attributes to construct a heterogeneous graph containing process parameter nodes and equipment status nodes. Nodes of different types are connected by edges, and the type of edge represents the relationship between nodes. For example, there is an edge between the cracking furnace outlet temperature node and the cracking furnace operation status node, indicating that there is an association relationship between them. The heterogeneous graph neural network uses a multi-layer graph attention layer to perform feature conversion on nodes of different types. The feature conversion matrix of the graph attention layer is implemented by the fully connected layer. By learning the features of the node and its neighboring nodes, a more expressive node representation is obtained. For example, through the graph attention layer, the association between the cracking furnace outlet temperature and the cracking furnace operation status can be learned, and this relationship can be reflected in the node representation.

[0092] A graph contrast learning method with dynamic negative sampling is used to extract node representation. Positive sample pairs are obtained by sampling from the time series neighborhood under the same operating condition. For example, the cracking furnace outlet temperature nodes at adjacent moments under the same operating condition are used as positive sample pairs. Negative sample pairs are obtained by random sampling from different operating conditions. For example, the cracking furnace outlet temperature nodes under different operating conditions are used as negative sample pairs. The node feature similarity is calculated based on cosine similarity, and the node representation is optimized by contrast loss function, so that the similarity between positive sample pairs is as high as possible, and the similarity between negative sample pairs is as low as possible.

[0093] The causal relationship between nodes is calculated through a neural causal discovery network based on a dual-stream architecture. The neural causal discovery network includes a temporal feature extraction stream and a causal inference stream. The temporal feature extraction stream uses a deep residual network to extract the temporal features of the nodes. The causal inference stream uses a graph attention network for causal inference. For example, the temporal features of the cracking furnace outlet temperature are extracted through a deep residual network, and the causal relationship between the cracking furnace outlet temperature and the separation tower pressure is inferred through a graph attention network. The features of the two streams are fused through a gating mechanism to calculate the causal scores between node pairs and establish a causal relationship graph.

[0094] The expert experience is converted into graph structure constraints and node attribute constraints through the knowledge distillation method of the teacher-student network structure. The teacher network is a knowledge graph constructed by expert experience, and the student network is a causal relationship graph constructed by a neural causal discovery network. Through knowledge distillation, the knowledge of the teacher network is transferred to the student network, and the structure and node attributes of the student network are optimized. For example, expert experience shows that the cracking furnace outlet temperature has an impact on the ethylene product flow rate. Through knowledge distillation, this experience can be converted into edge constraints in the causal relationship graph, strengthening the connection between the cracking furnace outlet temperature node and the ethylene product flow node.

[0095] A graph neural inference network is used to complete the missing relationships in the causal graph. The graph neural inference network consists of a feature propagation layer and a relational inference layer. The feature propagation layer updates the node representation in the causal graph through gated graph convolution. The relational inference layer calculates the confidence of the edge through the attention mechanism, filters the inference results according to the pre-set confidence threshold, and obtains the knowledge graph corresponding to the petrochemical production process. For example, through the graph neural inference network, the potential relationship between the operating status of the cracking furnace and the vibration data of the compressor can be inferred and added to the knowledge graph.

[0096] In this embodiment, knowledge is automatically extracted from data, avoiding the tedious process of manually constructing a knowledge graph, saving a lot of time and labor costs. By combining expert experience and data-driven methods, the constructed knowledge graph is more accurate and complete, and can better reflect the actual situation of the petrochemical production process. Through the mining of causal relationships, the causal relationship between various factors in the petrochemical production process is revealed, making the knowledge graph more interpretable and helpful for understanding and optimizing the petrochemical production process.

[0097] S2. Collect real-time monitoring data and perform feature extraction to obtain real-time features and add them to the pre-trained contrast diffusion model. Combine the contrast learning algorithm to determine the normal feature distribution corresponding to the normal working condition. Model the deviation between the real-time feature and the normal feature distribution through the graph diffusion network. If the deviation exceeds the preset deviation threshold, trigger anomaly detection, activate the causal reasoning engine according to the node where the abnormality is detected, combine the causal relationship in the knowledge graph and the neural structural equation model to analyze the propagation path of the anomaly in the graph structure, combine the multi-agent reinforcement learning method to determine the root cause of the anomaly, and combine the pre-set hierarchical anomaly prediction model to predict the trend of the current anomaly, and obtain the credibility assessment corresponding to the current anomaly.

[0098] The contrastive diffusion model is a generative model based on contrastive learning, which is specifically used to learn the latent spatial structure of images, time series or other data. The contrastive learning algorithm is an unsupervised learning method that optimizes the model by learning the similarities and differences between data samples. The graph diffusion network is a neural network for graph data, which is specially designed to capture the local and global structures of nodes in the graph. The causal inference engine is a computational framework based on causal inference theory, which is used to infer and understand the causal relationship between variables. The trend prediction is a time series analysis method that aims to predict future trends based on historical data.

[0099] In an optional embodiment,

[0100] Collect real-time monitoring data and perform feature extraction to obtain real-time features and add them to the pre-trained contrast diffusion model. Combine the contrast learning algorithm to determine the normal feature distribution corresponding to the normal working condition. Model the deviation between the real-time feature and the normal feature distribution through the graph diffusion network. If the deviation exceeds the preset deviation threshold, trigger anomaly detection. Activate the causal reasoning engine according to the node where the anomaly is detected. Combine the causal relationship in the knowledge graph and the neural structural equation model to analyze the propagation path of the anomaly in the graph structure. Combine the multi-agent reinforcement learning method to determine the root cause of the anomaly and combine the preset hierarchical anomaly prediction model to predict the trend of the current anomaly. Obtain the corresponding credibility assessment of the current anomaly, including:

[0101] Collecting real-time monitoring data, the real-time monitoring data includes real-time process parameters and real-time equipment status monitoring data, the real-time process parameters extract time domain features through a deep temporal convolutional network, the deep temporal convolutional network consists of an entry control layer, a feature extraction layer and a fusion layer, the entry control layer uses a gated linear unit to select effective information flow, the feature extraction layer uses a multi-scale hole convolution to extract features, and the fusion layer integrates features based on an attention mechanism, the real-time equipment status monitoring data extracts frequency domain features through a wavelet packet decomposition algorithm, and adaptively selects the optimal basis function to obtain real-time features;

[0102] The real-time features are input into a pre-trained contrast diffusion model, wherein the contrast diffusion model is constructed by a multi-stage pre-training method. In the first stage, unlabeled data is used for self-supervised training, a feature encoder adopts a residual network structure, and a contrast predictor implements feature projection based on a multi-layer perceptron. In the second stage, labeled data is introduced for supervised fine-tuning, and a normal feature distribution corresponding to a normal operating condition is determined in combination with a contrast learning algorithm.

[0103] Modeling the deviation between the real-time feature and the normal feature distribution based on a heterogeneous graph diffusion network, wherein the heterogeneous graph diffusion network includes process parameter nodes and equipment status nodes, the connection relationship between the nodes is determined according to the physical connection relationship and the process flow, the diffusion process is controlled by a heat kernel matrix, and feature propagation is performed based on multi-layer diffusion convolution and attention mechanism, and anomaly detection is triggered when the node deviation exceeds a preset deviation threshold;

[0104] The causal inference engine is activated according to the detected abnormal nodes, and a causal probability graph is constructed based on the conditional random field model. The nodes represent variables and the edges represent conditional dependencies. The variational inference algorithm is used to calculate the edge probability distribution, and the neural structural equation model is combined to fit the nonlinear relationship between variables and analyze the propagation path of the anomaly in the graph structure.

[0105] A hierarchical multi-agent reinforcement learning system is used to determine the root cause of the anomaly. The hierarchical multi-agent reinforcement learning system includes a strategy agent, an expert agent, and an execution agent. The strategy agent is responsible for global task allocation, the expert agent handles specific types of anomalies, and the execution agent completes specific diagnosis. The agents transmit messages and make collaborative decisions through a graph attention network.

