Method and system for intelligent auxiliary diagnosis and maintenance based on multi-channel recall

Abstract

The present invention provides a method and system for intelligent auxiliary diagnosis and maintenance based on multi-channel recall, which relates to the field of fault processing technology, including collecting equipment operation data, extracting and enhancing time series features after preprocessing, constructing a feature dependency graph, and generating a state feature vector. A multi-channel recall diagnosis network is used to recall fault features, and a fault propagation map is generated by fusion, and then a fault diagnosis result is generated. Finally, a preliminary maintenance plan is generated based on the fault diagnosis result and the equipment state transition probability matrix, and a hierarchical reinforcement learning method is used to optimize the maintenance plan in combination with historical cases and fault propagation maps to generate an optimal maintenance plan. The present invention can improve the accuracy of fault diagnosis and the effectiveness of maintenance plans, and reduce equipment maintenance costs.

Description

Technical Field

[0001] The present invention relates to fault handling technology, and in particular to a method and system for intelligent auxiliary diagnosis and maintenance based on multi-channel recall. Background Art

[0002] As the intelligence of industrial equipment continues to improve, the complexity of equipment fault diagnosis and maintenance is also increasing. Traditional fault diagnosis methods mainly rely on expert experience and fixed rules, which are difficult to cope with the dynamic changes of equipment operating status and fault modes, resulting in low diagnostic accuracy and inefficient maintenance plan formulation.

[0003] Most existing intelligent diagnosis methods use a single feature extraction and fault identification model, which fails to fully utilize the multi-source heterogeneous information contained in the equipment's historical data. At the same time, due to the lack of in-depth analysis of the fault propagation mechanism, it is difficult to accurately grasp the fault development trend, and the maintenance plan often fails to meet actual needs, resulting in a waste of maintenance resources.

[0004] Therefore, there is an urgent need for an intelligent auxiliary diagnosis and maintenance method that can comprehensively utilize multi-source information and integrate multiple diagnostic strategies. The accuracy of fault diagnosis can be improved through multi-channel feature recall and dynamic fusion mechanism, and the intelligent optimization of maintenance plans can be achieved based on fault propagation analysis and historical case matching, thereby improving equipment maintenance efficiency and reliability. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for intelligent auxiliary diagnosis and maintenance based on multi-channel recall, which can solve the problems in the prior art.

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

[0007] Provided is a method for intelligent auxiliary diagnosis and maintenance based on multi-channel recall, comprising:

[0008] Collect equipment operation data and perform adaptive filtering and noise reduction processing to obtain filtered data, use a recursive neural network to extract time series features from the filtered data, enhance the time series features through an adversarial learning network to obtain enhanced features, construct a feature dependency graph of the enhanced features based on a causal reasoning method, establish a device state transition probability matrix based on the feature dependency graph, and combine the enhanced features with the device state transition probability matrix to generate a state feature vector;

[0009] The state feature vector is input into a multi-channel recall diagnosis network for fault feature recall, wherein the multi-channel recall diagnosis network includes a fault mode recall channel based on a graph neural network, an expert rule recall channel based on causal reasoning, and a historical case recall channel based on deep reinforcement learning. The weight coefficient of the output results of each recall channel is calculated through a dynamic attention network and fused to obtain a fusion feature. A fault propagation map is constructed based on the fusion feature, and a graph convolutional network is used to extract fault propagation features from the fault propagation map. A fault diagnosis result including fault type information, fault degree information, and fault trend information is generated according to the fault propagation feature. The fault diagnosis result is input into a hierarchical decision model, and a meta-learning method is used to convert the equipment state transition probability matrix into a maintenance strategy parameter, and a preliminary maintenance plan is generated according to the maintenance strategy parameter.

[0010] The preliminary maintenance plan and the fault propagation graph are input into a case matching model. The case matching model uses a graph attention mechanism to extract repair paths from historical maintenance cases, and uses a temporal graph neural network to model the dependency relationship of the maintenance steps in the repair path to obtain a step dependency model. The fault diagnosis result is input into the step dependency model as a constraint condition, and a hierarchical reinforcement learning method is used to optimize the maintenance plan. In the maintenance strategy optimization submodule, the maintenance strategy parameters are used to plan the sequence of maintenance steps, and in the maintenance parameter optimization submodule, the maintenance parameters are adjusted according to the fault degree information, and the optimal maintenance plan is generated through a multi-objective optimization algorithm.

[0011] In an optional embodiment,

[0012] The equipment operation data is collected and adaptively filtered and denoised to obtain filtered data, a recursive neural network is used to extract time series features from the filtered data, and an adversarial learning network is used to enhance the time series features to obtain enhanced features including:

[0013] The vibration data, temperature data, pressure data and current data of the equipment are collected through the sensor array to obtain the equipment operation data, a coupling relationship matrix between the measurement variables is established based on the physical model of the equipment, the equipment operation data is input into the encoder of the variational autoencoder to obtain the latent variable distribution parameters, the latent variable distribution parameters are input into the decoder for data reconstruction, and the reconstruction error is optimized by maximizing the lower bound of evidence to obtain the reconstructed data;

[0014] Performing spectrum analysis on the reconstructed data to obtain frequency band energy distribution, calculating a signal-to-noise ratio evaluation value based on the frequency band energy distribution, jointly optimizing center frequency parameters, bandwidth parameters, and filter order parameters of low-pass, band-pass, and high-pass filters through an adaptive optimization algorithm using the signal-to-noise ratio evaluation value, updating a multiscale filter group according to the optimized parameters, and performing adaptive filtering on the reconstructed data to obtain filtered data;

[0015] Input the filtered data into a bidirectional gated recurrent unit, calculate the correlation matrix of the filtered data at different time steps to obtain the time dimension attention weight, use spatial position encoding to calculate the spatial dimension attention weight, recalibrate different channel features of the filtered data to obtain channel recalibration weights, adaptively fuse the time dimension attention weights, spatial dimension attention weights and channel recalibration weights to obtain comprehensive attention weights, and use the comprehensive attention weights to weight the filtered data to obtain time series features;

[0016] The time series features are input into the feature extractor of the conditional adversarial energy network to obtain intermediate features, the intermediate features are passed through the discriminator of the residual network structure to obtain discrimination results, positive and negative sample pairs are constructed based on the discrimination results, the network is trained by minimizing the energy difference of the positive sample pairs and maximizing the energy difference of the negative sample pairs, mutual information regularization terms and gradient penalty terms are introduced in the training process, and the trained conditional adversarial energy network is used to perform domain adaptation enhancement processing on the time series features to obtain enhanced features.

[0017] In an optional embodiment,

[0018] Constructing a feature dependency graph of the enhanced feature based on a causal reasoning method, establishing a device state transition probability matrix according to the feature dependency graph, and combining the enhanced feature with the device state transition probability matrix to generate a state feature vector includes:

[0019] Calculating conditional mutual information values ​​between feature pairs for enhanced features to obtain a feature correlation matrix, binarizing the feature correlation matrix based on a preset binarization threshold to obtain an adjacency matrix, and constructing an undirected graph structure according to the adjacency matrix;

[0020] Calculating a conditional independence score based on observed data for each node pair in the undirected graph structure, identifying a separation set in the undirected graph structure based on the conditional independence score, using the separation set to determine the causal relationship between nodes, identifying non-connected nodes with common child nodes as parent nodes of corresponding child nodes to obtain partial directed edges, and using a minimum directed rule to orient the remaining undirected edges to obtain a directed graph structure;

[0021] Inputting the directed graph structure into a graph attention network, generating a query vector, a key vector and a value vector for each node in the directed graph structure through a multi-layer perceptron, calculating the dot product of the query vector and the key vector to obtain an attention score, normalizing the attention score to obtain an attention weight, taking the product of the attention weight and the value vector as the causal strength between nodes, and constructing a weighted directed graph based on the causal strength;

[0022] Input the weighted directed graph into a dynamic Bayesian network, construct a conditional probability distribution for each time step, use a variational inference method to approximate the conditional probability distribution to a Gaussian distribution, optimize the mean and variance parameters of the Gaussian distribution by minimizing the variational free energy, calculate the characteristic transition probabilities between adjacent time steps based on the optimized distribution parameters, and combine the characteristic transition probabilities into a device state migration probability matrix;

[0023] The enhanced features and the device state transition probability matrix are projected onto a common feature space through feature mapping to obtain projection features, the correlation between the projection features is calculated to obtain modal correlation weights, and the projection features are weightedly fused based on the modal correlation weights to obtain a state feature vector.

[0024] In an optional embodiment,

[0025] The state feature vector is input into a multi-channel recall diagnosis network for fault feature recall. The multi-channel recall diagnosis network includes a fault mode recall channel based on a graph neural network, an expert rule recall channel based on causal reasoning, and a historical case recall channel based on deep reinforcement learning. The weight coefficient of the output result of each recall channel is calculated through a dynamic attention network and fused to obtain a fusion feature including:

[0026] In the fault mode recall channel of the graph neural network, a fault knowledge graph is constructed, the state feature vector is input into the fault knowledge graph, node embedding is performed based on the association relationship between the equipment component node, the fault mode node, the fault symptom node and the operating parameter node in the fault knowledge graph to obtain the node feature, the attention score between adjacent nodes is calculated by the graph attention network, the node feature is weightedly updated using the attention score to obtain the node embedding feature, and the cosine similarity between the node embedding feature and the fault feature in the preset fault mode library is calculated to obtain the first recall feature;

[0027] In the expert rule recall channel based on causal reasoning, historical operation data is collected and conditional independence test is performed to obtain the degree of dependence between feature pairs. An initial directed graph is constructed based on the degree of dependence. The causal intervention method is used to analyze the influence of features on the occurrence of faults to obtain the intervention effect. The intervention effect is used as the edge weight to construct a weighted directed causal graph.

[0028] Input the state feature vector into a weighted directed causal graph, calculate the causal effect strength between features based on the edge weights in the weighted directed causal graph, select key features according to the causal effect strength to construct a causal chain to obtain a causal feature, calculate the matching degree and confidence of the causal feature and the rule feature in the expert rule base to obtain a second recall feature;

[0029] In the historical case recall channel based on deep reinforcement learning, a multi-layer neural network structure is designed, wherein the number of input layer nodes is set to the state feature dimension, the number of output layer nodes is set to the number of retrieval actions, and the multi-layer neural network structure is pre-trained using historical retrieval data to obtain a deep Q network;

[0030] Input the state feature vector into the deep Q network to obtain the state action value of the retrieval action, perform feature matching in the historical case library according to the state action value to obtain the retrieval result, form the retrieval result and the state feature vector into a training sample to update the parameters of the deep Q network to obtain an optimized network, and input the state feature vector into the optimized network to obtain a third recall feature;

[0031] The first recall feature, the second recall feature and the third recall feature are respectively linearly transformed to obtain the corresponding query vector, key vector and value vector, the dot product of each group of query vectors and the key vector is calculated to obtain the attention score, the attention score is normalized to obtain the attention weight, the first recall feature, the second recall feature and the third recall feature are respectively multiplied by the corresponding attention weight and superimposed to obtain the fused feature.

