Power generation equipment anomaly prediction and adaptive maintenance system based on fusion deep learning

Abstract

This invention relates to the field of deep learning technology, specifically to a power generation equipment anomaly prediction and adaptive maintenance system based on fused deep learning. The system first acquires multimodal local and global data through a multimodal acquisition unit; then, a first neural network processing unit inputs the multimodal local data into a knowledge-embedded hybrid neural network model to obtain local anomaly predictions; a second neural network processing unit inputs the multimodal global data into an encoder-decoder network based on cross-modal attention to obtain global anomaly predictions; a reinforcement learning processing unit optimizes the maintenance strategy using a reinforcement learning model and performs hyperparameter optimization on the reinforcement learning model using a genetic algorithm, thus achieving model interpretability; finally, an output unit provides a comprehensive anomaly prediction result and an updated maintenance strategy optimized by reinforcement learning and genetic algorithms to improve the effectiveness of anomaly prediction and maintenance.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically to a power generation equipment anomaly prediction and adaptive maintenance system based on fused deep learning. Background Technology

[0002] In recent years, with the rapid development of deep learning, neural network models have achieved remarkable success in various fields by leveraging the nonlinear modeling capabilities of hidden layers and activation functions. Neural network models possess powerful feature extraction and representation capabilities, and can improve the accuracy of anomaly prediction by learning the operating patterns of power generation equipment.

[0003] Existing technologies use neural network models to extract equipment status and abnormal features from multi-source data to predict equipment anomalies. However, no additional equipment association knowledge is introduced into the neural network processing unit to improve prediction accuracy. At the same time, existing technologies do not utilize reinforcement learning-based neural network models to adaptively optimize maintenance schemes, which reduces the effectiveness of power generation equipment anomaly prediction and adaptive maintenance.

[0004] To address this, a power generation equipment anomaly prediction and adaptive maintenance system based on fused deep learning is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a power generation equipment anomaly prediction and adaptive maintenance system based on fused deep learning. This system provides an integrated information system service for equipment anomaly monitoring and adaptive maintenance. First, a multimodal acquisition unit acquires multimodal local and global data. A first neural network processing unit inputs the multimodal local data into a knowledge-embedded hybrid neural network model to obtain local anomaly predictions. A second neural network processing unit inputs the multimodal global data into an encoder-decoder network based on cross-modal attention to obtain global anomaly predictions. A reinforcement learning processing unit optimizes the maintenance strategy using a reinforcement learning model and performs hyperparameter optimization on the reinforcement learning model using a genetic algorithm, thus achieving model interpretability. Finally, an output unit provides a comprehensive anomaly prediction result and an updated maintenance strategy optimized by reinforcement learning and genetic algorithms to improve the effectiveness of anomaly prediction and maintenance.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A power generation equipment anomaly prediction and adaptive maintenance system based on fusion deep learning includes:

[0008] The multimodal acquisition unit is used to acquire multimodal local data and multimodal global data of the power generation equipment;

[0009] The first neural network processing unit is used to input the multimodal local data into a knowledge-embedded hybrid neural network model to obtain local prediction values ​​of device anomalies.

[0010] The second neural network processing unit is used to input the multimodal global data into an encoder-decoder network based on cross-modal attention to obtain a global prediction value of device anomaly.

[0011] The reinforcement learning processing unit is used to optimize the equipment maintenance strategy using a reinforcement learning model and to perform hyperparameter optimization on the reinforcement learning model using a genetic algorithm to obtain an updated equipment maintenance strategy.

[0012] The output unit is used to determine a comprehensive equipment anomaly prediction value based on the weighted result of the local equipment anomaly prediction value and the global equipment anomaly prediction value, and output the anomaly prediction result corresponding to the comprehensive equipment anomaly prediction value and the updated equipment maintenance strategy.