[0106] Based on the hierarchical anomaly prediction model, the trend of the current anomaly is predicted. The hierarchical anomaly prediction model includes a short-term prediction module, a medium-term prediction module and a long-term prediction module. The short-term prediction module uses a bidirectional long short-term memory network to model local features, the medium-term prediction module uses a causal convolutional network to expand the receptive field, and the long-term prediction module uses a multi-head self-attention mechanism to capture long-range dependencies to obtain anomaly diagnosis results;

[0107] A credibility evaluation system is used to evaluate the abnormal diagnosis results. The credibility evaluation system includes an abnormal scoring module, a confidence calculation module and a decision optimization module. The abnormal scoring module integrates multiple basic detectors for multi-dimensional evaluation. The confidence calculation module fuses multiple evaluation results based on an integrated learning framework. The decision optimization module optimizes the evaluation results using a Bayesian decision method to obtain a credibility evaluation result corresponding to the current abnormality.

[0108] The deep temporal convolutional network is a deep learning network for processing time series data. Local patterns and trend information in time series are extracted through multiple convolutional layers. The effective information flow refers to key information that can effectively improve task performance through calculation and propagation in the model. The supervised fine-tuning is a training method based on a pre-trained model. The specific performance of the model is improved by fine-tuning on a specific task. The heat kernel matrix is ​​a mathematical matrix used to represent nonlinear relationships or similarities between variables.

[0109] Collect real-time monitoring data. Real-time monitoring data includes real-time process parameters and real-time equipment status monitoring data. For real-time process parameters, a deep temporal convolutional network is used to extract time domain features. The network consists of three levels: the entry control layer uses gated linear units to control the flow of effective information and avoid interference from invalid information; the feature extraction layer uses multi-scale hole convolution to extract process parameter features of different time scales, for example, hole convolutions of scales 1, 2, and 4 are used to extract short-term, medium-term, and long-term features respectively; the fusion layer is based on the attention mechanism to weightedly fuse features of different scales to obtain more representative time domain features. For example, weights are adaptively assigned according to the importance of the features to fuse short-term, medium-term, and long-term features. For real-time equipment status monitoring data, a wavelet packet decomposition algorithm is used to extract frequency domain features. By adaptively selecting the optimal basis function, the original signal is decomposed into different frequency bands, and the energy features of each frequency band are extracted as frequency domain features. For example, the Daubechies wavelet basis is adaptively selected for decomposition according to the spectral characteristics of the signal to extract energy features of different frequency bands. Assuming that the collected process parameters are temperature, pressure and flow, and the equipment status monitoring data is a vibration signal, their time domain features and frequency domain features are extracted respectively to obtain a real-time feature vector.

[0110] The extracted real-time features are input into the pre-trained contrast diffusion model, which is constructed through two-stage training: the first stage uses unlabeled data for self-supervised training, the feature encoder uses a residual network structure to extract features, and the contrast predictor implements feature projection based on a multi-layer perceptron, and is trained by minimizing the consistency loss between features; the second stage introduces labeled data for supervised fine-tuning, and the features extracted by the feature encoder are input into the classifier for classification, and fine-tuned by minimizing the classification loss. Combined with the contrast learning algorithm, the normal feature distribution is learned using data under normal conditions. For example, the mean and variance of the features under normal conditions are calculated to construct the normal feature distribution.

[0111] Based on the heterogeneous graph diffusion network, the deviation between the real-time features and the normal feature distribution is modeled, including process parameter nodes and equipment status nodes. The connection relationship between the nodes is determined according to the physical connection relationship and the process flow. For example, there is a connection between the temperature sensor node and the pressure sensor node, indicating that there is a physical connection or process association between them. The diffusion process of the feature on the graph is controlled by the heat kernel matrix, and the feature propagation is performed using multi-layer diffusion convolution and attention mechanism. For example, the attention weight is calculated based on the connection strength and feature similarity between the nodes, and the feature is propagated to the adjacent nodes. When the node deviation exceeds the preset deviation threshold, anomaly detection is triggered. For example, when the temperature feature of a node deviates from the normal mean by 2 times the standard deviation, the node is considered to be abnormal.

[0112] If an abnormal node is detected, the causal reasoning engine is activated. A causal probability graph is constructed based on the conditional random field model, where nodes represent variables and edges represent conditional dependencies. For example, if an increase in temperature leads to an increase in pressure, there is a directed edge between the temperature node and the pressure node. The variational inference algorithm is used to calculate the edge probability distribution, and the neural structural equation model is combined to fit the nonlinear relationship between variables and analyze the propagation path of the anomaly in the graph structure. For example, the degree of influence of increased temperature on increased pressure is calculated, and the path of the anomaly propagating from the temperature node to the pressure node is analyzed.

[0113] In order to determine the root cause of the anomaly, a hierarchical multi-agent reinforcement learning system is used. The system includes a policy agent, an expert agent, and an executive agent. The policy agent is responsible for global task allocation, for example, assigning diagnostic tasks to the corresponding expert agent; the expert agent handles specific types of anomalies, for example, the temperature anomaly expert is responsible for diagnosing temperature-related anomalies; the executive agent completes specific diagnosis, for example, collecting sensor data and analyzing it. The agents communicate and make collaborative decisions through a graph attention network. For example, the expert agent passes the diagnosis results to the policy agent, and the policy agent makes the final decision based on the opinions of all experts.

[0114] Based on the hierarchical anomaly prediction model, the trend of the current anomaly is predicted, including short-term prediction module, medium-term prediction module and long-term prediction module. The short-term prediction module uses a bidirectional long short-term memory network to model local features; the medium-term prediction module uses a causal convolutional network to expand the receptive field; the long-term prediction module uses a multi-head self-attention mechanism to capture long-range dependencies. For example, the short-term prediction module predicts the temperature change in the next hour, the medium-term prediction module predicts the temperature change in the next day, and the long-term prediction module predicts the temperature change in the next week. Finally, the abnormal diagnosis result is obtained.

[0115] The credibility evaluation system is used to evaluate the abnormal diagnosis results. The credibility evaluation system includes an abnormal scoring module, a confidence calculation module and a decision optimization module. The abnormal scoring module integrates multiple basic detectors for multi-dimensional evaluation, for example, scoring based on changes in temperature, pressure and flow; the confidence calculation module integrates multiple evaluation results based on an integrated learning framework, for example, weighted averaging the scoring results of multiple detectors to obtain the final confidence; the decision optimization module uses the Bayesian decision method to optimize the evaluation results, for example, adjusting the confidence based on historical data and prior knowledge to obtain the final credibility evaluation results.

[0116] In this embodiment, by combining contrastive learning, graph diffusion network, causal reasoning and other technologies, anomalies can be identified more accurately and the root cause of anomalies can be quickly located. Through the hierarchical anomaly prediction model and credibility assessment system, the development trend of anomalies can be predicted and the credibility of the diagnosis results can be evaluated, providing a more reliable basis for decision-making. By adaptively selecting wavelet basis functions, multi-scale dilated convolution and multi-agent reinforcement learning and other technologies, it can adapt to different working conditions and environments and improve the robustness of the system.

[0117] In an optional embodiment,

[0118] Based on the hierarchical anomaly prediction model, the trend of the current anomaly is predicted. The hierarchical anomaly prediction model includes a short-term prediction module, a medium-term prediction module and a long-term prediction module. The short-term prediction module uses a bidirectional long short-term memory network to model local features, the medium-term prediction module uses a causal convolutional network to expand the receptive field, and the long-term prediction module uses a multi-head self-attention mechanism to capture long-range dependencies. The abnormal diagnosis results include:

[0119] Obtain an abnormal state data sequence corresponding to the current abnormality, perform adaptive segmentation based on the data change rate, and dynamically window divide the abnormal state data sequence to obtain a segmented data sequence;

[0120] Inputting the segmented data sequence into an enhanced bidirectional long short-term memory network for short-term prediction, wherein the enhanced bidirectional long short-term memory network comprises four bidirectional long short-term memory layers, each layer comprises a residual connection mechanism and an adaptive Dropout mechanism, and weights are assigned to features of different time steps through a temporal attention mechanism to obtain a short-term prediction result;

[0121] Input the segmented data sequence into a multi-branch causal convolutional network for medium-term prediction, wherein the multi-branch causal convolutional network comprises a time domain branch, a frequency domain branch and a trend branch, wherein the time domain branch extracts time series features through progressive dilated convolution, the frequency domain branch extracts spectrum features through wavelet transform, and the trend branch extracts main trend features through an adaptive smoothing filter, and the time series features, spectrum features and main trend features are dynamically fused through an attention mechanism to obtain a medium-term prediction result;

[0122] Input the segmented data sequence, the short-term prediction result and the mid-term prediction result into the improved Transformer network for long-term prediction, store historical abnormal pattern information through multi-scale receptive field mechanism and learning position embedding, combine memory enhancement module and time-aware mask mechanism, and fuse the historical abnormal pattern information through multi-head attention mechanism to obtain long-term prediction result;

[0123] A dynamic confidence assessment network is established, and basic confidence is assigned to the short-term prediction results, the medium-term prediction results, and the long-term prediction results based on the current abnormal characteristics and the historical prediction accuracy. The prediction results are fused through a multi-layer attention network to obtain a fused prediction result. The uncertainty of the fused prediction result is estimated according to the Bayesian neural network, and a probability distribution model of the prediction result is established to obtain the abnormal diagnosis result.