[0032] In an optional embodiment,

[0033] A fault propagation map is constructed based on the fusion features, a graph convolutional network is used to extract fault propagation features from the fault propagation map, a fault diagnosis result including fault type information, fault degree information and fault trend information is generated according to the fault propagation features, the fault diagnosis result is input into a hierarchical decision model, a meta-learning method is used to convert the equipment state transition probability matrix into maintenance strategy parameters, and a preliminary maintenance plan is generated according to the maintenance strategy parameters, including:

[0034] Deconstructing the fused features into a node feature matrix and an adjacency matrix, wherein the node feature matrix includes the operation status information, fault type information and component topology relationship information of the equipment components, and the adjacency matrix includes the physical connection relationship and logical dependency relationship between the components;

[0035] Calculating mutual information entropy and conditional mutual information entropy for component pairs in the node feature matrix, determining a fault propagation direction based on the mutual information entropy, determining a propagation strength based on the conditional mutual information entropy, and constructing a fault propagation graph by taking the fault propagation direction as the direction of an edge and taking the propagation strength as the weight of an edge;

[0036] Construct a graph convolution network including a spatial convolution layer and a temporal convolution layer, calculate the attention weight based on the propagation strength in the spatial convolution layer, use the attention weight to weightedly aggregate the features of the neighborhood nodes to obtain spatial features, perform sliding window segmentation on the spatial features in the temporal convolution layer to obtain temporal segments, perform a one-dimensional convolution operation on the temporal segments to obtain local temporal features, concatenate the local temporal features with the spatial features to obtain a single-layer spatiotemporal feature, perform residual connection and feature fusion on each single-layer spatiotemporal feature to obtain a fault propagation feature;

[0037] A multi-task shared encoder and a task-specific decoder are used to analyze the fault propagation characteristics, a classification decoder is used to identify the fault type and fault location, a regression decoder is used to evaluate the fault severity and remaining life, and a sequence decoder is used to predict the fault development trend and propagation path to obtain a fault diagnosis result;

[0038] Based on the fault diagnosis results, a Markov decision process model is constructed, the equipment state is used as the state space, the maintenance operation is used as the action space, a state transition probability matrix is ​​established, and a meta-learning framework including a policy network and a value network is designed. In the inner loop, the policy gradient method is used to optimize the timing selection and method selection parameters of the maintenance strategy based on the state transition probability matrix, and in the outer loop, the Monte Carlo method is used to update the state transition probability based on the actual maintenance effect;

[0039] According to the optimized maintenance strategy parameters, a preliminary maintenance plan is generated in combination with equipment maintenance specifications and resource constraints. The preliminary maintenance plan includes maintenance items, maintenance time and maintenance resource configuration information.

[0040] In an optional embodiment,

[0041] The preliminary maintenance plan and the fault propagation graph are input into a case matching model. The case matching model uses a graph attention mechanism to extract a repair path from historical maintenance cases, and uses a temporal graph neural network to model the dependency relationship of the maintenance steps in the repair path to obtain a step dependency model including:

[0042] The preliminary maintenance plan and the fault propagation graph are input into a case matching model, and the case matching model performs graph matching through a multi-layer graph attention network, wherein the node matching layer calculates the matching degree of the component fault feature vector by cosine similarity to obtain a node similarity matrix, and the edge matching layer calculates the edge matching degree by the Euclidean distance of the propagation feature vector to obtain an edge similarity matrix;

[0043] Based on the node similarity matrix and the edge similarity matrix, a multi-head attention mechanism including fault type, fault degree, propagation direction and propagation intensity is used to calculate the similarity score of each dimension, and the similarity scores of each dimension are fused through adaptive weights to obtain a comprehensive similarity vector.

[0044] Retrieving a similar case set from a historical case library based on the comprehensive similarity vector, extracting a maintenance operation sequence from the similar case set to form a repair path library, wherein each repair path in the repair path library includes a maintenance step identifier, an execution sequence, resource configuration, and time information;

[0045] A bidirectional temporal graph neural network is constructed, and the maintenance step identifiers in the repair path library are mapped to graph nodes. In the forward propagation, a gated recurrent unit is used to extract the sequential dependency features of the maintenance steps. In the backward propagation, an attention mechanism is used to capture the long-range dependency relationship of the maintenance steps. The sequential dependency features and the long-range dependency relationship are concatenated and then the step dependency features are obtained through the graph attention layer.

[0046] A message passing framework is constructed based on the step dependency features to generate a node message including the step completion status, resource occupancy status, and associated step status, and the node message is iteratively updated multiple times through a message aggregation function to obtain a target node representation;

[0047] A graph convolutional network is used to extract features from the target node representation, and a step dependency model including explicit timing constraints and implicit resource constraints is constructed, wherein the explicit timing constraints represent the prerequisite relationship and parallel relationship of the maintenance steps, and the implicit resource constraints represent the resource exclusive relationship of the maintenance steps.

[0048] In an optional embodiment,

[0049] The fault diagnosis result is input as a constraint condition into a step-dependent model, and a hierarchical reinforcement learning method is used to optimize the maintenance plan, wherein the maintenance strategy parameters are used in the maintenance strategy optimization submodule to plan the maintenance step sequence, and the maintenance parameters are adjusted according to the fault degree information in the maintenance parameter optimization submodule. The optimal maintenance plan is generated by a multi-objective optimization algorithm, including:

[0050] The fault diagnosis result is input as a constraint condition into a step dependency model, wherein the step dependency model includes a maintenance step node and dependency edges between steps, wherein the maintenance step node includes a resource requirement vector, an execution duration, and a priority weight attribute, and the dependency edge includes a preceding relationship identifier and a parallel relationship identifier, and the priority weight of the maintenance step node is updated according to the fault diagnosis result, and a priority execution constraint is set for the relevant maintenance steps based on the fault diagnosis result;

[0051] Constructing a maintenance strategy optimization objective function, wherein the maintenance strategy optimization objective function includes a total maintenance time item, a maintenance cost item, a residual risk item, and is subject to a preceding relationship identifier constraint of the dependency relationship edge between the steps, a resource capacity constraint, a time window constraint, and the priority execution constraint;

[0052] A mixed integer programming method is used to solve the maintenance strategy optimization objective function to obtain an initial maintenance step execution sequence and an initial resource allocation plan;

[0053] Constructing a maintenance parameter optimization objective function, wherein the maintenance parameter optimization objective function includes a maintenance quality index item, a step execution time item, and a step cost item;

[0054] Based on the fault degree information in the fault diagnosis result, a gradient-based multi-objective optimization algorithm is used to solve the maintenance parameter optimization objective function to obtain maintenance depth parameters, maintenance process parameters, detection parameters and quality control parameters;

[0055] The actual execution time is calculated according to the maintenance parameters, and a deviation analysis is performed between the actual execution time and the planned time in the initial maintenance step execution sequence. Based on the deviation analysis results, the initial maintenance step execution sequence is adjusted, and the maintenance parameters of the affected steps are triggered to be re-optimized until the optimization results of the maintenance strategy optimization objective function and the maintenance parameter optimization objective function converge. According to the converged optimization results, an optimal maintenance plan including step execution timing, parameter configuration plan, resource scheduling plan and quality control requirements is generated.

[0056] A second aspect of an embodiment of the present invention provides a system for intelligent auxiliary diagnosis and maintenance based on multi-path recall, comprising:

[0057] The first unit is used to collect equipment operation data and perform adaptive filtering and noise reduction processing to obtain filtered data, extract time series features from the filtered data using a recursive neural network, enhance the time series features through an adversarial learning network to obtain enhanced features, construct a feature dependency graph of the enhanced features based on a causal reasoning method, establish a device state transition probability matrix according to the feature dependency graph, and combine the enhanced features with the device state transition probability matrix to generate a state feature vector;

[0058] The second unit is used to input the state feature vector into a multi-channel recall diagnosis network for fault feature recall, wherein the multi-channel recall diagnosis network includes a fault mode recall channel based on a graph neural network, an expert rule recall channel based on causal reasoning, and a historical case recall channel based on deep reinforcement learning. The weight coefficient of the output results of each recall channel is calculated by a dynamic attention network and fused to obtain a fusion feature, a fault propagation map is constructed based on the fusion feature, a graph convolutional network is used to extract fault propagation features from the fault propagation map, a fault diagnosis result including fault type information, fault degree information and fault trend information is generated according to the fault propagation feature, the fault diagnosis result is input into a hierarchical decision model, a meta-learning method is used to convert the equipment state transition probability matrix into maintenance strategy parameters, and a preliminary maintenance plan is generated according to the maintenance strategy parameters;

[0059] The third unit is used to input the preliminary maintenance plan and the fault propagation map into a case matching model, the case matching model uses a graph attention mechanism to extract repair paths from historical maintenance cases, and uses a temporal graph neural network to model the dependency relationship of the maintenance steps in the repair path to obtain a step dependency model, and inputs the fault diagnosis result as a constraint condition into the step dependency model, and uses a hierarchical reinforcement learning method to optimize the maintenance plan, wherein the maintenance strategy parameters are used to plan the maintenance step sequence in the maintenance strategy optimization submodule, and the maintenance parameters are adjusted according to the fault degree information in the maintenance parameter optimization submodule, and the optimal maintenance plan is generated through a multi-objective optimization algorithm.

[0060] A third aspect of the embodiments of the present invention

[0061] An electronic device is provided, comprising:

[0062] processor;

[0063] a memory for storing processor-executable instructions;

[0064] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0065] A fourth aspect of the embodiments of the present invention is:

[0066] 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.