[0013] Furthermore, the multimodal local data refers to the multimodal state data of the equipment components; the multimodal global data includes: equipment operating parameter data, environmental data, power consumption data, and maintenance log data.

[0014] Furthermore, the knowledge-embedded hybrid neural network model implemented by the first neural network processing unit is configured to calculate the local predicted value of the device anomaly in the following manner:

[0015] The knowledge-embedded hybrid neural network model includes: an input layer, a modality feature extraction layer, a knowledge embedding fusion layer, and a feature fusion and prediction layer;

[0016] The input layer is used to receive preprocessed multimodal local data;

[0017] The modality feature extraction layer is used to process the preprocessed multimodal local data using multiple parallel neural network models to generate feature vectors for different modalities.

[0018] The knowledge embedding fusion layer is used to fuse entity embedding vectors with feature vectors of different modalities. The fusion can be achieved through vector concatenation, attention weighting, or gating mechanisms.

[0019] The feature fusion and prediction layer is used to integrate the fused feature vectors of different modalities using a neural network layer, and output the local prediction value of the device anomaly.

[0020] Furthermore, the encoder-decoder network based on cross-modal attention implemented by the second neural network processing unit is configured to compute the global prediction value of the device anomaly in the following manner:

[0021] The encoder-decoder network based on cross-modal attention includes an encoder and a decoder;

[0022] The encoder includes:

[0023] Multiple modality-independent encoding subnetworks are used to each process a global data modality to generate an initial vector representation;

[0024] The cross-modal attention mechanism module is used to calculate and apply attention weights, so that the vector representations of each modality can be adjusted and weighted according to information from other modalities to capture the inter-modal dependencies.

[0025] The context aggregation module is used to aggregate the modal representations after cross-modal attention adjustment into a context vector, which contains comprehensive state information of the multimodal global data;

[0026] The decoder receives the context vector and maps the context vector using a multilayer perceptron combined with an activation function to output the global prediction value of the device anomaly.

[0027] Furthermore, the process by which the reinforcement learning processing unit obtains the updated equipment maintenance strategy by combining reinforcement learning and genetic algorithms includes:

[0028] Define the state space, action space, and reward function in the environment model;

[0029] DQN was chosen as the reinforcement learning agent model and interacted with a simulated environment to learn the optimal policy.

[0030] The key hyperparameters of the reinforcement learning agent model are optimized using a genetic algorithm. This process includes: encoding hyperparameter combinations into individuals in the genetic algorithm; determining the fitness of the individuals by evaluating the learning performance of the reinforcement learning agent model under the corresponding hyperparameter configuration; and iteratively evolving the hyperparameter combinations by performing selection, crossover, and mutation operations on the individual population to search for the optimal hyperparameter configuration.

[0031] The reinforcement learning agent model is trained using the optimal hyperparameter configuration to obtain the updated equipment maintenance strategy.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] 1. This invention proposes a first neural network processing unit for calculating local predicted values ​​of equipment anomalies using a hybrid neural network model based on knowledge embedding. This neural network model integrates the multimodal features of power generation equipment components with domain knowledge by utilizing a knowledge embedding fusion layer, which can improve the accuracy of local anomaly prediction and thus improve the effectiveness of power generation equipment anomaly prediction and adaptive maintenance.

[0034] 2. This invention proposes a second neural network processing unit for calculating the global prediction value of equipment anomalies using an encoder-decoder network based on cross-modal attention. This encoder-decoder model automatically learns the correlation between different global data modes by utilizing the cross-modal attention mechanism module, and combined with the attention mechanism, it can make the model pay more attention to the modal information related to anomaly prediction, suppress the interference of irrelevant information, improve the accuracy of global anomaly prediction, and thus improve the effectiveness of power generation equipment anomaly prediction and adaptive maintenance.