[0124] The multi-branch causal convolutional network is a network that learns causal relationships through multiple branch convolutional structures. Each branch is responsible for capturing causal information in the data from a different angle or level, and ultimately forms an overall causal inference result by merging this information. It is suitable for causal inference and analysis tasks in complex systems. The time-aware mask mechanism is a mechanism for introducing time context information when processing time series data.

[0125] Get the abnormal status data sequence corresponding to the current abnormality. For example, the change of CPU usage of a server over a period of time can be expressed as a time series, such as [30, 32, 35, 38, 42, 45, 40, 35, 32, 30, 35, 40, 45, 50, 55, 50, 45, 40, 35, 30].

[0126] Adaptive segmentation based on data change rate. Calculate the change rate between adjacent data points and set a threshold. When the change rate exceeds the threshold, it is considered that a new segment has appeared. Assuming the threshold is 5, the above sequence can be divided into the following segments: [30, 32, 35, 38, 42, 45], [45, 40, 35, 32, 30], [30, 35, 40, 45, 50, 55], [55, 50, 45, 40, 35, 30].

[0127] Perform dynamic window division on the segmented data sequence. According to the characteristics of abnormal development, dynamically adjust the window size, for example, for segments with faster changes, use a smaller window; for segments with slower changes, use a larger window. Assuming the window sizes are 3 and 5 respectively, the first segment [30, 32, 35, 38, 42, 45] can be divided into [30, 32, 35], [32, 35, 38], [35, 38, 42], [38, 42, 45], etc.

[0128] The segmented data sequence is input into the enhanced bidirectional long short-term memory network for short-term prediction. The network contains four layers of bidirectional long short-term memory layers, each layer contains a residual connection mechanism and an adaptive dropout mechanism, and weights are assigned to features at different time steps through a temporal attention mechanism. For example, for the sequence [30, 32, 35], the network learns its short-term trend and predicts the next few values.

[0129] The segmented data sequence is input into a multi-branch causal convolutional network for medium-term prediction. The multi-branch causal convolutional network includes a time domain branch, a frequency domain branch, and a trend branch. The time domain branch extracts time series features through progressive dilated convolution, the frequency domain branch extracts spectrum features through wavelet transform, and the trend branch extracts main trend features through adaptive smoothing filter. These features are dynamically fused through the attention mechanism to obtain the medium-term prediction result.

[0130] The segmented data sequence, short-term prediction results, and mid-term prediction results are input into the improved Transformer network for long-term prediction. The network uses a multi-scale receptive field mechanism and learned position embedding, combined with a memory enhancement module and a time-aware mask mechanism to store historical abnormal pattern information, and fuses historical abnormal pattern information through a multi-head attention mechanism to obtain long-term prediction results.

[0131] Establish a dynamic confidence assessment network. Assign basic confidence to short-term, medium-term and long-term prediction results based on current abnormal characteristics and historical prediction accuracy. Fuse the prediction results through a multi-layer attention network to obtain a fused prediction result. Estimate the uncertainty of the fused prediction result based on the Bayesian neural network, establish a probability distribution model of the prediction result, and obtain the final abnormal diagnosis result.

[0132] In this embodiment, short-term, medium-term and long-term prediction results are combined and fused through a dynamic confidence evaluation mechanism, which can more comprehensively capture the development trend of abnormalities, thereby improving the accuracy of predictions. A variety of advanced deep learning technologies are used, such as residual connections, adaptive Dropout, attention mechanisms, etc., which can effectively improve the robustness and generalization ability of the model. The uncertainty of the prediction results is estimated based on the Bayesian neural network, which can provide the probability distribution of the prediction results, thereby better assisting decision-making.

[0133] S3. Extract the constraints and optimization objectives related to the current anomaly from the knowledge graph to construct a multi-objective optimization problem, construct an optimization strategy generator based on the meta-learning period and adjust the weights of the optimization objectives according to the abnormal scenario, solve the local optimization strategy and the global optimization strategy respectively through a hybrid optimization method based on deep reinforcement learning and evolutionary algorithm, combine and execute the initial optimization strategy, construct a predictive controller based on Neural Ordinary Differential Equations and dynamically constrain and dynamically adjust the optimization process, obtain the initial optimization result and add it to the hybrid model based on knowledge distillation, perform reliability evaluation on the initial optimization result according to the pre-acquired mechanism knowledge and historical data, construct a progressive optimization execution strategy according to the evaluation results, collect effect data when executing the progressive optimization execution strategy and judge whether the optimization goal is achieved, and if so, update the knowledge graph according to the progressive optimization execution strategy.

[0134] The hybrid optimization method is a strategy that combines multiple optimization algorithms, aiming to improve the efficiency and accuracy of problem solving by complementing the advantages of different algorithms. The hybrid model based on knowledge distillation is an optimization method that combines knowledge distillation technology with multi-model fusion. The dynamic constraint means that during the optimization process, as the problem solving progresses, the constraint conditions will be dynamically adjusted according to real-time data or model status. The mechanism knowledge refers to theoretical knowledge in the fields of physics, chemistry, biology, etc., which describes the inherent laws or behaviors of the system. The progressive optimization execution strategy is a step-by-step optimization method, which usually gradually improves the solution at each stage according to the results of the previous stage.

[0135] In an optional embodiment,

[0136] Extracting constraints and optimization objectives related to the current anomaly from the knowledge graph to construct a multi-objective optimization problem, constructing an optimization strategy generator based on the meta-learning period and adjusting the weight of the optimization objective according to the abnormal scenario, solving the local optimization strategy and the global optimization strategy respectively through a hybrid optimization method based on deep reinforcement learning and evolutionary algorithm, combining and executing the initial optimization strategy, constructing a predictive controller based on Neural Ordinary Differential Equations and dynamically constraining and dynamically adjusting the optimization process, obtaining the initial optimization result and adding it to the hybrid model based on knowledge distillation, performing reliability evaluation on the initial optimization result according to the pre-acquired mechanism knowledge and historical data, constructing a progressive optimization execution strategy according to the evaluation result, collecting effect data when executing the progressive optimization execution strategy and judging whether the optimization objective is achieved, and if achieved, updating the knowledge graph according to the progressive optimization execution strategy, including:

[0137] Extract the constraints and optimization goals related to the current anomaly from the knowledge graph, and obtain multi-channel features through the feature extraction layer, wherein the feature extraction layer includes a process parameter channel, an equipment status channel, and an operating condition channel, and integrate the features extracted from each channel to obtain multi-channel feature data;

[0138] A meta-learning optimization strategy generator is constructed based on the multi-channel feature data, wherein the meta-learning optimization strategy generator includes a strategy generation layer, wherein the strategy generation layer sets an input gate and a forget gate, wherein the input gate controls the feature inflow intensity, and the forget gate manages the retention ratio of historical information, and dynamically allocates optimization target weights according to abnormal scene features through a soft attention mechanism;

[0139] A hybrid optimization method is used to solve the optimization strategy. The hybrid optimization method includes local optimization and global optimization. The local optimization adopts a hierarchical reinforcement learning structure, including a high-level strategy network and a dual value evaluation mechanism. The high-level strategy network includes four fully connected layers and uses a LeakyReLU activation function. The dual value evaluation mechanism includes an immediate value network and a long-term value network. The immediate value network evaluates the single-step optimization effect. The long-term value network uses a long-short-term memory network to predict long-term benefits. The optimized samples are stored in a hierarchical experience pool according to priority to obtain a local optimization strategy.