[0067] In this embodiment, through a multi-channel recall diagnosis network, combined with graph neural networks, causal reasoning, deep reinforcement learning and other technical means, fault feature recall is performed from different angles, and the results of each channel are fused through a dynamic attention network to construct a fault propagation map, which can more comprehensively and accurately identify the type, degree and trend of equipment faults, thereby improving the accuracy of fault diagnosis. Meta-learning is used to convert the equipment state transition probability matrix into maintenance strategy parameters, and combined with case matching models and hierarchical reinforcement learning methods, the maintenance step sequence and parameters are optimized according to the fault diagnosis results and historical maintenance cases, and finally the optimal maintenance plan is generated to avoid unnecessary maintenance operations and improve maintenance efficiency. A variety of advanced artificial intelligence technologies are comprehensively utilized to realize the intelligent diagnosis and maintenance of equipment faults, reduce dependence on manual experience, and improve the intelligent level of equipment maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a flowchart of a method for intelligent auxiliary diagnosis and maintenance based on multi-channel recall according to an embodiment of the present invention;

[0069] Figure 2 The present invention is a schematic diagram of the structure of a system for intelligent auxiliary diagnosis and maintenance based on multi-channel recall according to an embodiment of the present invention. DETAILED DESCRIPTION

[0070] 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.

[0071] 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.

[0072] Figure 1 FIG. 1 is a flow chart of a method for intelligent auxiliary diagnosis and maintenance based on multi-channel recall according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0073] S101. Collect equipment operation data and perform adaptive filtering and noise reduction processing to obtain filtered data, use a recursive neural network to extract time series features from the filtered data, enhance the time series features through an adversarial learning network to obtain enhanced features, construct a feature dependency graph of the enhanced features based on a causal reasoning method, establish a device state transition probability matrix based on the feature dependency graph, and combine the enhanced features with the device state transition probability matrix to generate a state feature vector;

[0074] S102. Input the state feature vector into a multi-channel recall diagnosis network for fault feature recall, wherein the multi-channel recall diagnosis network includes a fault mode recall channel based on a graph neural network, an expert rule recall channel based on causal reasoning, and a historical case recall channel based on deep reinforcement learning. The weight coefficient of the output results of each recall channel is calculated through a dynamic attention network and fused to obtain a fusion feature. A fault propagation map is constructed based on the fusion feature, and a graph convolutional network is used to extract fault propagation features from the fault propagation map. A fault diagnosis result including fault type information, fault degree information, and fault trend information is generated according to the fault propagation feature. The fault diagnosis result is input into a hierarchical decision model, and a meta-learning method is used to convert the equipment state transition probability matrix into maintenance strategy parameters, and a preliminary maintenance plan is generated according to the maintenance strategy parameters.

[0075] S103. The preliminary maintenance plan and the fault propagation map are input into the case matching model. The case matching model uses a graph attention mechanism to extract repair paths from historical maintenance cases, and uses a temporal graph neural network to model the dependency relationship of the maintenance steps in the repair path to obtain a step dependency model. The fault diagnosis result is input into the step dependency model as a constraint condition, and a hierarchical reinforcement learning method is used to optimize the maintenance plan. In the maintenance strategy optimization submodule, the maintenance strategy parameters are used to plan the maintenance step sequence, and in the maintenance parameter optimization submodule, the maintenance parameters are adjusted according to the fault degree information, and the optimal maintenance plan is generated through a multi-objective optimization algorithm.

[0076] The equipment is operated to collect data, including multi-dimensional sensor signals such as vibration, temperature, pressure, and current, with a sampling frequency of 1kHz. The Kalman filter is used to perform adaptive noise reduction on the collected data, and the noise covariance matrix is ​​dynamically adjusted to reduce the impact of measurement noise.

[0077] The long short-term memory network is used to extract time series features from the filtered data. The network contains a three-layer bidirectional LSTM structure with a hidden layer dimension of 128. The feature extraction of long sequence data is achieved through a sliding window mechanism. The extracted time series features are input into the adversarial learning network for feature enhancement. Both the generator and the discriminator use a four-layer convolutional neural network structure to achieve adversarial learning of feature representation.

[0078] In an optional implementation, collecting device operation data and performing adaptive filtering and noise reduction processing to obtain filtered data, extracting time series features from the filtered data using a recursive neural network, and enhancing the time series features through an adversarial learning network to obtain enhanced features include:

[0079] The vibration data, temperature data, pressure data and current data of the equipment are collected through the sensor array to obtain the equipment operation data, a coupling relationship matrix between the measurement variables is established based on the physical model of the equipment, the equipment operation data is input into the encoder of the variational autoencoder to obtain the latent variable distribution parameters, the latent variable distribution parameters are input into the decoder for data reconstruction, and the reconstruction error is optimized by maximizing the lower bound of evidence to obtain the reconstructed data;

[0080] Performing spectrum analysis on the reconstructed data to obtain frequency band energy distribution, calculating a signal-to-noise ratio evaluation value based on the frequency band energy distribution, jointly optimizing center frequency parameters, bandwidth parameters, and filter order parameters of low-pass, band-pass, and high-pass filters through an adaptive optimization algorithm using the signal-to-noise ratio evaluation value, updating a multiscale filter group according to the optimized parameters, and performing adaptive filtering on the reconstructed data to obtain filtered data;

[0081] Input the filtered data into a bidirectional gated recurrent unit, calculate the correlation matrix of the filtered data at different time steps to obtain the time dimension attention weight, use spatial position encoding to calculate the spatial dimension attention weight, recalibrate different channel features of the filtered data to obtain channel recalibration weights, adaptively fuse the time dimension attention weights, spatial dimension attention weights and channel recalibration weights to obtain comprehensive attention weights, and use the comprehensive attention weights to weight the filtered data to obtain time series features;

[0082] The time series features are input into the feature extractor of the conditional adversarial energy network to obtain intermediate features, the intermediate features are passed through the discriminator of the residual network structure to obtain discrimination results, positive and negative sample pairs are constructed based on the discrimination results, the network is trained by minimizing the energy difference of the positive sample pairs and maximizing the energy difference of the negative sample pairs, mutual information regularization terms and gradient penalty terms are introduced in the training process, and the trained conditional adversarial energy network is used to perform domain adaptation enhancement processing on the time series features to obtain enhanced features.

[0083] Exemplarily, first, collect equipment operation data. Vibration data, temperature data, pressure data and current data during equipment operation are collected by arranging sensor arrays at key parts of the equipment. For example, acceleration sensors, temperature sensors, pressure sensors and current sensors are arranged on the bearing seat, motor housing and pipeline of a rotating mechanical equipment, and equipment operation data is synchronously collected at a sampling frequency of 10kHz. Each collection time is 10 seconds, and an equipment operation data set containing 4 types of sensor data and 100,000 sampling points for each type of data is obtained.

[0084] Then, the collected data is reconstructed based on the physical model of the device. According to the structure and working principle of the device, a matrix describing the coupling relationship between the measured variables is established. For example, according to the dynamic model of the rotating machinery, a coupling relationship matrix between vibration, temperature, pressure and current is established, which reflects the mutual influence and transmission relationship between different physical quantities. The collected equipment operation data is input into the encoder of the variational autoencoder to obtain the hidden variable distribution parameters, such as mean and variance. These parameters are input into the decoder for data reconstruction. The reconstruction process is optimized by minimizing the difference between the reconstructed data and the original data, such as the mean square error, to obtain the reconstructed equipment operation data.

[0085] Next, the reconstructed data is subjected to adaptive filtering and denoising. The reconstructed data is subjected to spectral analysis to obtain the energy distribution of different frequency bands. The signal-to-noise ratio evaluation value is calculated based on the frequency band energy distribution, for example, the ratio of the signal frequency band energy to the noise frequency band energy is used as the signal-to-noise ratio evaluation value. Using the signal-to-noise ratio evaluation value, the center frequency parameters, bandwidth parameters and filter order parameters of the low-pass, band-pass and high-pass filters are jointly optimized through an adaptive optimization algorithm, such as a particle swarm optimization algorithm. For example, based on the signal-to-noise ratio evaluation value, the best filter parameter combination is searched to maximize the signal-to-noise ratio of the filtered data. The multi-scale filter group is updated according to the optimized parameters, and the reconstructed data is adaptively filtered to obtain the filtered data.

[0086] Subsequently, temporal features are extracted from the filtered data. The filtered data is input into a bidirectional gated recurrent unit network. The correlation matrix of the filtered data at different time steps is calculated to obtain the time dimension attention weight. The spatial dimension attention weight is calculated using spatial position encoding. The different channel features of the filtered data are recalibrated to obtain the channel recalibration weight. For example, the features of different channels are normalized using a Softmax function to obtain the channel recalibration weight. The time dimension attention weight, the spatial dimension attention weight, and the channel recalibration weight are adaptively fused, for example, by weighted averaging, to obtain a comprehensive attention weight. The filtered data is weighted using the comprehensive attention weight to obtain the temporal features.

[0087] Finally, the extracted time series features are enhanced. The time series features are input into the feature extractor of the conditional adversarial energy network to obtain intermediate features. The intermediate features are input into the discriminator of the residual network structure to obtain the discrimination results. Positive and negative sample pairs are constructed based on the discrimination results. The network is trained by minimizing the energy difference of the positive sample pairs and maximizing the energy difference of the negative sample pairs. Mutual information regularization terms and gradient penalty terms are introduced during the training process to improve the generalization ability and stability of the network. The time series features are domain-adaptively enhanced using the trained conditional adversarial energy network to obtain enhanced features. These enhanced features can be used for subsequent equipment status monitoring and fault diagnosis.

[0088] In this embodiment, through reconstruction and adaptive filtering based on the physical model of the equipment, noise interference is effectively removed, and the signal-to-noise ratio and reliability of the data are improved. By combining the attention mechanism of time, space and channel dimensions, more representative time series features are extracted, and enhanced through adversarial learning networks, the expressiveness and discrimination of the features are improved. By introducing mutual information regularization terms and gradient penalty terms, the generalization ability and robustness of the model are enhanced, so that it can adapt to equipment operation status monitoring and fault diagnosis under different working conditions.