[0035] 3. This invention proposes a reinforcement learning processing unit to combine reinforcement learning and genetic algorithms to obtain updated equipment maintenance strategies. This unit, by combining reinforcement learning and genetic algorithms, can adaptively adjust the maintenance strategy according to the equipment's state and operating environment. Among them, the genetic algorithm can automatically search for the optimal hyperparameter combination, avoiding the tedious process of manual parameter tuning, thereby improving the learning efficiency and performance of DQN, enabling DQN to converge to the optimal strategy more quickly. This can improve the effectiveness of power generation equipment anomaly prediction and adaptive maintenance. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the structure of the power generation equipment anomaly prediction and adaptive maintenance system based on fused deep learning proposed in this invention;

[0037] Figure 2 This is a schematic diagram of the structure of the knowledge-embedded hybrid neural network model of the present invention;

[0038] Figure 3 This is a schematic diagram illustrating the process by which the reinforcement learning processing unit of the present invention obtains an updated device maintenance strategy. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Please see Figures 1 to 3 This invention provides a power generation equipment anomaly prediction and adaptive maintenance system based on fused deep learning, the technical solution of which is as follows:

[0041] Example 1:

[0042] To improve the effectiveness and efficiency of power generation equipment anomaly prediction and adaptive maintenance, a company used the power generation equipment anomaly prediction and adaptive maintenance system based on fused deep learning proposed in this invention. The structure of the system is as follows: Figure 1 As shown, the details are as follows:

[0043] The multimodal acquisition unit is used to acquire multimodal local data and multimodal global data of the power generation equipment;

[0044] Furthermore, multimodal local data refers to the multimodal state data of equipment components; multimodal global data includes: equipment operating parameter data, environmental data, power consumption data, and maintenance log data.

[0045] Furthermore, the multimodal state data includes: temperature, sound, oil, voltage, pressure, image, and other data; equipment operating parameter data includes: generator temperature, equipment load, and generator voltage, etc.; environmental data includes ambient temperature and ambient humidity.

[0046] Furthermore, multimodal status data and environmental data are collected through sensor arrays; equipment operating parameter data, power consumption data, and maintenance log data can be obtained from the background access of the power generation equipment status management platform.

[0047] By introducing multimodal local and global data, we can provide data support for subsequent anomaly prediction models and anomaly prediction models. Introducing multimodal data can analyze modal information related to equipment anomaly prediction from different dimensions, which can improve the effectiveness of power generation equipment anomaly prediction and adaptive maintenance.

[0048] The first neural network processing unit is used to input multimodal local data into a knowledge-embedded hybrid neural network model to obtain local prediction values ​​of device anomalies.

[0049] Furthermore, the knowledge-embedded hybrid neural network model implemented by the first neural network processing unit is configured to calculate local predicted values ​​of device anomalies in the following manner:

[0050] The structure of a knowledge-embedded hybrid neural network model is as follows: Figure 2 As shown, it includes: an input layer, a modal feature extraction layer, a knowledge embedding and fusion layer, and a feature fusion and prediction layer;

[0051] The input layer is used to receive preprocessed multimodal local data;

[0052] The preprocessing process includes: data alignment, data cleaning, and data normalization.

[0053] The modality feature extraction layer is used to process preprocessed multimodal local data using multiple parallel neural network models to generate feature vectors for different modalities;

[0054] The knowledge embedding fusion layer is used to fuse entity embedding vectors with feature vectors from different modalities. Fusion can be achieved through vector concatenation, attention weighting, or gating mechanisms.

[0055] The feature fusion and prediction layer is used to integrate fused feature vectors from different modalities using neural network layers, and outputs local prediction values ​​for device anomalies.

[0056] Furthermore, the modal feature extraction layer selects the appropriate neural network model for different modal data. For example, a one-dimensional convolutional neural network is used for sound data, a two-dimensional convolutional neural network is used for image data, a long short-term memory network is used for temperature, pressure and voltage data, and an autoencoder network is used for oil data.