[0140] The global optimization adopts a multi-scale evolutionary framework, divides the population into multiple sub-populations, performs population mutation through Gaussian mutation operator and Cauchy mutation operator, wherein the mutation intensity is negatively correlated with the individual fitness, adopts adaptive factorial crossover to perform population crossover, determines the crossover site through a probabilistic graphical model, maintains population diversity based on crowding sorting, obtains a global optimization strategy and combines it with the local optimization strategy, obtains the initial optimization strategy and executes it;

[0141] In the process of executing the initial optimization strategy, a prediction controller based on a neural ordinary differential equation is constructed and a discrete observation sequence is obtained. The prediction controller includes a state encoding network, a dynamic approximation network and a decoding network. The state encoding network maps the discrete observation sequence into a continuous state vector. The control sequence is restored through the residual block structure in the dynamic approximation network and the deconvolution operation in the decoding network. The process limit is converted into a penalty term in combination with a constraint processing module to obtain an initial optimization result.

[0142] Input the initial optimization result into a hybrid model based on knowledge distillation, wherein the hybrid model includes a teacher network and a student network, wherein the teacher network integrates a process mechanism model, and the student network adopts a deep residual network structure, and performs knowledge migration through an attention migration mechanism, and constructs a three-layer evaluation system to perform reliability evaluation on the optimization scheme, wherein the three-layer evaluation system includes a static evaluation layer, a dynamic evaluation layer, and a comprehensive evaluation layer, wherein the static evaluation layer evaluates the feasibility of the initial optimization strategy based on fuzzy rule reasoning, the dynamic evaluation layer performs random perturbation simulation through a Monte Carlo method, and the comprehensive evaluation layer uses a deep belief network to output a reliability score and outputs the reliability score as an evaluation result;

[0143] A progressive optimization execution strategy is constructed based on the evaluation results, execution effect data is collected and it is determined whether the optimization goal is achieved. When the optimization goal is achieved, the progressive optimization execution strategy is updated to the knowledge graph.

[0144] The soft attention mechanism is a mechanism for weighting input data, which is commonly found in neural networks. The Gaussian mutation operator is a mutation operator used in evolutionary algorithms, which generates new individuals through Gaussian distribution. The Cauchy mutation operator is a mutation operator based on Cauchy distribution, which is similar to the Gaussian mutation operator but has a larger tail probability. The penalty term is an additional term added to the optimization objective function, which is usually used to limit illegal solutions or solutions that violate constraints in the solution space. The Monte Carlo method is a numerical calculation method based on random sampling, which is used to estimate the numerical solutions of complex systems. The deep belief network is a multi-layer neural network, which is usually used for unsupervised learning tasks.

[0145] Extract constraints and optimization goals related to the current anomaly from the knowledge graph. For example, the knowledge graph of a chemical plant contains information such as process parameters, equipment status, and operating conditions of various chemical reactions. When the reaction temperature rises abnormally, constraints related to temperature control, such as cooling water flow restrictions, reactor pressure restrictions, etc., and optimization goals, such as reducing the temperature to a normal range and maintaining product quality, can be extracted from the knowledge graph. Multi-channel features are obtained through the feature extraction layer, which contains a process parameter channel, an equipment status channel, and an operating condition channel. For example, the process parameter channel extracts data such as reaction temperature, pressure, and flow, the equipment status channel extracts data such as the speed of the cooling water pump and the opening of the valve, and the operating condition channel extracts data such as ambient temperature and humidity. Integrate the features extracted from each channel to obtain multi-channel feature data, such as integrating data such as temperature, pressure, flow, speed, opening, ambient temperature, and humidity into a multidimensional vector.

[0146] A meta-learning optimization strategy generator is constructed based on multi-channel feature data, which includes a strategy generation layer, which sets an input gate and a forget gate. The input gate controls the strength of feature inflow, such as adjusting the feature weight according to the severity of the current anomaly. The forget gate manages the proportion of historical information retention, such as retaining different proportions of historical data according to different anomaly types. The soft attention mechanism dynamically allocates optimization target weights according to the characteristics of the abnormal scene. For example, if the abnormal temperature rise is serious, the weight of lowering the temperature will be increased, while the weight of maintaining product quality will be lowered. Assuming that the current anomaly is that the reaction temperature is too high, the multi-channel feature data shows that the temperature is 10 degrees higher than the normal value, the pressure is slightly increased, and the flow is normal. Through the soft attention mechanism, the optimization target weight for lowering the temperature is set to 0.8, and the weight for maintaining product quality is set to 0.2.

[0147] A hybrid optimization method is used to solve the optimization strategy. The hybrid optimization method includes local optimization and global optimization. The local optimization adopts a hierarchical reinforcement learning structure, including a high-level policy network and a dual value evaluation mechanism. The high-level policy network includes four fully connected layers and uses a LeakyReLU activation function to output local optimization actions, such as adjusting the cooling water flow. The dual value evaluation mechanism includes an immediate value network and a long-term value network. The immediate value network evaluates the effect of single-step optimization, such as evaluating the degree of temperature reduction after adjusting the cooling water flow. The long-term value network uses a long short-term memory network to predict long-term benefits, such as predicting the temperature change trend within a period of time after adjusting the cooling water flow. The optimized samples are stored in a hierarchical experience pool according to priority, for example, samples with obvious temperature reduction effects are given a higher priority. The local optimization strategy is obtained through reinforcement learning training. The global optimization adopts a multi-scale evolutionary framework to divide the population into multiple sub-populations, and the population mutation is performed by Gaussian mutation operators and Cauchy mutation operators. The mutation intensity is negatively correlated with the individual fitness. Adaptive factorial crossover is used for population crossover, and the crossover site is determined by a probabilistic graph model. The population diversity is maintained based on the crowding ranking to obtain the global optimization strategy. The global optimization strategy is combined with the local optimization strategy to obtain the initial optimization strategy and execute it. For example, the global optimization strategy determines the approximate adjustment range of the cooling water flow rate, and the local optimization strategy makes fine adjustments within this range.

[0148] In the process of executing the initial optimization strategy, a predictive controller based on the Neural Ordinary Differential Equation is constructed and a discrete observation sequence is obtained. The predictive controller contains a state encoding network, a dynamic approximation network, and a decoding network. The state encoding network maps the discrete observation sequence into a continuous state vector. The dynamic approximation network simulates the system dynamics through a residual block structure, and the decoding network restores the control sequence through a deconvolution operation. Combined with the constraint processing module, the process limit is converted into a penalty term, such as converting the cooling water flow limit into a penalty term to prevent exceeding the limit. The initial optimization result is obtained. For example, the predictive controller predicts the temperature change trend in the future period of time based on the current temperature, pressure, flow and other data, and outputs the adjustment value of the cooling water flow.

[0149] The initial optimization results are input into a hybrid model based on knowledge distillation. The model includes a teacher network and a student network. The teacher network integrates the process mechanism model, and the student network adopts a deep residual network structure. Knowledge transfer is performed through the attention transfer mechanism. A three-layer evaluation system is constructed to evaluate the reliability of the optimization scheme. The system includes a static evaluation layer, a dynamic evaluation layer, and a comprehensive evaluation layer. The static evaluation layer evaluates the feasibility of the initial optimization strategy based on fuzzy rule reasoning. The dynamic evaluation layer performs random perturbation simulation through the Monte Carlo method. The comprehensive evaluation layer uses a deep belief network to output the reliability score and outputs the reliability score as the evaluation result. For example, if the reliability score is lower than the threshold, the optimization scheme is considered unreliable.