[0089] In an optional implementation, constructing a feature dependency graph of the enhanced feature based on a causal reasoning method, establishing a device state transition probability matrix according to the feature dependency graph, and combining the enhanced feature with the device state transition probability matrix to generate a state feature vector includes:

[0090] Calculating conditional mutual information values ​​between feature pairs for enhanced features to obtain a feature correlation matrix, binarizing the feature correlation matrix based on a preset binarization threshold to obtain an adjacency matrix, and constructing an undirected graph structure according to the adjacency matrix;

[0091] Calculating a conditional independence score based on observed data for each node pair in the undirected graph structure, identifying a separation set in the undirected graph structure based on the conditional independence score, using the separation set to determine the causal relationship between nodes, identifying non-connected nodes with common child nodes as parent nodes of corresponding child nodes to obtain partial directed edges, and using a minimum directed rule to orient the remaining undirected edges to obtain a directed graph structure;

[0092] Inputting the directed graph structure into a graph attention network, generating a query vector, a key vector and a value vector for each node in the directed graph structure through a multi-layer perceptron, calculating the dot product of the query vector and the key vector to obtain an attention score, normalizing the attention score to obtain an attention weight, taking the product of the attention weight and the value vector as the causal strength between nodes, and constructing a weighted directed graph based on the causal strength;

[0093] Input the weighted directed graph into a dynamic Bayesian network, construct a conditional probability distribution for each time step, use a variational inference method to approximate the conditional probability distribution to a Gaussian distribution, optimize the mean and variance parameters of the Gaussian distribution by minimizing the variational free energy, calculate the characteristic transition probabilities between adjacent time steps based on the optimized distribution parameters, and combine the characteristic transition probabilities into a device state migration probability matrix;

[0094] The enhanced features and the device state transition probability matrix are projected onto a common feature space through feature mapping to obtain projection features, the correlation between the projection features is calculated to obtain modal correlation weights, and the projection features are weightedly fused based on the modal correlation weights to obtain a state feature vector.

[0095] Exemplarily, first, the enhanced features are preprocessed, such as normalization or standardization, to eliminate the dimensional differences between different features. Then, the conditional mutual information value between each pair of enhanced features is calculated, which is used to measure the degree of dependence between two features given other features. These conditional mutual information values ​​are combined into a feature correlation matrix. A binarization threshold is set to convert the feature correlation matrix into an adjacency matrix. Among them, the binarization threshold refers to a boundary for converting continuous mutual information values ​​in the feature correlation matrix into discrete values. If the conditional mutual information value is greater than or equal to the binarization threshold, the value of the corresponding position of the matrix is ​​set to 1, indicating that there is a significant correlation between the features. If the conditional mutual information value is less than the binarization threshold, the value of the corresponding position of the matrix is ​​set to 0, indicating that the correlation between the features is weak or non-existent. The elements in the adjacency matrix are 0 or 1, indicating whether there is a connection relationship between the features. An undirected graph structure is constructed based on the adjacency matrix, in which nodes represent features and edges represent connections between features.

[0096] Next, for each node pair in the undirected graph structure, a conditional independence score based on the observed data is calculated. This score reflects whether two nodes are independent of each other given other nodes. Based on the conditional independence score, a disjunction set in the undirected graph structure is identified. A disjunction set is a set of nodes that can block the information transfer between two nodes. The disjunction set is used to determine the causal relationship between nodes. Non-connected nodes with common child nodes are identified as parent nodes of the corresponding child nodes, thereby obtaining some directed edges. The minimum directed rule is used to direct the remaining undirected edges, and finally a directed graph structure is obtained.

[0097] The obtained directed graph structure is input into the graph attention network. For each node in the directed graph structure, a query vector, a key vector, and a value vector are generated by a multi-layer perceptron. The dot product of the query vector and the key vector is calculated to obtain the attention score. The attention score is normalized to obtain the attention weight, which reflects the strength of the causal relationship between the nodes. The product of the attention weight and the value vector is used as the causal strength between the nodes, and a weighted directed graph is constructed based on the causal strength.

[0098] The weighted directed graph is input into the dynamic Bayesian network. For each time step, a conditional probability distribution is constructed, which describes the probability distribution of the child node state given the parent node state. The conditional probability distribution is approximated as a Gaussian distribution using a variational inference method. The mean and variance parameters of the Gaussian distribution are optimized by minimizing the variational free energy. Based on the optimized distribution parameters, the feature transition probability between adjacent time steps is calculated, which represents the probability of the feature value changing from one time step to the next. These feature transition probabilities are combined into a device state migration probability matrix.

[0099] Finally, the enhanced features and the device state transition probability matrix are projected into the common feature space through feature mapping to obtain the projected features. For example, principal component analysis or other dimensionality reduction techniques can be used for projection. The correlation between the projected features is calculated to obtain the modal correlation weight. The projected features are weighted fused based on the modal correlation weight to obtain the final state feature vector.

[0100] Assume that there are three enhanced features: temperature, pressure, and vibration. Through the above method, a weighted directed graph can be constructed, in which temperature is the parent node of pressure and vibration. Assume that an increase in temperature will lead to an increase in pressure and enhanced vibration. The device state transition probability matrix can describe the transition probability of temperature, pressure, and vibration at different time steps. The final state feature vector will contain temperature, pressure, vibration, and the causal relationship and time dependency information between them.

[0101] In this embodiment, by considering the causal relationship and time dependency between features, the real state of the equipment can be more accurately reflected, thereby improving the accuracy of condition monitoring. By constructing the equipment state transition probability matrix, the state of the equipment in the future time step can be predicted, so that preventive measures can be taken in advance to avoid equipment failure. Through more accurate condition monitoring and prediction, unnecessary maintenance work can be reduced, thereby reducing equipment maintenance costs.

[0102] In an optional implementation, the state feature vector is input into a multi-channel recall diagnosis network for fault feature recall, wherein the multi-channel recall diagnosis network includes a fault mode recall channel based on a graph neural network, an expert rule recall channel based on causal reasoning, and a historical case recall channel based on deep reinforcement learning. The weight coefficient of the output result of each recall channel is calculated by a dynamic attention network and the fusion feature is obtained by fusion, including:

[0103] In the fault mode recall channel of the graph neural network, a fault knowledge graph is constructed, the state feature vector is input into the fault knowledge graph, node embedding is performed based on the association relationship between the equipment component node, the fault mode node, the fault symptom node and the operating parameter node in the fault knowledge graph to obtain the node feature, the attention score between adjacent nodes is calculated by the graph attention network, the node feature is weightedly updated using the attention score to obtain the node embedding feature, and the cosine similarity between the node embedding feature and the fault feature in the preset fault mode library is calculated to obtain the first recall feature;

[0104] In the expert rule recall channel based on causal reasoning, historical operation data is collected and conditional independence test is performed to obtain the degree of dependence between feature pairs. An initial directed graph is constructed based on the degree of dependence. The causal intervention method is used to analyze the influence of features on the occurrence of faults to obtain the intervention effect. The intervention effect is used as the edge weight to construct a weighted directed causal graph.

[0105] Input the state feature vector into a weighted directed causal graph, calculate the causal effect strength between features based on the edge weights in the weighted directed causal graph, select key features according to the causal effect strength to construct a causal chain to obtain a causal feature, calculate the matching degree and confidence of the causal feature and the rule feature in the expert rule base to obtain a second recall feature;

[0106] In the historical case recall channel based on deep reinforcement learning, a multi-layer neural network structure is designed, wherein the number of input layer nodes is set to the state feature dimension, the number of output layer nodes is set to the number of retrieval actions, and the multi-layer neural network structure is pre-trained using historical retrieval data to obtain a deep Q network;

[0107] Input the state feature vector into the deep Q network to obtain the state action value of the retrieval action, perform feature matching in the historical case library according to the state action value to obtain the retrieval result, form the retrieval result and the state feature vector into a training sample to update the parameters of the deep Q network to obtain an optimized network, and input the state feature vector into the optimized network to obtain a third recall feature;

[0108] The first recall feature, the second recall feature and the third recall feature are respectively linearly transformed to obtain the corresponding query vector, key vector and value vector, the dot product of each group of query vectors and the key vector is calculated to obtain the attention score, the attention score is normalized to obtain the attention weight, the first recall feature, the second recall feature and the third recall feature are respectively multiplied by the corresponding attention weight and superimposed to obtain the fused feature.

[0109] Exemplarily, first, a fault knowledge graph is constructed. The graph contains equipment component nodes, fault mode nodes, fault symptom nodes, and operating parameter nodes, as well as the associations between them. For example, a knowledge graph of a pump may contain nodes such as "pump shaft", "vibration", "flow drop", "speed", and the association between "pump shaft" and "vibration". Then, the state feature vector of the current device is input into the fault knowledge graph. Assuming that the state feature vector is [temperature increase, pressure drop, current fluctuation], these features are used as initial node features. Next, the attention scores between adjacent nodes are calculated using the graph attention network. For example, the attention scores of the "temperature increase" node and the "pump shaft" node can be obtained by calculating their feature similarity. The node features are weighted updated using the attention score to obtain the node embedding feature. For example, the feature of the "temperature increase" node is weighted summed with the node features adjacent to it (such as "pump shaft" and "pressure drop") to obtain the updated "temperature increase" node embedding feature. Finally, the cosine similarity between the node embedding feature and the fault feature in the preset fault mode library is calculated to obtain the first recall feature of the fault mode recall channel based on the graph neural network. Assuming that the fault mode library contains fault features such as "pump shaft wear" and "motor failure", the similarity between the updated "temperature rise" node embedding feature and these fault features is calculated to obtain the first recall feature, for example [0.8, 0.2], which means that the similarity with "pump shaft wear" is 0.8 and the similarity with "motor failure" is 0.2.

[0110] Secondly, a weighted directed causal graph is constructed. First, historical operation data is collected, such as temperature, pressure, current and other data at different time points. Then, a conditional independence test is performed to obtain the degree of dependence between feature pairs. For example, the correlation coefficient of temperature and pressure can be calculated to represent the degree of dependence between them. An initial directed graph is constructed based on the degree of dependence. If temperature and pressure are highly correlated, a directed edge is established between them. The causal intervention method is used to analyze the impact of features on the occurrence of faults and obtain the intervention effect. For example, the impact of temperature increase on the probability of fault occurrence can be observed by simulating temperature increase, thereby quantifying the intervention effect of temperature on faults. The intervention effect is used as the edge weight to construct a weighted directed causal graph. If the temperature increase has a greater impact on the occurrence of faults, the weight of the edge from temperature to fault is set to a larger value. The state feature vector is input into the weighted directed causal graph, and the causal effect strength between features is calculated based on the edge weight. The causal effect strength of temperature, pressure and current on faults is calculated based on the current values ​​of temperature, pressure and current and their connections and edge weights in the causal graph. According to the causal effect strength, key features are selected to construct a causal chain to obtain causal features. If the causal effect strength of temperature and current is high, they are used as key features to construct the causal chain "temperature rise -> current fluctuation -> fault". The matching degree and confidence of the causal feature and the rule feature in the expert rule base are calculated to obtain the second recall feature of the expert rule recall channel based on causal reasoning. For example, if there is a rule in the expert rule base that "temperature rise and current fluctuation may be a motor fault", the matching degree and confidence of the current causal feature and the rule are calculated to obtain the second recall feature, such as [0.7, 0.9], which means the matching degree is 0.7 and the confidence is 0.9.