[0057] Furthermore, the process for obtaining entity embedding vectors includes: constructing a knowledge graph of power generation equipment; inputting the knowledge graph of power generation equipment into a pre-trained knowledge graph embedding model to obtain entity embedding vectors;

[0058] Furthermore, the entity relationships in the power generation equipment knowledge graph can be extracted from sources such as equipment manuals, maintenance records, and experience summaries; entity relationships include: equipment components, anomaly types, connection relationships between components, and causal relationships between components and faults / anomalies, etc.

[0059] Furthermore, the knowledge graph embedding model can be selected from TransE model, ComplEx model, RotatE model, etc.;

[0060] Furthermore, the neural network layer used in the feature fusion and prediction layer is a multilayer perceptron;

[0061] Furthermore, the process of obtaining local predicted values ​​of device anomalies using a hybrid neural network model based on knowledge embedding can be represented as:

[0062]

[0063] Where JBYC represents the local predicted value of device anomaly; MLP represents the multilayer perceptron network used in the feature fusion and prediction layer; Fu represents the fusion process of the knowledge embedding fusion layer, which can be vector concatenation, attention weighting or gating mechanism; This represents the neural network model corresponding to the i-th modality of data, where θ represents the model parameters; x i represents the i-th modal data; ev represents the entity embedding vector input to the knowledge embedding fusion layer;

[0064] Furthermore, the range of the predicted local values ​​for equipment anomalies is between (0, 1).

[0065] By utilizing the knowledge embedding fusion layer in a knowledge-embedded hybrid neural network model to embed and fuse the multimodal features of power generation equipment components with domain knowledge, the accuracy of local anomaly prediction can be improved, thereby enhancing the effectiveness of power generation equipment anomaly prediction and adaptive maintenance.

[0066] The second neural network processing unit is used to input multimodal global data into an encoder-decoder network based on cross-modal attention to obtain a global prediction value of device anomalies.

[0067] Furthermore, the encoder-decoder network based on cross-modal attention implemented by the second neural network processing unit is configured to compute global predictions of device anomalies in the following manner:

[0068] Encoder-decoder networks based on cross-modal attention include an encoder and a decoder;

[0069] The encoder includes:

[0070] Multiple modality-independent encoding subnetworks are used to each process a global data modality to generate an initial vector representation;

[0071] The cross-modal attention mechanism module is used to calculate and apply attention weights, so that the vector representations of each modality can be adjusted and weighted according to information from other modalities to capture the inter-modal dependencies.

[0072] The context aggregation module is used to aggregate the modal vector representations after cross-modal attention adjustment into a context vector, which contains comprehensive state information of the multimodal global data;

[0073] The decoder receives the context vector and maps it using a multilayer perceptron and activation function to output a global prediction value for device anomalies.

[0074] Furthermore, the process by which the cross-modal attention mechanism module uses cross-modal attention to obtain the adjusted representations of each modality vector can be represented as follows:

[0075] z i ′=LN(z i +A i *V);

[0076]

[0077] Among them, z i ′ represents the adjusted vector representation of the i-th mode; LN represents the layer normalization operation; z i A represents the initial vector representation of the i-th mode; i V represents the attention weight matrix for the i-th modality; V represents the value matrix for all modalities, specifically... Where T is the transpose operation, N is the total number of modal types; softmax represents the normalized exponential function; Q i Let K represent the query matrix for the i-th modality, which is obtained by multiplying the initial vector representation of the i-th modality with the learnable parameter matrix; K represents the key matrix for all modalities, specifically... d k This represents the dimension of the key matrix.

[0078] Furthermore, the context aggregation module aggregates the adjusted modal vector representations into a single context vector using a channel concatenation operation;

[0079] Furthermore, the range of the global predicted value for equipment anomalies is between (0, 1).