[0150] Based on the evaluation results, a progressive optimization execution strategy is constructed, and the execution effect data is collected to determine whether the optimization goal has been achieved. When the optimization goal is achieved, the progressive optimization execution strategy is updated to the knowledge graph. For example, if the temperature has dropped to the normal range, the optimization goal is considered to be achieved.

[0151] In this embodiment, through the combination of knowledge graph and deep reinforcement learning, anomalies can be identified quickly and accurately and optimization strategies can be formulated, thereby shortening the exception handling time and improving production efficiency. The meta-learning optimization strategy generator can dynamically adjust the optimization target weight according to the abnormal scenario, the hybrid optimization method can take into account local and global optimization, the predictive controller can dynamically constrain and adjust the optimization process, and the reliability evaluation system can evaluate the reliability of the optimization scheme, thereby improving the accuracy and reliability of exception handling. The progressive optimization execution strategy and the knowledge graph update mechanism enable the system to continuously learn new exception handling experience and improve the system's adaptability and robustness.

[0152] In an optional embodiment,

[0153] The global optimization adopts a multi-scale evolutionary framework, divides the population into multiple sub-populations, and performs population mutation through Gaussian mutation operator and Cauchy mutation operator, wherein the mutation intensity is negatively correlated with the individual fitness, adopts adaptive factorial crossover to perform population crossover, determines the crossover site through a probabilistic graph model, maintains population diversity based on crowding sorting, and obtains the global optimization strategy including:

[0154] Divide the process parameter space to obtain multiple search areas of different scales, establish the corresponding relationship between each search area and a sub-population, allocate the search space to each sub-population according to the value range and accuracy requirements of the process parameters, and construct the initial population;

[0155] Performing a mutation operation on each individual in the initial population to obtain a mutant individual, using a Gaussian mutation operator to perform a local fine search on the individual, using a Cauchy mutation operator to perform a global wide-area search on the individual, detecting the relationship between the fitness value of the individual and the average fitness value of the initial population, and when it is detected that the fitness value of the individual is lower than the average fitness value of the initial population, increasing the mutation strength to a preset first mutation strength, and when it is detected that the fitness value of the individual is higher than the average fitness value of the initial population, reducing the mutation strength to a preset second mutation strength;

[0156] Constructing a parameter correlation network and calculating the correlation strength between process parameters, identifying parameter pairs whose correlation strength is greater than a preset threshold to form a strongly correlated parameter group, performing an adaptive factorial crossover operation on the parameters in the strongly correlated parameter group, detecting the fitness value of the parent individual and adjusting the crossover probability according to the detection result;

[0157] Count the evolutionary generations and trigger population migration when the preset cycle is reached, calculate the fitness values ​​of all individuals in the population and sort them, randomly select individuals from the top ten percent of the individuals in fitness as high-quality individuals and migrate the high-quality individuals to the sub-population of the adjacent scale, replace the individuals with the lowest fitness in the target sub-population with the high-quality individuals, count the number of migrations of the high-quality individuals, and remove the high-quality individuals whose migration times exceed the preset number;

[0158] Calculate the distance between each individual in the population and other individuals in the parameter space and the target space, perform normalization and weighted summation to obtain the crowding degree of each individual, select and retain individuals based on the crowding degree, detect the population diversity index, and increase the mutation intensity and generate random individuals when the population diversity index is detected to be lower than a preset threshold;

[0159] The mutation operation, crossover operation, population migration and diversity maintenance are performed cyclically until the preset number of iterations is reached to obtain the global optimization strategy.

[0160] The population diversity index is an indicator used to measure the degree of population diversity. In an optimization algorithm, the diversity of a population usually refers to the distribution breadth of individuals in the solution space. The population diversity index can evaluate whether the algorithm has an early convergence problem.

[0161] According to the value range and accuracy requirements of the process parameters, the entire parameter space is divided into multiple search areas of different scales. For example, if the value range of a parameter is 0 to 100, it can be divided into 10 sub-areas, and the range of each sub-area is 10. Then, a subpopulation is assigned to each search area, and initial individuals are randomly generated to form an initial population. Each individual represents a set of process parameter combinations. Assuming that a process has three parameters, each individual is a vector containing three parameter values. For example, an initial individual can be represented as [25, 55, 85].

[0162] Perform mutation operation on each subpopulation. Use Gaussian mutation operator for local fine search, and Cauchy mutation operator for global wide search. The mutation intensity is adaptively adjusted according to the individual fitness. For example, if the fitness value of an individual is lower than the average fitness value of the subpopulation, its mutation intensity is increased to the preset first mutation intensity, such as 0.1; otherwise, its mutation intensity is reduced to the preset second mutation intensity, such as 0.01. Suppose an individual is [25, 55, 85], after Gaussian mutation, it may become [25.1, 54.9, 85.2], and after Cauchy mutation, it may become [30, 60, 90].

[0163] Perform a crossover operation to construct a parameter correlation network. For example, the Pearson correlation coefficient between each pair of parameters can be calculated, and its absolute value can be used as the correlation strength. A threshold value, such as 0.8, is set to identify parameter pairs with a correlation strength greater than the threshold as a strongly correlated parameter group. Perform an adaptive factorial crossover operation on the parameters within the strongly correlated parameter group. The crossover probability is adjusted according to the fitness value of the parent individual. For example, if the fitness value of the parent individual is high, the crossover probability is increased, such as 0.9; otherwise, the crossover probability is reduced, such as 0.5. Suppose there are two parent individuals [25, 55, 85] and [30, 60, 90], the first and second parameters constitute a strongly correlated parameter group, and the crossover probability is 0.7, then new individuals [25, 60, 90] and [30, 55, 85] may be generated.

[0164] Perform population migration. Population migration is triggered every certain number of evolutionary generations, such as every 10 generations. Randomly select a portion of the top 10% of individuals in each subpopulation and migrate them to the subpopulation of the adjacent scale to replace the individuals with the lowest fitness in the target subpopulation. The number of migrations of each individual is counted. If the number of migrations of an individual exceeds the preset number, such as 5 times, it will be eliminated.

[0165] Maintain diversity. Calculate the distance between each individual and other individuals in the parameter space and the target space, normalize them, and then perform weighted summation to obtain the crowding degree of each individual. Screen and retain individuals based on the crowding degree to maintain population diversity. If the population diversity index is lower than the preset threshold, such as 0.5, increase the mutation intensity and generate some random individuals to increase population diversity.

[0166] The above mutation, crossover, migration and diversity maintenance operations are executed cyclically until a preset number of iterations, such as 100, is reached, and finally a global optimization strategy, that is, a set of optimal process parameter combinations, is obtained.

[0167] In this embodiment, through multi-scale search strategy and adaptive parameter adjustment, the global optimal solution can be found quickly, the calculation time is reduced, and the combination of Gaussian mutation and Cauchy mutation, taking into account local fine search and global wide-area search, can effectively jump out of the local optimum and find the global optimal solution. Through crowding sorting and diversity maintenance mechanism, the premature maturity of the population can be effectively avoided, ensuring the robustness and stability of the algorithm.