[0111] Secondly, a historical case recall channel based on deep reinforcement learning is constructed. A multi-layer neural network structure is designed, and the number of input layer nodes is set to the state feature dimension, and the number of output layer nodes is set to the number of retrieval actions. For example, if the state feature dimension is 3 (temperature, pressure, current) and the number of retrieval actions is 10 (corresponding to 10 different case categories), the number of input layer nodes is 3 and the number of output layer nodes is 10. The network is pre-trained using historical retrieval data to obtain a deep Q network. The state feature vector is input into the deep Q network to obtain the state action value of the retrieval action. For example, [temperature rises, pressure drops, current fluctuates] is input to obtain 10 state action values, which correspond to the retrieval scores of 10 different case categories. Feature matching is performed in the historical case library based on the state action value to obtain the retrieval result. For example, the case category with the highest score is selected, and the case that is most similar to the current state feature vector is retrieved in this category. The retrieval result and the state feature vector are combined into a training sample to update the parameters of the deep Q network to obtain an optimized network. The state feature vector is input into the optimized network to obtain the third recall feature of the historical case recall channel based on deep reinforcement learning. For example, [temperature increase, pressure drop, current fluctuation] is input into the optimization network to obtain the third recall feature, such as [0.6, 0.5, 0.3], which represents the similarity with the three most similar cases in the historical case library.

[0112] Finally, the three recall features are fused. The first, second, and third recall features are linearly transformed to obtain the corresponding query vector, key vector, and value vector. The dot product of each group of query vectors and key vectors is calculated to obtain the attention score, and the attention score is normalized to obtain the attention weight. The first, second, and third recall features are multiplied by the corresponding attention weights and superimposed to obtain the fused feature. For example, assuming that the attention weights of the first, second, and third recall features are 0.5, 0.3, and 0.2, respectively, the fused feature is 0.5×first recall feature+0.3×second recall feature+0.2×third recall feature.

[0113] In this embodiment, by fusing the information of multiple recall channels, various possible fault modes can be considered more comprehensively, thereby improving the accuracy of fault diagnosis. By adaptively adjusting the weights of each recall channel through the dynamic attention network, the information of different channels can be more effectively utilized, thereby improving the efficiency of fault diagnosis. The causal relationship between features can be analyzed through causal reasoning methods, so as to better understand the mechanism of fault occurrence and enhance the interpretability of fault diagnosis.

[0114] In an optional implementation, a fault propagation map is constructed based on the fusion features, a graph convolutional network is used to extract fault propagation features from the fault propagation map, a fault diagnosis result including fault type information, fault degree information and fault trend information is generated according to the fault propagation features, the fault diagnosis result is input into a hierarchical decision model, a meta-learning method is used to convert the equipment state transition probability matrix into maintenance strategy parameters, and a preliminary maintenance plan is generated according to the maintenance strategy parameters, including:

[0115] Deconstructing the fused features into a node feature matrix and an adjacency matrix, wherein the node feature matrix includes the operation status information, fault type information and component topology relationship information of the equipment components, and the adjacency matrix includes the physical connection relationship and logical dependency relationship between the components;

[0116] Calculating mutual information entropy and conditional mutual information entropy for component pairs in the node feature matrix, determining a fault propagation direction based on the mutual information entropy, determining a propagation strength based on the conditional mutual information entropy, and constructing a fault propagation graph by taking the fault propagation direction as the direction of an edge and taking the propagation strength as the weight of an edge;

[0117] Construct a graph convolution network including a spatial convolution layer and a temporal convolution layer, calculate the attention weight based on the propagation strength in the spatial convolution layer, use the attention weight to weightedly aggregate the features of the neighborhood nodes to obtain spatial features, perform sliding window segmentation on the spatial features in the temporal convolution layer to obtain temporal segments, perform a one-dimensional convolution operation on the temporal segments to obtain local temporal features, concatenate the local temporal features with the spatial features to obtain a single-layer spatiotemporal feature, perform residual connection and feature fusion on each single-layer spatiotemporal feature to obtain a fault propagation feature;

[0118] A multi-task shared encoder and a task-specific decoder are used to analyze the fault propagation characteristics, a classification decoder is used to identify the fault type and fault location, a regression decoder is used to evaluate the fault severity and remaining life, and a sequence decoder is used to predict the fault development trend and propagation path to obtain a fault diagnosis result;

[0119] Based on the fault diagnosis results, a Markov decision process model is constructed, the equipment state is used as the state space, the maintenance operation is used as the action space, a state transition probability matrix is ​​established, and a meta-learning framework including a policy network and a value network is designed. In the inner loop, the policy gradient method is used to optimize the timing selection and method selection parameters of the maintenance strategy based on the state transition probability matrix, and in the outer loop, the Monte Carlo method is used to update the state transition probability based on the actual maintenance effect;

[0120] According to the optimized maintenance strategy parameters, a preliminary maintenance plan is generated in combination with equipment maintenance specifications and resource constraints. The preliminary maintenance plan includes maintenance items, maintenance time and maintenance resource configuration information.

[0121] Exemplarily, first, collect various operating data of equipment components, such as temperature, pressure, current, vibration, etc., as well as the topological structure information of the equipment and the physical connection relationship and logical dependency relationship between components. For example, a power transformer contains oil temperature sensors, oil level sensors, winding temperature sensors, etc. The data of these sensors constitute the operating status information; at the same time, record the component type to which each sensor belongs, such as "oil temperature sensor" belongs to "cooling system"; and record the connection relationship between components, such as the physical connection between the oil pump and the radiator, and the logical dependency relationship between the main transformer and the cooling system. This information together constitutes a node feature matrix, in which each row represents a component and each column represents a feature. At the same time, construct an adjacency matrix, in which both rows and columns of the matrix represent components. If there is a physical connection or logical dependency relationship between two components, the value of the corresponding position is 1, otherwise it is 0.

[0122] Next, the mutual information entropy and conditional mutual information entropy between component pairs are calculated. For example, by calculating the mutual information entropy between the oil temperature sensor and the oil pump, the degree of correlation between the operating state changes of the two can be measured. If the increase in oil temperature is always accompanied by a decrease in the speed of the oil pump, the mutual information entropy between the two is high. Further calculating the conditional mutual information entropy between the oil temperature sensor and the oil pump when the state of the oil level sensor is known can measure the degree of correlation between the oil temperature and the oil pump after excluding the influence of the oil level. By comparing the mutual information entropy and conditional mutual information entropy of different component pairs, the direction and intensity of fault propagation can be determined. For example, if the mutual information entropy between the oil temperature sensor and the oil pump is high, and the conditional mutual information entropy between the oil temperature sensor and the oil pump is still high when the state of the oil level sensor is known, it can be determined that the fault may propagate from the oil temperature sensor to the oil pump. The higher the mutual information entropy, the greater the propagation intensity. The propagation direction is used as the direction of the edge, and the propagation intensity is used as the weight of the edge to construct a fault propagation map.

[0123] Then, a graph convolutional network is constructed, which includes a spatial convolutional layer and a temporal convolutional layer. In the spatial convolutional layer, the attention weight is calculated according to the propagation strength. For example, if the propagation strength from the oil temperature sensor to the oil pump is large, the oil pump will give a greater weight to the oil temperature sensor when aggregating the features of neighboring nodes. The features of the neighboring nodes are weightedly aggregated using the attention weights to obtain spatial features. In the temporal convolutional layer, the spatial features are segmented by sliding windows, for example, the data of the past hour is divided into 12 5-minute segments. A one-dimensional convolution operation is performed on each segment to extract local temporal features. The local temporal features are concatenated with the spatial features to obtain a single-layer spatiotemporal feature. All single-layer spatiotemporal features are residually connected and feature fused to obtain the final fault propagation features.

[0124] Next, the fault propagation features are parsed using a multi-task learning framework. The framework consists of a shared encoder and multiple task-specific decoders. The shared encoder encodes the fault propagation features into a higher-level representation. The classification decoder identifies the fault type and fault location, for example, judging the fault type as "oil temperature is too high" and the fault location as "cooling system". The regression decoder evaluates the fault severity and remaining life, for example, predicting the severity of the oil temperature fault to be "medium" and the remaining life to be "100 hours". The sequence decoder predicts the fault development trend and propagation path, for example, predicting that the oil temperature fault will lead to oil pump failure and gives the specific propagation path. Together, this information constitutes the fault diagnosis result.

[0125] Finally, a Markov decision process model is constructed based on the fault diagnosis results. The equipment state (e.g., normal, minor fault, major fault) is used as the state space, and the maintenance operation (e.g., non-stop maintenance, shutdown maintenance, and component replacement) is used as the action space. A state transition probability matrix is ​​established. For example, if no maintenance is performed, there is an 80% probability that a minor fault will develop into a major fault, and a 20% probability that it will remain unchanged. A meta-learning framework is designed. In the inner loop, the policy gradient method is used to optimize the maintenance strategy parameters according to the state transition probability matrix, such as determining when to perform maintenance and which maintenance method to choose. In the outer loop, the Monte Carlo method is used to update the state transition probability matrix according to the actual maintenance effect, so that the model is closer to the actual situation. Based on the optimized maintenance strategy parameters, combined with the equipment maintenance specifications and resource constraints, a preliminary maintenance plan is generated, such as recommending non-stop maintenance of the cooling system within 10 hours, and allocating corresponding maintenance personnel and spare parts.

[0126] In this embodiment, by constructing a fault propagation graph and applying a graph convolutional network, the correlation between equipment components and the fault propagation law can be captured more comprehensively, thereby improving the accuracy of fault diagnosis. At the same time, automated feature extraction and fault diagnosis processes can significantly improve the efficiency of fault diagnosis. By constructing a Markov decision process model and applying a meta-learning method, the optimal maintenance strategy can be automatically generated according to the equipment status, fault diagnosis results and resource constraints, thereby avoiding unnecessary maintenance or delayed maintenance and improving the accuracy of maintenance decisions. It can reduce unnecessary maintenance while ensuring the safe and reliable operation of the equipment, thereby reducing maintenance costs and spare parts consumption, and ultimately reducing the overall operation and maintenance costs of the equipment.

[0127] In an optional implementation, the preliminary maintenance plan and the fault propagation graph are input into a case matching model, the case matching model uses a graph attention mechanism to extract a repair path from historical maintenance cases, and uses a temporal graph neural network to model the dependency relationship of the maintenance steps in the repair path to obtain a step dependency model, including:

[0128] The preliminary maintenance plan and the fault propagation graph are input into a case matching model, and the case matching model performs graph matching through a multi-layer graph attention network, wherein the node matching layer calculates the matching degree of the component fault feature vector through cosine similarity to obtain a node similarity matrix, and the edge matching layer calculates the edge matching degree through the Euclidean distance of the propagation feature vector to obtain an edge similarity matrix;

[0129] Based on the node similarity matrix and the edge similarity matrix, a multi-head attention mechanism including fault type, fault degree, propagation direction and propagation intensity is used to calculate the similarity score of each dimension, and the similarity scores of each dimension are fused through adaptive weights to obtain a comprehensive similarity vector.