[0080] By utilizing the cross-modal attention mechanism module in the encoder-decoder network based on cross-modal attention to automatically learn the correlation between different global data modalities, and combining it with the attention mechanism, the model can pay more attention to modal information related to anomaly prediction, suppress the interference of irrelevant information, improve the accuracy of global anomaly prediction, and thus improve the effectiveness of anomaly prediction and adaptive maintenance of power generation equipment.

[0081] The reinforcement learning processing unit is used to optimize the equipment maintenance strategy using a reinforcement learning model and to optimize the hyperparameters of the reinforcement learning model using a genetic algorithm to obtain an updated equipment maintenance strategy.

[0082] Furthermore, the reinforcement learning processing unit obtains the updated equipment maintenance strategy by combining reinforcement learning and genetic algorithms, as follows: Figure 3 As shown, the specific process is as follows:

[0083] Define the state space, action space, and reward function in the environment model;

[0084] DQN was chosen as the reinforcement learning agent model and interacted with a simulated environment to learn the optimal policy.

[0085] The key hyperparameters of a reinforcement learning agent model are optimized using a genetic algorithm. The process includes: encoding hyperparameter combinations into individuals in the genetic algorithm; determining the fitness of individuals by evaluating the learning performance of the reinforcement learning agent model under the corresponding hyperparameter configuration; and iteratively evolving the hyperparameter combinations by performing selection, crossover, and mutation operations on the individual population to search for the optimal hyperparameter configuration.

[0086] By training a reinforcement learning agent model using optimal hyperparameter configuration, an updated equipment maintenance strategy can be obtained.

[0087] Furthermore, the state space includes: local anomaly prediction values, global anomaly prediction values, equipment operating parameters, and historical maintenance information; the action space includes: maintenance status information, component replacement information, and operating parameter adjustment information; the reward function is constructed using power generation, equipment health status, maintenance costs, and downtime.

[0088] Furthermore, the process of training a reinforcement learning agent model using the optimal hyperparameter configuration includes: initializing the DQN model parameters using the optimal hyperparameter configuration; obtaining the current state, selecting and executing actions based on the ε-greedy policy using the ε-parameter parameter in the optimal hyperparameter configuration, and observing the reward and the next state from the environment; storing the experience in the experience replay buffer and randomly sampling a batch of experience from it; calculating the target Q-value and the loss function; updating the DQN parameters using the learning rate α in the optimal hyperparameter configuration; periodically updating the target network; and obtaining the final reinforcement learning agent model after training converges or reaches a predetermined condition.

[0089] By employing reinforcement learning, maintenance strategies can be adaptively adjusted based on the equipment's status and operating environment. Simultaneously, genetic algorithms are used to automatically search for the optimal hyperparameter combination, avoiding the tedious process of manual parameter tuning. This improves the learning efficiency and performance of the DQN model, enabling it to converge to the optimal strategy more quickly. This enhances the effectiveness of anomaly prediction and adaptive maintenance for power generation equipment.

[0090] The output unit is used to determine the comprehensive equipment anomaly prediction value based on the weighted result of the local and global equipment anomaly prediction values, and output the anomaly prediction result corresponding to the comprehensive equipment anomaly prediction value and the updated equipment maintenance strategy.

[0091] Furthermore, the formula for calculating the predicted value of integrated equipment anomalies is as follows:

[0092] ZHYC=ω loc *JBYC+ω glo *QJYC;

[0093] ω loc +ω glo =1.0;

[0094] Where ZHYC represents the predicted value of integrated equipment anomalies; ω loc Represents the local prediction weight; JBYC represents the local prediction value of equipment anomalies; ω glo QJYC represents the global prediction weight; QJYC represents the global prediction value for equipment anomalies.

[0095] Furthermore, the local prediction weight and the global prediction weight are set to 0.5; the weight setting can be adjusted according to actual needs and is not unique.