[0168] Figure 2 The structure diagram of the petrochemical production process abnormality diagnosis and optimization system integrated with the knowledge graph according to the embodiment of the present invention is as follows: Figure 2 As shown, the system comprises:

[0169] The first unit is used to collect process parameters and equipment status monitoring data in the petrochemical production process, pre-process the process parameters through adaptive wavelet threshold denoising, select the optimal decomposition scale and threshold in combination with the attention mechanism to obtain denoised process parameters, extract multi-scale features of the equipment status monitoring data through deep variational mode decomposition, and adaptively weight them in combination with the multi-head self-attention mechanism to obtain equipment status feature vectors, construct a heterogeneous graph neural network based on the denoised process parameters and the equipment status feature vectors, extract node representations through graph contrast learning and calculate the causal strength between different nodes in combination with the neural causal discovery network, construct a causal relationship graph, convert expert experience into graph structure constraints and node attribute constraints through the knowledge distillation method and optimize the causal relationship graph, and complete the missing relationships in combination with graph structure reasoning to obtain the knowledge graph corresponding to the petrochemical production process;

[0170] The second unit is used to collect real-time monitoring data and perform feature extraction to obtain real-time features and add them to the pre-trained contrast diffusion model, determine the normal feature distribution corresponding to the normal working condition in combination with the contrast learning algorithm, and model the deviation between the real-time features and the normal feature distribution through the graph diffusion network. If the deviation exceeds the preset deviation threshold, anomaly detection is triggered, and the causal reasoning engine is activated according to the node where the abnormality is detected. The propagation path of the anomaly in the graph structure is analyzed in combination with the causal relationship in the knowledge graph and the neural structural equation model, and the root cause of the anomaly is determined in combination with the multi-agent reinforcement learning method, and the trend of the current anomaly is predicted in combination with the pre-set hierarchical anomaly prediction model, so as to obtain the credibility evaluation corresponding to the current anomaly;

[0171] The third unit is used to extract constraints and optimization objectives related to the current anomaly from the knowledge graph to construct a multi-objective optimization problem, construct an optimization strategy generator based on the meta-learning period and adjust the weights of the optimization objectives according to the abnormal scenario, solve the local optimization strategy and the global optimization strategy respectively through a hybrid optimization method based on deep reinforcement learning and evolutionary algorithm, combine and execute the initial optimization strategy, construct a predictive controller based on Neural Ordinary Differential Equations and dynamically constrain and dynamically adjust the optimization process, obtain the initial optimization result and add it to the hybrid model based on knowledge distillation, perform reliability evaluation on the initial optimization result according to the pre-acquired mechanism knowledge and historical data, construct a progressive optimization execution strategy according to the evaluation results, collect effect data when executing the progressive optimization execution strategy and judge whether the optimization goal is achieved, and if completed, update the knowledge graph according to the progressive optimization execution strategy.

[0172] According to a third aspect of the embodiments of the present invention,

[0173] An electronic device is provided, comprising:

[0174] processor;

[0175] a memory for storing processor-executable instructions;

[0176] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0177] According to a fourth aspect of the embodiments of the present invention,

[0178] 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.

[0179] 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.

[0180] 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 petrochemical production process abnormality diagnosis and optimization method integrating knowledge graph, characterized in that: include: Collect process parameters and equipment status monitoring data in the petrochemical production process, pre-process the process parameters through adaptive wavelet threshold denoising, select the optimal decomposition scale and threshold in combination with the attention mechanism to obtain denoised process parameters, extract multi-scale features of the equipment status monitoring data through deep variational mode decomposition, and adaptively weight them in combination with the multi-head self-attention mechanism to obtain the equipment status feature vector, build a heterogeneous graph neural network based on the denoised process parameters and the equipment status feature vector, extract node representations through graph contrast learning and calculate the causal strength between different nodes in combination with the neural causal discovery network, build a causal relationship graph, convert expert experience into graph structure constraints and node attribute constraints through knowledge distillation method and optimize the causal relationship graph, and complete the missing relationship in combination with graph structure reasoning to obtain the knowledge graph corresponding to the petrochemical production process; Collect real-time monitoring data and perform feature extraction to obtain real-time features and add them to the pre-trained contrast diffusion model. Combine the contrast learning algorithm to determine the normal feature distribution corresponding to the normal working condition. Model the deviation between the real-time feature and the normal feature distribution through the graph diffusion network. If the deviation exceeds the preset deviation threshold, trigger anomaly detection, activate the causal reasoning engine according to the node where the anomaly is detected, combine the causal relationship in the knowledge graph and the neural structural equation model to analyze the propagation path of the anomaly in the graph structure, combine the multi-agent reinforcement learning method to determine the root cause of the anomaly, and combine the pre-set hierarchical anomaly prediction model to predict the trend of the current anomaly, and obtain the credibility evaluation corresponding to the current anomaly; The constraints and optimization objectives related to the current anomaly are extracted from the knowledge graph to construct a multi-objective optimization problem, an optimization strategy generator based on the meta-learning period is constructed, and the weights of the optimization objectives are adjusted according to the abnormal scenarios, the local optimization strategy and the global optimization strategy are solved respectively by a hybrid optimization method based on deep reinforcement learning and evolutionary algorithm, the initial optimization strategy is obtained by combination and executed, a predictive controller based on Neural Ordinary Differential Equations is constructed, and the optimization process is dynamically constrained and dynamically adjusted, the initial optimization result is obtained and added to the hybrid model based on knowledge distillation, the reliability of the initial optimization result is evaluated according to the pre-acquired mechanism knowledge and historical data, a progressive optimization execution strategy is constructed according to the evaluation results, the effect data when executing the progressive optimization execution strategy is collected and it is determined whether the optimization goal is achieved, and if completed, the knowledge graph is updated according to the progressive optimization execution strategy.

2. The method according to claim 1, characterized in that Collect process parameters and equipment status monitoring data in the petrochemical production process, pre-process the process parameters through adaptive wavelet threshold denoising, select the optimal decomposition scale and threshold in combination with the attention mechanism, obtain denoised process parameters, extract multi-scale features of the equipment status monitoring data through deep variational mode decomposition, and perform adaptive weighting in combination with the multi-head self-attention mechanism to obtain the equipment status feature vector including: Collecting process parameters and equipment status monitoring data in the petrochemical production process through a distributed data acquisition network, wherein the process parameters include temperature parameters, pressure parameters and flow parameters, and the equipment status monitoring data includes equipment vibration data and corresponding timestamps; The process parameters are preprocessed by an adaptive wavelet threshold denoising network based on Transformer. The adaptive wavelet threshold denoising network includes an encoder and a decoder. The encoder analyzes the time-frequency characteristics of the signal through a multi-layer self-attention module. The decoder adaptively determines the soft threshold parameters according to the noise characteristics of different frequency bands. The self-attention module in the encoder retains the signal timing information through position encoding and analyzes the correlation between time domain and frequency domain characteristics. The optimal wavelet decomposition parameters are adaptively selected according to the characteristics of different types of process parameters. The temperature parameter is decomposed by the db4 wavelet basis function, the pressure parameter is decomposed by the sym8 wavelet basis function, and the flow parameter is decomposed by the coif5 wavelet basis function. The denoising thresholds of different frequency bands are adaptively calculated by residual connection and layer normalization algorithm combined with Bayesian optimization. The signal energy, variance and entropy value are used as optimization targets to obtain the denoising process parameters. A deep variational mode decomposition network is used to extract features from the equipment status monitoring data. The deep variational mode decomposition network performs preliminary feature extraction through multi-layer one-dimensional convolution, decomposes the equipment vibration data into multiple characteristic mode functions, each characteristic mode function characterizes the equipment characteristics of a specific frequency band, the number of channels of the one-dimensional convolution increases successively, and a multi-head self-attention mechanism is used to perform weighted fusion on the characteristic mode functions. The multi-head self-attention mechanism calculates the attention scores of different frequency bands through dot product operations, adaptively allocates weights according to the importance of frequency band features, and generates equipment status feature vectors.

3. The method according to claim 1, characterized in that A heterogeneous graph neural network is constructed based on the denoising process parameters and the equipment state feature vector, node representations are extracted through graph contrast learning, and the causal strength between different nodes is calculated in combination with a neural causal discovery network, a causal relationship graph is constructed, expert experience is converted into graph structure constraints and node attribute constraints through a knowledge distillation method, and the causal relationship graph is optimized, and the missing relationship is completed in combination with graph structure reasoning to obtain a knowledge graph corresponding to the petrochemical production process, including: Constructing a heterogeneous graph neural network based on the denoising process parameters and the equipment state feature vector, wherein the heterogeneous graph neural network includes process parameter nodes and equipment state nodes, and performing feature conversion on different types of nodes through a multi-layer graph attention layer, wherein the feature conversion matrix of the graph attention layer is implemented by a fully connected layer; A graph contrast learning method with dynamic negative sampling is used to extract node representations. Positive sample pairs are obtained by sampling from the temporal neighborhood under the same working condition, and negative sample pairs are obtained by random sampling from different working conditions. The node feature similarity is calculated based on cosine similarity, and the node representation is optimized through a contrast loss function. The causal relationship between nodes is calculated through a neural causal discovery network based on a dual-stream architecture. The neural causal discovery network includes a temporal feature extraction stream and a causal inference stream. The temporal feature extraction stream uses a deep residual network to extract temporal features, and the causal inference stream uses a graph attention network to perform causal inference. The features of the two streams are fused through a gating mechanism to calculate the causal scores between node pairs and establish a causal relationship graph. Through the knowledge distillation method of the teacher-student network structure, the expert experience is converted into graph structure constraints and node attribute constraints, and a graph neural reasoning network is used to complete the missing relationships in the causal graph. The graph neural reasoning network includes a feature propagation layer and a relationship reasoning layer. The feature propagation layer updates the node representation in the causal graph through gated graph convolution, and the relationship reasoning layer calculates the confidence of the edge through the attention mechanism. The reasoning results are filtered according to the pre-set confidence threshold to obtain the knowledge graph corresponding to the petrochemical production process.