[0130] Retrieving a similar case set from a historical case library based on the comprehensive similarity vector, extracting a maintenance operation sequence from the similar case set to form a repair path library, wherein each repair path in the repair path library includes a maintenance step identifier, an execution sequence, resource configuration, and time information;

[0131] A bidirectional temporal graph neural network is constructed, and the maintenance step identifiers in the repair path library are mapped to graph nodes. In the forward propagation, a gated recurrent unit is used to extract the sequential dependency features of the maintenance steps. In the backward propagation, an attention mechanism is used to capture the long-range dependency relationship of the maintenance steps. The sequential dependency features and the long-range dependency relationship are concatenated and then the step dependency features are obtained through the graph attention layer.

[0132] A message passing framework is constructed based on the step dependency features to generate a node message including the step completion status, resource occupancy status, and associated step status, and the node message is iteratively updated multiple times through a message aggregation function to obtain a target node representation;

[0133] A graph convolutional network is used to extract features from the target node representation, and a step dependency model including explicit timing constraints and implicit resource constraints is constructed, wherein the explicit timing constraints represent the prerequisite relationship and parallel relationship of the maintenance steps, and the implicit resource constraints represent the resource exclusive relationship of the maintenance steps.

[0134] Exemplarily, first, the fault information of the equipment is collected and preprocessed. The fault type, fault component, fault degree and other information of the equipment to be repaired, as well as the propagation path of the fault in the equipment system, are collected to construct a fault propagation map. For example, a power transformer fails, the fault type is insulation fault, the fault component is bushing, the fault degree is serious, and the fault propagation path is bushing-high voltage winding-low voltage winding. This information is converted into a structured data format, for example, a graph data structure is used to represent the fault propagation map, where nodes represent equipment components and edges represent the direction and intensity of fault propagation.

[0135] Then, a preliminary maintenance plan is developed. Based on the equipment failure information and expert experience, the scope and steps of the maintenance operation are preliminarily determined. For example, for the above transformer failure, the preliminary maintenance plan may include replacing the bushing, checking the insulation status of the high-voltage winding and the low-voltage winding, etc.

[0136] Next, case matching is performed. The preliminary maintenance plan and fault propagation graph are input into the case matching model. The model uses the graph attention mechanism to extract the repair path from the historical maintenance cases. Specifically, the model first calculates the matching degree of the component fault feature vector to obtain the node similarity matrix. Then the Euclidean distance of the propagation feature vector is calculated to obtain the edge similarity matrix. Next, the model uses a multi-head attention mechanism including fault type, fault degree, propagation direction and propagation strength to calculate the similarity score of each dimension, and fuses the similarity scores of each dimension through adaptive weights to obtain a comprehensive similarity vector. For example, the similarity between the current case and the historical case A in fault type is 0.9, the similarity in fault degree is 0.8, the similarity in propagation direction is 0.7, and the similarity in propagation strength is 0.6, then the comprehensive similarity may be 0.85. Finally, based on the comprehensive similarity vector, a collection of similar cases is retrieved from the historical case library, for example, the top 5 cases in comprehensive similarity are selected.

[0137] After that, the maintenance operation sequence is extracted from the collection of similar cases to form a repair path library. Each repair path in the repair path library contains the maintenance step identification, execution order, resource configuration and time information. For example, the repair path of historical case A is: Step 1: Dismantle the faulty bushing (requires a wrench and a crane, and is expected to take 2 hours); Step 2: Install a new bushing (requires a wrench and a crane, and is expected to take 2 hours); Step 3: Test the insulation performance (requires an insulation tester, and is expected to take 1 hour).

[0138] Then, a step dependency model is constructed. A bidirectional temporal graph neural network is used to model the dependency of the maintenance steps in the repair path library. The maintenance step identifiers are mapped to graph nodes. In the forward propagation, a gated recurrent unit is used to extract the sequential dependency features of the maintenance steps. In the backward propagation, an attention mechanism is used to capture the long-range dependencies of the maintenance steps. After concatenating the sequential dependency features and the long-range dependencies, the step dependency features are obtained through the graph attention layer. For example, there is a strong sequential dependency between step 1 and step 2, and a weak sequential dependency between step 2 and step 3.

[0139] Next, a message passing framework is constructed based on the step dependency feature. A node message containing the step completion status, resource occupation status, and associated step status is generated. For example, step 1 has been completed, the wrench and crane are occupied, and the associated step 2 has not yet started. The node message is iteratively updated multiple times through the message aggregation function to obtain the target node representation.

[0140] Finally, a graph convolutional network is used to extract features from the target node representation, and a step dependency model containing explicit timing constraints and implicit resource constraints is constructed. Among them, the explicit timing constraints represent the pre-relationship and parallel relationship of the maintenance steps. For example, step 1 must be completed before step 2, and step 2 and step 3 can be executed in parallel. Implicit resource constraints represent the resource exclusive relationship of the maintenance steps. For example, both step 1 and step 2 require the use of wrenches and cranes, so they cannot be executed at the same time.

[0141] In this embodiment, through case matching and step dependency modeling, the best maintenance plan can be quickly determined to reduce maintenance time and cost. By drawing on the experience of historical cases, repeated errors can be avoided and maintenance quality can be improved. By considering the resource constraints of maintenance steps, resource allocation can be optimized and resource utilization can be improved.

[0142] In an optional implementation, the fault diagnosis result is input into a step-dependent model as a constraint condition, and a hierarchical reinforcement learning method is used to optimize the maintenance plan, wherein the maintenance strategy parameters are used in the maintenance strategy optimization submodule to plan the maintenance step sequence, and the maintenance parameters are adjusted according to the fault degree information in the maintenance parameter optimization submodule. Generating an optimal maintenance plan through a multi-objective optimization algorithm includes:

[0143] The fault diagnosis result is input as a constraint condition into a step dependency model, wherein the step dependency model includes a maintenance step node and dependency edges between steps, wherein the maintenance step node includes a resource requirement vector, an execution duration, and a priority weight attribute, and the dependency edge includes a preceding relationship identifier and a parallel relationship identifier, and the priority weight of the maintenance step node is updated according to the fault diagnosis result, and a priority execution constraint is set for the relevant maintenance steps based on the fault diagnosis result;

[0144] Constructing a maintenance strategy optimization objective function, wherein the maintenance strategy optimization objective function includes a total maintenance time item, a maintenance cost item, a residual risk item, and is subject to a preceding relationship identifier constraint of the dependency relationship edge between the steps, a resource capacity constraint, a time window constraint, and the priority execution constraint;

[0145] A mixed integer programming method is used to solve the maintenance strategy optimization objective function to obtain an initial maintenance step execution sequence and an initial resource allocation plan;

[0146] Constructing a maintenance parameter optimization objective function, wherein the maintenance parameter optimization objective function includes a maintenance quality index item, a step execution time item, and a step cost item;

[0147] Based on the fault degree information in the fault diagnosis result, a gradient-based multi-objective optimization algorithm is used to solve the maintenance parameter optimization objective function to obtain maintenance depth parameters, maintenance process parameters, detection parameters and quality control parameters;

[0148] The actual execution time is calculated according to the maintenance parameters, and a deviation analysis is performed between the actual execution time and the planned time in the initial maintenance step execution sequence. Based on the deviation analysis results, the initial maintenance step execution sequence is adjusted, and the maintenance parameters of the affected steps are triggered to be re-optimized until the optimization results of the maintenance strategy optimization objective function and the maintenance parameter optimization objective function converge. According to the converged optimization results, an optimal maintenance plan including step execution timing, parameter configuration plan, resource scheduling plan and quality control requirements is generated.

[0149] Exemplarily, first, a fault diagnosis result of the device is obtained, which includes information such as fault type, fault location, fault degree, etc. For example, the diagnosis result shows that the motor bearing of device A is worn, and the degree of wear is moderate.

[0150] Then, the fault diagnosis results are input into the pre-established step dependency model. The model describes all the steps in the maintenance process and their dependencies. Each step node contains attributes such as resource requirement vector (for example, 2 technicians and 1 crane are required), execution time (for example, 2 hours), priority weight, etc. The dependency edge between steps contains the predecessor relationship identifier (for example, step A must be completed before step B) and the parallel relationship identifier (for example, step C and step D can be performed at the same time). According to the fault diagnosis results, the priority weight of the step node is updated. For example, due to the wear of the motor bearing, the priority weight of the step related to the bearing replacement is increased. At the same time, according to the diagnosis results, priority execution constraints are set for the relevant steps, for example, the bearing must be inspected before the bearing can be replaced.

[0151] Next, the maintenance strategy optimization objective function is constructed. This objective function aims to minimize the total maintenance time, maintenance cost, and residual risk. The objective function is subject to the dependencies between steps, resource capacity constraints (e.g., only 1 crane is available), time window constraints (e.g., all maintenance must be completed within 8 hours), and priority execution constraints set according to the diagnosis results. The objective function is solved using the mixed integer programming method to obtain the initial maintenance step execution order and the initial resource allocation plan. For example, the initial plan is: first perform bearing inspection (1 hour, 2 technicians), and then perform bearing replacement (3 hours, 2 technicians, 1 crane).

[0152] Subsequently, an overhaul parameter optimization objective function is constructed. This objective function aims to maximize the overhaul quality index while minimizing the step execution time and step cost. The objective function considers overhaul depth parameters (e.g., replacement of bearings or lubrication only), overhaul process parameters (e.g., which welding process to use), inspection parameters (e.g., which nondestructive testing method to use), and quality control parameters (e.g., quality inspection standards).

[0153] According to the fault degree information (such as moderate wear) in the fault diagnosis results, a gradient-based multi-objective optimization algorithm is used to solve the maintenance parameter optimization objective function and obtain the specific parameter configuration scheme for each step. For example, since the wear degree is moderate, the bearing is replaced and a specific type of bearing and corresponding installation process are used.

[0154] According to the optimized maintenance parameters, calculate the actual execution time of each step. Perform deviation analysis on the actual execution time and the planned time in the initial maintenance step execution sequence. For example, the actual execution time of the bearing replacement is 3.5 hours, which is 0.5 hours more than the originally planned 3 hours. Based on the deviation analysis results, adjust the initial maintenance step execution sequence, and trigger re-optimization of the maintenance parameters of the affected steps. For example, due to the extension of the bearing replacement time, the start time of the subsequent steps needs to be adjusted accordingly, and re-evaluate whether the parameter configuration of the subsequent steps needs to be adjusted. Repeat this process until the optimization results of the maintenance strategy optimization objective function and the maintenance parameter optimization objective function converge.

[0155] Finally, based on the converged optimization results, the optimal maintenance plan including step execution timing, parameter configuration plan, resource scheduling plan and quality control requirements is generated. For example, the final plan is: first perform bearing inspection (1 hour, 2 technicians), then perform bearing replacement (3.5 hours, 2 technicians, 1 crane), and finally perform equipment testing (0.5 hours, 1 technician).

[0156] In this embodiment, by optimizing the execution sequence of maintenance steps and resource allocation scheme, the total maintenance time can be shortened and the maintenance efficiency can be improved. By optimizing the maintenance parameter configuration scheme, the maintenance cost can be reduced, such as selecting more economical maintenance materials and processes. By considering the fault diagnosis results and optimizing the maintenance depth parameters, the maintenance quality can be improved and the risk of equipment failure can be reduced.

[0157] Figure 2 FIG. 1 is a schematic diagram of a structure of a system for intelligent auxiliary diagnosis and maintenance based on multi-channel recall according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0158] The first unit is used to collect equipment operation data and perform adaptive filtering and noise reduction processing to obtain filtered data, extract time series features from the filtered data using a recursive neural network, enhance the time series features through an adversarial learning network to obtain enhanced features, construct a feature dependency graph of the enhanced features based on a causal reasoning method, establish a device state transition probability matrix according to the feature dependency graph, and combine the enhanced features with the device state transition probability matrix to generate a state feature vector;

[0159] The second unit is used to input the state feature vector into a multi-channel recall diagnosis network for fault feature recall, wherein the multi-channel recall diagnosis network includes a fault mode recall channel based on a graph neural network, an expert rule recall channel based on causal reasoning, and a historical case recall channel based on deep reinforcement learning. The weight coefficient of the output results of each recall channel is calculated by a dynamic attention network and fused to obtain a fusion feature, a fault propagation map is constructed based on the fusion feature, a graph convolutional network is used to extract fault propagation features from the fault propagation map, a fault diagnosis result including fault type information, fault degree information and fault trend information is generated according to the fault propagation feature, the fault diagnosis result is input into a hierarchical decision model, a meta-learning method is used to convert the equipment state transition probability matrix into maintenance strategy parameters, and a preliminary maintenance plan is generated according to the maintenance strategy parameters;

[0160] The third unit is used to input the preliminary maintenance plan and the fault propagation map into a case matching model, the case matching model uses a graph attention mechanism to extract repair paths from historical maintenance cases, and uses a temporal graph neural network to model the dependency relationship of the maintenance steps in the repair path to obtain a step dependency model, and inputs the fault diagnosis result as a constraint condition into the step dependency model, and uses a hierarchical reinforcement learning method to optimize the maintenance plan, wherein the maintenance strategy parameters are used to plan the maintenance step sequence in the maintenance strategy optimization submodule, and the maintenance parameters are adjusted according to the fault degree information in the maintenance parameter optimization submodule, and the optimal maintenance plan is generated through a multi-objective optimization algorithm.

[0161] According to a third aspect of the embodiments of the present invention,

[0162] An electronic device is provided, comprising:

[0163] processor;

[0164] a memory for storing processor-executable instructions;

[0165] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0166] A fourth aspect of the embodiments of the present invention is:

[0167] 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.

[0168] 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.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent auxiliary diagnosis and maintenance based on multi-channel recall, characterized in that: include: Collect equipment operation data and perform adaptive filtering and noise reduction processing to obtain filtered data, use a recursive neural network to extract time series features from the filtered data, enhance the time series features through an adversarial learning network to obtain enhanced features, construct a feature dependency graph of the enhanced features based on a causal reasoning method, establish a device state transition probability matrix based on the feature dependency graph, and combine the enhanced features with the device state transition probability matrix to generate a state feature vector; The state feature vector is input into a multi-channel recall diagnosis network for fault feature recall, wherein the multi-channel recall diagnosis network includes a fault mode recall channel based on a graph neural network, an expert rule recall channel based on causal reasoning, and a historical case recall channel based on deep reinforcement learning. The weight coefficient of the output results of each recall channel is calculated through a dynamic attention network and fused to obtain a fusion feature. A fault propagation map is constructed based on the fusion feature, and a graph convolutional network is used to extract fault propagation features from the fault propagation map. A fault diagnosis result including fault type information, fault degree information, and fault trend information is generated according to the fault propagation feature. The fault diagnosis result is input into a hierarchical decision model, and a meta-learning method is used to convert the equipment state transition probability matrix into a maintenance strategy parameter, and a preliminary maintenance plan is generated according to the maintenance strategy parameter. The preliminary maintenance plan and the fault propagation graph are input into a case matching model. The case matching model uses a graph attention mechanism to extract repair paths from historical maintenance cases, and uses a temporal graph neural network to model the dependency relationship of the maintenance steps in the repair path to obtain a step dependency model. The fault diagnosis result is input into the step dependency model as a constraint condition, and a hierarchical reinforcement learning method is used to optimize the maintenance plan. In the maintenance strategy optimization submodule, the maintenance strategy parameters are used to plan the sequence of maintenance steps, and in the maintenance parameter optimization submodule, the maintenance parameters are adjusted according to the fault degree information, and the optimal maintenance plan is generated through a multi-objective optimization algorithm.

2. The method according to claim 1, characterized in that The equipment operation data is collected and adaptively filtered and denoised to obtain filtered data, a recursive neural network is used to extract time series features from the filtered data, and an adversarial learning network is used to enhance the time series features to obtain enhanced features including: The vibration data, temperature data, pressure data and current data of the equipment are collected through the sensor array to obtain the equipment operation data, a coupling relationship matrix between the measurement variables is established based on the physical model of the equipment, the equipment operation data is input into the encoder of the variational autoencoder to obtain the latent variable distribution parameters, the latent variable distribution parameters are input into the decoder for data reconstruction, and the reconstruction error is optimized by maximizing the lower bound of evidence to obtain the reconstructed data; Performing spectrum analysis on the reconstructed data to obtain frequency band energy distribution, calculating a signal-to-noise ratio evaluation value based on the frequency band energy distribution, jointly optimizing center frequency parameters, bandwidth parameters, and filter order parameters of low-pass, band-pass, and high-pass filters through an adaptive optimization algorithm using the signal-to-noise ratio evaluation value, updating a multiscale filter group according to the optimized parameters, and performing adaptive filtering on the reconstructed data to obtain filtered data; Input the filtered data into a bidirectional gated recurrent unit, calculate the correlation matrix of the filtered data at different time steps to obtain the time dimension attention weight, use spatial position encoding to calculate the spatial dimension attention weight, recalibrate different channel features of the filtered data to obtain channel recalibration weights, adaptively fuse the time dimension attention weights, spatial dimension attention weights and channel recalibration weights to obtain comprehensive attention weights, and use the comprehensive attention weights to weight the filtered data to obtain time series features; The time series features are input into the feature extractor of the conditional adversarial energy network to obtain intermediate features, the intermediate features are passed through the discriminator of the residual network structure to obtain discrimination results, positive and negative sample pairs are constructed based on the discrimination results, the network is trained by minimizing the energy difference of the positive sample pairs and maximizing the energy difference of the negative sample pairs, mutual information regularization terms and gradient penalty terms are introduced in the training process, and the trained conditional adversarial energy network is used to perform domain adaptation enhancement processing on the time series features to obtain enhanced features.

3. The method according to claim 1, characterized in that Constructing a feature dependency graph of the enhanced feature based on a causal reasoning method, establishing a device state transition probability matrix according to the feature dependency graph, and combining the enhanced feature with the device state transition probability matrix to generate a state feature vector includes: Calculating conditional mutual information values ​​between feature pairs for enhanced features to obtain a feature correlation matrix, binarizing the feature correlation matrix based on a preset binarization threshold to obtain an adjacency matrix, and constructing an undirected graph structure according to the adjacency matrix; Calculating a conditional independence score based on observed data for each node pair in the undirected graph structure, identifying a separation set in the undirected graph structure based on the conditional independence score, using the separation set to determine the causal relationship between nodes, identifying non-connected nodes with common child nodes as parent nodes of corresponding child nodes to obtain partial directed edges, and using a minimum directed rule to orient the remaining undirected edges to obtain a directed graph structure; Inputting the directed graph structure into a graph attention network, generating a query vector, a key vector and a value vector for each node in the directed graph structure through a multi-layer perceptron, calculating the dot product of the query vector and the key vector to obtain an attention score, normalizing the attention score to obtain an attention weight, taking the product of the attention weight and the value vector as the causal strength between nodes, and constructing a weighted directed graph based on the causal strength; Input the weighted directed graph into a dynamic Bayesian network, construct a conditional probability distribution for each time step, use a variational inference method to approximate the conditional probability distribution to a Gaussian distribution, optimize the mean and variance parameters of the Gaussian distribution by minimizing the variational free energy, calculate the characteristic transition probabilities between adjacent time steps based on the optimized distribution parameters, and combine the characteristic transition probabilities into a device state migration probability matrix; The enhanced features and the device state transition probability matrix are projected onto a common feature space through feature mapping to obtain projection features, the correlation between the projection features is calculated to obtain modal correlation weights, and the projection features are weightedly fused based on the modal correlation weights to obtain a state feature vector.