4. The method according to claim 1, characterized in that: Collect real-time monitoring data and perform feature extraction to obtain real-time features and add them to the pre-trained contrast diffusion model. Combine the contrast learning algorithm to determine the normal feature distribution corresponding to the normal working condition. Model the deviation between the real-time feature and the normal feature distribution through the graph diffusion network. If the deviation exceeds the preset deviation threshold, trigger anomaly detection. Activate the causal reasoning engine according to the node where the anomaly is detected. Combine the causal relationship in the knowledge graph and the neural structural equation model to analyze the propagation path of the anomaly in the graph structure. Combine the multi-agent reinforcement learning method to determine the root cause of the anomaly and combine the preset hierarchical anomaly prediction model to predict the trend of the current anomaly. Obtain the corresponding credibility assessment of the current anomaly, including: Collecting real-time monitoring data, the real-time monitoring data includes real-time process parameters and real-time equipment status monitoring data, the real-time process parameters extract time domain features through a deep temporal convolutional network, the deep temporal convolutional network consists of an entry control layer, a feature extraction layer and a fusion layer, the entry control layer uses a gated linear unit to select effective information flow, the feature extraction layer uses a multi-scale hole convolution to extract features, and the fusion layer integrates features based on an attention mechanism, the real-time equipment status monitoring data extracts frequency domain features through a wavelet packet decomposition algorithm, and adaptively selects the optimal basis function to obtain real-time features; The real-time features are input into a pre-trained contrast diffusion model, wherein the contrast diffusion model is constructed by a multi-stage pre-training method. In the first stage, unlabeled data is used for self-supervised training, a feature encoder adopts a residual network structure, and a contrast predictor implements feature projection based on a multi-layer perceptron. In the second stage, labeled data is introduced for supervised fine-tuning, and a normal feature distribution corresponding to a normal operating condition is determined in combination with a contrast learning algorithm. Modeling the deviation between the real-time feature and the normal feature distribution based on a heterogeneous graph diffusion network, wherein the heterogeneous graph diffusion network includes process parameter nodes and equipment status nodes, the connection relationship between the nodes is determined according to the physical connection relationship and the process flow, the diffusion process is controlled by a heat kernel matrix, and feature propagation is performed based on multi-layer diffusion convolution and attention mechanism, and anomaly detection is triggered when the node deviation exceeds a preset deviation threshold; The causal inference engine is activated according to the detected abnormal nodes, and a causal probability graph is constructed based on the conditional random field model. The nodes represent variables and the edges represent conditional dependencies. The variational inference algorithm is used to calculate the edge probability distribution, and the neural structural equation model is combined to fit the nonlinear relationship between variables and analyze the propagation path of the anomaly in the graph structure. A hierarchical multi-agent reinforcement learning system is used to determine the root cause of the anomaly. The hierarchical multi-agent reinforcement learning system includes a strategy agent, an expert agent, and an execution agent. The strategy agent is responsible for global task allocation, the expert agent handles specific types of anomalies, and the execution agent completes specific diagnosis. The agents transmit messages and make collaborative decisions through a graph attention network. Based on the hierarchical anomaly prediction model, the trend of the current anomaly is predicted. The hierarchical anomaly prediction model includes a short-term prediction module, a medium-term prediction module and a long-term prediction module. The short-term prediction module uses a bidirectional long short-term memory network to model local features, the medium-term prediction module uses a causal convolutional network to expand the receptive field, and the long-term prediction module uses a multi-head self-attention mechanism to capture long-range dependencies to obtain anomaly diagnosis results; A credibility evaluation system is used to evaluate the abnormal diagnosis results. The credibility evaluation system includes an abnormal scoring module, a confidence calculation module and a decision optimization module. The abnormal scoring module integrates multiple basic detectors for multi-dimensional evaluation. The confidence calculation module fuses multiple evaluation results based on an integrated learning framework. The decision optimization module optimizes the evaluation results using a Bayesian decision method to obtain a credibility evaluation result corresponding to the current abnormality.

5. The method according to claim 4, characterized in that Based on the hierarchical anomaly prediction model, the trend of the current anomaly is predicted. The hierarchical anomaly prediction model includes a short-term prediction module, a medium-term prediction module and a long-term prediction module. The short-term prediction module uses a bidirectional long short-term memory network to model local features, the medium-term prediction module uses a causal convolutional network to expand the receptive field, and the long-term prediction module uses a multi-head self-attention mechanism to capture long-range dependencies. The abnormal diagnosis results include: Obtain an abnormal state data sequence corresponding to the current abnormality, perform adaptive segmentation based on the data change rate, and dynamically window divide the abnormal state data sequence to obtain a segmented data sequence; Inputting the segmented data sequence into an enhanced bidirectional long short-term memory network for short-term prediction, wherein the enhanced bidirectional long short-term memory network comprises four bidirectional long short-term memory layers, each layer comprises a residual connection mechanism and an adaptive Dropout mechanism, and weights are assigned to features of different time steps through a temporal attention mechanism to obtain a short-term prediction result; Input the segmented data sequence into a multi-branch causal convolutional network for medium-term prediction, wherein the multi-branch causal convolutional network comprises a time domain branch, a frequency domain branch and a trend branch, wherein the time domain branch extracts time series features through progressive dilated convolution, the frequency domain branch extracts spectrum features through wavelet transform, and the trend branch extracts main trend features through an adaptive smoothing filter, and the time series features, spectrum features and main trend features are dynamically fused through an attention mechanism to obtain a medium-term prediction result; Input the segmented data sequence, the short-term prediction result and the mid-term prediction result into the improved Transformer network for long-term prediction, store historical abnormal pattern information through multi-scale receptive field mechanism and learning position embedding, combine memory enhancement module and time-aware mask mechanism, and fuse the historical abnormal pattern information through multi-head attention mechanism to obtain long-term prediction result; A dynamic confidence assessment network is established, and basic confidence is assigned to the short-term prediction results, the medium-term prediction results, and the long-term prediction results based on the current abnormal characteristics and the historical prediction accuracy. The prediction results are fused through a multi-layer attention network to obtain a fused prediction result. The uncertainty of the fused prediction result is estimated according to the Bayesian neural network, and a probability distribution model of the prediction result is established to obtain the abnormal diagnosis result.