4. The method according to claim 1, characterized in that: The state feature vector is input into a multi-channel recall diagnosis network for fault feature recall. The multi-channel recall diagnosis network includes a fault mode recall channel based on a graph neural network, an expert rule recall channel based on causal reasoning, and a historical case recall channel based on deep reinforcement learning. The weight coefficient of the output result of each recall channel is calculated through a dynamic attention network and fused to obtain a fusion feature including: In the fault mode recall channel of the graph neural network, a fault knowledge graph is constructed, the state feature vector is input into the fault knowledge graph, node embedding is performed based on the association relationship between the equipment component node, the fault mode node, the fault symptom node and the operating parameter node in the fault knowledge graph to obtain the node feature, the attention score between adjacent nodes is calculated by the graph attention network, the node feature is weightedly updated using the attention score to obtain the node embedding feature, and the cosine similarity between the node embedding feature and the fault feature in the preset fault mode library is calculated to obtain the first recall feature; In the expert rule recall channel based on causal reasoning, historical operation data is collected and conditional independence test is performed to obtain the degree of dependence between feature pairs. An initial directed graph is constructed based on the degree of dependence. The causal intervention method is used to analyze the influence of features on the occurrence of faults to obtain the intervention effect. The intervention effect is used as the edge weight to construct a weighted directed causal graph. Input the state feature vector into a weighted directed causal graph, calculate the causal effect strength between features based on the edge weights in the weighted directed causal graph, select key features according to the causal effect strength to construct a causal chain to obtain a causal feature, calculate the matching degree and confidence of the causal feature and the rule feature in the expert rule base to obtain a second recall feature; In the historical case recall channel based on deep reinforcement learning, a multi-layer neural network structure is designed, wherein the number of input layer nodes is set to the state feature dimension, the number of output layer nodes is set to the number of retrieval actions, and the multi-layer neural network structure is pre-trained using historical retrieval data to obtain a deep Q network; Input the state feature vector into the deep Q network to obtain the state action value of the retrieval action, perform feature matching in the historical case library according to the state action value to obtain the retrieval result, form the retrieval result and the state feature vector into a training sample to update the parameters of the deep Q network to obtain an optimized network, and input the state feature vector into the optimized network to obtain a third recall feature; The first recall feature, the second recall feature and the third recall feature are respectively linearly transformed to obtain the corresponding query vector, key vector and value vector, the dot product of each group of query vectors and the key vector is calculated to obtain the attention score, the attention score is normalized to obtain the attention weight, the first recall feature, the second recall feature and the third recall feature are respectively multiplied by the corresponding attention weight and superimposed to obtain the fused feature.

5. The method according to claim 1, characterized in that A fault propagation map is constructed based on the fusion features, a graph convolutional network is used to extract fault propagation features from the fault propagation map, a fault diagnosis result including fault type information, fault degree information and fault trend information is generated according to the fault propagation features, the fault diagnosis result is input into a hierarchical decision model, a meta-learning method is used to convert the equipment state transition probability matrix into maintenance strategy parameters, and a preliminary maintenance plan is generated according to the maintenance strategy parameters, including: Deconstructing the fused features into a node feature matrix and an adjacency matrix, wherein the node feature matrix includes the operation status information, fault type information and component topology relationship information of the equipment components, and the adjacency matrix includes the physical connection relationship and logical dependency relationship between the components; Calculating mutual information entropy and conditional mutual information entropy for component pairs in the node feature matrix, determining a fault propagation direction based on the mutual information entropy, determining a propagation strength based on the conditional mutual information entropy, and constructing a fault propagation graph by taking the fault propagation direction as the direction of an edge and taking the propagation strength as the weight of an edge; Construct a graph convolution network including a spatial convolution layer and a temporal convolution layer, calculate the attention weight based on the propagation strength in the spatial convolution layer, use the attention weight to weightedly aggregate the features of the neighborhood nodes to obtain spatial features, perform sliding window segmentation on the spatial features in the temporal convolution layer to obtain temporal segments, perform a one-dimensional convolution operation on the temporal segments to obtain local temporal features, concatenate the local temporal features with the spatial features to obtain a single-layer spatiotemporal feature, perform residual connection and feature fusion on each single-layer spatiotemporal feature to obtain a fault propagation feature; A multi-task shared encoder and a task-specific decoder are used to analyze the fault propagation characteristics, a classification decoder is used to identify the fault type and fault location, a regression decoder is used to evaluate the fault severity and remaining life, and a sequence decoder is used to predict the fault development trend and propagation path to obtain a fault diagnosis result; Based on the fault diagnosis results, a Markov decision process model is constructed, the equipment state is used as the state space, the maintenance operation is used as the action space, a state transition probability matrix is ​​established, and a meta-learning framework including a policy network and a value network is designed. In the inner loop, the policy gradient method is used to optimize the timing selection and method selection parameters of the maintenance strategy based on the state transition probability matrix, and in the outer loop, the Monte Carlo method is used to update the state transition probability based on the actual maintenance effect; According to the optimized maintenance strategy parameters, a preliminary maintenance plan is generated in combination with equipment maintenance specifications and resource constraints. The preliminary maintenance plan includes maintenance items, maintenance time and maintenance resource configuration information.

6. The method according to claim 1, characterized in that The preliminary maintenance plan and the fault propagation graph are input into a case matching model. The case matching model uses a graph attention mechanism to extract a repair path from historical maintenance cases, and uses a temporal graph neural network to model the dependency relationship of the maintenance steps in the repair path to obtain a step dependency model including: The preliminary maintenance plan and the fault propagation graph are input into a case matching model, and the case matching model performs graph matching through a multi-layer graph attention network, wherein the node matching layer calculates the matching degree of the component fault feature vector by cosine similarity to obtain a node similarity matrix, and the edge matching layer calculates the edge matching degree by the Euclidean distance of the propagation feature vector to obtain an edge similarity matrix; Based on the node similarity matrix and the edge similarity matrix, a multi-head attention mechanism including fault type, fault degree, propagation direction and propagation intensity is used to calculate the similarity score of each dimension, and the similarity scores of each dimension are fused through adaptive weights to obtain a comprehensive similarity vector. Retrieving a similar case set from a historical case library based on the comprehensive similarity vector, extracting a maintenance operation sequence from the similar case set to form a repair path library, wherein each repair path in the repair path library includes a maintenance step identifier, an execution sequence, resource configuration, and time information; A bidirectional temporal graph neural network is constructed, and the maintenance step identifiers in the repair path library are mapped to graph nodes. In the forward propagation, a gated recurrent unit is used to extract the sequential dependency features of the maintenance steps. In the backward propagation, an attention mechanism is used to capture the long-range dependency relationship of the maintenance steps. The sequential dependency features and the long-range dependency relationship are concatenated and then the step dependency features are obtained through the graph attention layer. A message passing framework is constructed based on the step dependency features to generate a node message including the step completion status, resource occupancy status, and associated step status, and the node message is iteratively updated multiple times through a message aggregation function to obtain a target node representation; A graph convolutional network is used to extract features from the target node representation, and a step dependency model including explicit timing constraints and implicit resource constraints is constructed, wherein the explicit timing constraints represent the prerequisite relationship and parallel relationship of the maintenance steps, and the implicit resource constraints represent the resource exclusive relationship of the maintenance steps.

7. The method according to claim 1, characterized in that The fault diagnosis result is input as a constraint condition into a step-dependent model, and a hierarchical reinforcement learning method is used to optimize the maintenance plan, wherein the maintenance strategy parameters are used in the maintenance strategy optimization submodule to plan the maintenance step sequence, and the maintenance parameters are adjusted according to the fault degree information in the maintenance parameter optimization submodule. The optimal maintenance plan is generated by a multi-objective optimization algorithm, including: The fault diagnosis result is input as a constraint condition into a step dependency model, wherein the step dependency model includes a maintenance step node and dependency edges between steps, wherein the maintenance step node includes a resource requirement vector, an execution duration, and a priority weight attribute, and the dependency edge includes a preceding relationship identifier and a parallel relationship identifier, and the priority weight of the maintenance step node is updated according to the fault diagnosis result, and a priority execution constraint is set for the relevant maintenance steps based on the fault diagnosis result; Constructing a maintenance strategy optimization objective function, wherein the maintenance strategy optimization objective function includes a total maintenance time item, a maintenance cost item, a residual risk item, and is subject to a preceding relationship identifier constraint of the dependency relationship edge between the steps, a resource capacity constraint, a time window constraint, and the priority execution constraint; A mixed integer programming method is used to solve the maintenance strategy optimization objective function to obtain an initial maintenance step execution sequence and an initial resource allocation plan; Constructing a maintenance parameter optimization objective function, wherein the maintenance parameter optimization objective function includes a maintenance quality index item, a step execution time item, and a step cost item; Based on the fault degree information in the fault diagnosis result, a gradient-based multi-objective optimization algorithm is used to solve the maintenance parameter optimization objective function to obtain maintenance depth parameters, maintenance process parameters, detection parameters and quality control parameters; The actual execution time is calculated according to the maintenance parameters, and a deviation analysis is performed between the actual execution time and the planned time in the initial maintenance step execution sequence. Based on the deviation analysis results, the initial maintenance step execution sequence is adjusted, and the maintenance parameters of the affected steps are triggered to be re-optimized until the optimization results of the maintenance strategy optimization objective function and the maintenance parameter optimization objective function converge. According to the converged optimization results, an optimal maintenance plan including step execution timing, parameter configuration plan, resource scheduling plan and quality control requirements is generated.

8. A system for intelligent auxiliary diagnosis and maintenance based on multi-channel recall, 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 equipment operation data and perform adaptive filtering and noise reduction processing to obtain filtered data, extract time series features from the filtered data using a recursive neural network, enhance the time series features through an adversarial learning network to obtain enhanced features, construct a feature dependency graph of the enhanced features based on a causal reasoning method, establish a device state transition probability matrix according to the feature dependency graph, and combine the enhanced features with the device state transition probability matrix to generate a state feature vector; The second unit is used to input the state feature vector into a multi-channel recall diagnosis network for fault feature recall, wherein the multi-channel recall diagnosis network includes a fault mode recall channel based on a graph neural network, an expert rule recall channel based on causal reasoning, and a historical case recall channel based on deep reinforcement learning. The weight coefficient of the output results of each recall channel is calculated by a dynamic attention network and fused to obtain a fusion feature, a fault propagation map is constructed based on the fusion feature, a graph convolutional network is used to extract fault propagation features from the fault propagation map, a fault diagnosis result including fault type information, fault degree information and fault trend information is generated according to the fault propagation feature, the fault diagnosis result is input into a hierarchical decision model, a meta-learning method is used to convert the equipment state transition probability matrix into maintenance strategy parameters, and a preliminary maintenance plan is generated according to the maintenance strategy parameters; The third unit is used to input the preliminary maintenance plan and the fault propagation map into a case matching model, the case matching model uses a graph attention mechanism to extract repair paths from historical maintenance cases, and uses a temporal graph neural network to model the dependency relationship of the maintenance steps in the repair path to obtain a step dependency model, and inputs the fault diagnosis result as a constraint condition into the step dependency model, and uses a hierarchical reinforcement learning method to optimize the maintenance plan, wherein the maintenance strategy parameters are used to plan the maintenance step sequence in the maintenance strategy optimization submodule, and the maintenance parameters are adjusted according to the fault degree information in the maintenance parameter optimization submodule, and the optimal maintenance plan is generated through a multi-objective optimization algorithm.

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