6. The method according to claim 1, characterized in that Extracting constraints and optimization objectives related to the current anomaly from the knowledge graph to construct a multi-objective optimization problem, constructing an optimization strategy generator based on the meta-learning period and adjusting the weight of the optimization objective according to the abnormal scenario, solving the local optimization strategy and the global optimization strategy respectively through a hybrid optimization method based on deep reinforcement learning and evolutionary algorithm, combining and executing the initial optimization strategy, constructing a predictive controller based on Neural Ordinary Differential Equations and dynamically constraining and dynamically adjusting the optimization process, obtaining the initial optimization result and adding it to the hybrid model based on knowledge distillation, performing reliability evaluation on the initial optimization result according to the pre-acquired mechanism knowledge and historical data, constructing a progressive optimization execution strategy according to the evaluation result, collecting effect data when executing the progressive optimization execution strategy and judging whether the optimization objective is achieved, and if achieved, updating the knowledge graph according to the progressive optimization execution strategy, including: Extract the constraints and optimization goals related to the current anomaly from the knowledge graph, and obtain multi-channel features through the feature extraction layer, wherein the feature extraction layer includes a process parameter channel, an equipment status channel, and an operating condition channel, and integrate the features extracted from each channel to obtain multi-channel feature data; A meta-learning optimization strategy generator is constructed based on the multi-channel feature data, wherein the meta-learning optimization strategy generator includes a strategy generation layer, wherein the strategy generation layer sets an input gate and a forget gate, wherein the input gate controls the feature inflow intensity, and the forget gate manages the retention ratio of historical information, and dynamically allocates optimization target weights according to abnormal scene features through a soft attention mechanism; A hybrid optimization method is used to solve the optimization strategy. The hybrid optimization method includes local optimization and global optimization. The local optimization adopts a hierarchical reinforcement learning structure, including a high-level strategy network and a dual value evaluation mechanism. The high-level strategy network includes four fully connected layers and uses a LeakyReLU activation function. The dual value evaluation mechanism includes an immediate value network and a long-term value network. The immediate value network evaluates the single-step optimization effect. The long-term value network uses a long-short-term memory network to predict long-term benefits. The optimized samples are stored in a hierarchical experience pool according to priority to obtain a local optimization strategy. The global optimization adopts a multi-scale evolutionary framework, divides the population into multiple sub-populations, performs population mutation through Gaussian mutation operator and Cauchy mutation operator, wherein the mutation intensity is negatively correlated with the individual fitness, adopts adaptive factorial crossover to perform population crossover, determines the crossover site through a probabilistic graphical model, maintains population diversity based on crowding sorting, obtains a global optimization strategy and combines it with the local optimization strategy, obtains the initial optimization strategy and executes it; In the process of executing the initial optimization strategy, a prediction controller based on a neural ordinary differential equation is constructed and a discrete observation sequence is obtained. The prediction controller includes a state encoding network, a dynamic approximation network and a decoding network. The state encoding network maps the discrete observation sequence into a continuous state vector. The control sequence is restored through the residual block structure in the dynamic approximation network and the deconvolution operation in the decoding network. The process limit is converted into a penalty term in combination with a constraint processing module to obtain an initial optimization result. Input the initial optimization result into a hybrid model based on knowledge distillation, wherein the hybrid model includes a teacher network and a student network, wherein the teacher network integrates a process mechanism model, and the student network adopts a deep residual network structure, and performs knowledge migration through an attention migration mechanism, and constructs a three-layer evaluation system to perform reliability evaluation on the optimization scheme, wherein the three-layer evaluation system includes a static evaluation layer, a dynamic evaluation layer, and a comprehensive evaluation layer, wherein the static evaluation layer evaluates the feasibility of the initial optimization strategy based on fuzzy rule reasoning, the dynamic evaluation layer performs random perturbation simulation through a Monte Carlo method, and the comprehensive evaluation layer uses a deep belief network to output a reliability score and outputs the reliability score as an evaluation result; A progressive optimization execution strategy is constructed based on the evaluation results, execution effect data is collected and it is determined whether the optimization goal is achieved. When the optimization goal is achieved, the progressive optimization execution strategy is updated to the knowledge graph.

7. The method according to claim 6, characterized in that The global optimization adopts a multi-scale evolutionary framework, divides the population into multiple sub-populations, and performs population mutation through Gaussian mutation operator and Cauchy mutation operator, wherein the mutation intensity is negatively correlated with the individual fitness, adopts adaptive factorial crossover to perform population crossover, determines the crossover site through a probabilistic graph model, maintains population diversity based on crowding sorting, and obtains the global optimization strategy including: Divide the process parameter space to obtain multiple search areas of different scales, establish the corresponding relationship between each search area and a sub-population, allocate the search space to each sub-population according to the value range and accuracy requirements of the process parameters, and construct the initial population; Performing a mutation operation on each individual in the initial population to obtain a mutant individual, using a Gaussian mutation operator to perform a local fine search on the individual, using a Cauchy mutation operator to perform a global wide-area search on the individual, detecting the relationship between the fitness value of the individual and the average fitness value of the initial population, and when it is detected that the fitness value of the individual is lower than the average fitness value of the initial population, increasing the mutation strength to a preset first mutation strength, and when it is detected that the fitness value of the individual is higher than the average fitness value of the initial population, reducing the mutation strength to a preset second mutation strength; Constructing a parameter correlation network and calculating the correlation strength between process parameters, identifying parameter pairs whose correlation strength is greater than a preset threshold to form a strongly correlated parameter group, performing an adaptive factorial crossover operation on the parameters in the strongly correlated parameter group, detecting the fitness value of the parent individual and adjusting the crossover probability according to the detection result; Count the evolutionary generations and trigger population migration when the preset cycle is reached, calculate the fitness values ​​of all individuals in the population and sort them, randomly select individuals from the top ten percent of the individuals in fitness as high-quality individuals and migrate the high-quality individuals to the sub-population of the adjacent scale, replace the individuals with the lowest fitness in the target sub-population with the high-quality individuals, count the number of migrations of the high-quality individuals, and remove the high-quality individuals whose migration times exceed the preset number; Calculate the distance between each individual in the population and other individuals in the parameter space and the target space, perform normalization and weighted summation to obtain the crowding degree of each individual, select and retain individuals based on the crowding degree, detect the population diversity index, and increase the mutation intensity and generate random individuals when the population diversity index is detected to be lower than a preset threshold; The mutation operation, crossover operation, population migration and diversity maintenance are performed cyclically until the preset number of iterations is reached to obtain the global optimization strategy.

8. A petrochemical production process abnormality diagnosis and optimization system integrated with a knowledge graph, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect process parameters and equipment status monitoring data in the petrochemical production process, pre-process the process parameters through adaptive wavelet threshold denoising, select the optimal decomposition scale and threshold in combination with the attention mechanism to obtain denoised process parameters, extract multi-scale features of the equipment status monitoring data through deep variational mode decomposition, and adaptively weight them in combination with the multi-head self-attention mechanism to obtain equipment status feature vectors, construct a heterogeneous graph neural network based on the denoised process parameters and the equipment status feature vectors, extract node representations through graph contrast learning and calculate the causal strength between different nodes in combination with the neural causal discovery network, construct a causal relationship graph, convert expert experience into graph structure constraints and node attribute constraints through the knowledge distillation method and optimize the causal relationship graph, and complete the missing relationships in combination with graph structure reasoning to obtain the knowledge graph corresponding to the petrochemical production process; The second unit is used to collect real-time monitoring data and perform feature extraction to obtain real-time features and add them to the pre-trained contrast diffusion model, determine the normal feature distribution corresponding to the normal working condition in combination with the contrast learning algorithm, and model the deviation between the real-time features and the normal feature distribution through the graph diffusion network. If the deviation exceeds the preset deviation threshold, anomaly detection is triggered, and the causal reasoning engine is activated according to the node where the abnormality is detected. The propagation path of the anomaly in the graph structure is analyzed in combination with the causal relationship in the knowledge graph and the neural structural equation model, and the root cause of the anomaly is determined in combination with the multi-agent reinforcement learning method, and the trend of the current anomaly is predicted in combination with the pre-set hierarchical anomaly prediction model, so as to obtain the credibility evaluation corresponding to the current anomaly; The third unit is used to extract constraints and optimization objectives related to the current anomaly from the knowledge graph to construct a multi-objective optimization problem, construct an optimization strategy generator based on the meta-learning period and adjust the weights of the optimization objectives according to the abnormal scenario, solve the local optimization strategy and the global optimization strategy respectively through a hybrid optimization method based on deep reinforcement learning and evolutionary algorithm, combine and execute the initial optimization strategy, construct a predictive controller based on Neural Ordinary Differential Equations and dynamically constrain and dynamically adjust the optimization process, obtain the initial optimization result and add it to the hybrid model based on knowledge distillation, perform reliability evaluation on the initial optimization result according to the pre-acquired mechanism knowledge and historical data, construct a progressive optimization execution strategy according to the evaluation results, collect effect data when executing the progressive optimization execution strategy and judge whether the optimization goal is achieved, and if completed, update the knowledge graph according to the progressive optimization execution strategy.

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