[0096] To illustrate the comprehensive equipment anomaly prediction value proposed in this invention, multimodal data from three different power generation devices were randomly selected for testing, denoted as Sample 1, Sample 2, and Sample 3, respectively. The multimodal data included multimodal local data and multimodal global data. The multimodal data from each group were input into the first neural network processing unit and the second neural network processing unit for processing, yielding the local and global equipment anomaly prediction values ​​for each group. Combining this with the calculation formula for the comprehensive equipment anomaly prediction value, the test results for the comprehensive equipment anomaly prediction value were obtained, as shown in Table 1.

[0097] Table 1. Test Results of Comprehensive Equipment Anomaly Prediction Values

[0098]

[0099] This embodiment proposes a power generation equipment anomaly prediction and adaptive maintenance system based on fused deep learning. The system first acquires multimodal local and global data through a multimodal acquisition unit. Then, a first neural network processing unit inputs the multimodal local data into a knowledge-embedded hybrid neural network model to obtain local anomaly predictions. Next, a second neural network processing unit inputs the multimodal global data into an encoder-decoder network based on cross-modal attention to obtain global anomaly predictions. A reinforcement learning processing unit optimizes the maintenance strategy using a reinforcement learning model and performs hyperparameter optimization on the reinforcement learning model using a genetic algorithm, thus achieving model interpretability. Finally, an output unit provides comprehensive anomaly prediction results and updated maintenance strategies optimized by reinforcement learning and genetic algorithms to improve the effectiveness of anomaly prediction and maintenance.

[0100] Example 2:

[0101] This invention proposes a power generation equipment anomaly prediction and adaptive maintenance system based on fused deep learning. To further verify the effectiveness of the proposed equipment anomaly prediction model and anomaly judgment scheme, this invention conducts effectiveness tests on different models and schemes. This invention selects two companies, A and B, to conduct the above two sets of tests respectively.

[0102] This invention selects historical equipment multimodal data of Company A over the past 3 years as the dataset for the model in the first neural network processing unit, with data from the first and second years used as the training set and data from the third year used as the validation set. The model dataset is sampled according to the following rules: 20 sets of data are randomly selected each week.

[0103] This invention inputs the training set collected from Company A into different local anomaly prediction models for training, and obtains their respective pre-trained models; then, the validation set is input into each pre-trained model to obtain the local anomaly prediction values ​​of each model; then, by combining historical local anomaly results and using manual verification methods, the proportion of local anomaly prediction values ​​of each group within a reasonable range is obtained, thereby obtaining the effectiveness test results of the equipment anomaly prediction model.

[0104] The models used in the test are as follows: the local anomaly prediction model proposed in this invention, denoted as Model 1; the sub-network of the modal feature extraction layer in the model is unified, denoted as Model 2; the knowledge embedding fusion layer in the local anomaly prediction model is removed, denoted as Model 3; the effectiveness test results of the equipment anomaly prediction model are shown in Table 2.

[0105] Table 2. Results of the effectiveness test of the equipment anomaly prediction model.

[0106]

[0107] The results in Table 2 show that the model proposed in this invention performs better than other models in the test results of equipment anomaly prediction model in terms of local anomaly prediction. This demonstrates that the equipment anomaly prediction model proposed in this invention can obtain reasonable and accurate anomaly prediction values, further improving the effectiveness of power generation equipment anomaly prediction and adaptive maintenance.

[0108] To further test the effectiveness of the equipment anomaly prediction scheme, this embodiment collected historical multimodal equipment data from Company B over the past two years. Following the same dataset partitioning method and data sampling rules, different anomaly prediction schemes were tested on the same data to obtain their respective equipment anomaly prediction values. The prediction values ​​were then combined with historical anomaly results and manually verified to obtain the prediction accuracy of each scheme. The equipment anomaly prediction schemes for each group are denoted as Scheme 1, Scheme 2, and Scheme 3, respectively. Scheme 1 is the scheme proposed in this invention, which is a weighted sum of local and global equipment anomaly prediction values. Scheme 2 uses only local equipment anomaly prediction values ​​for prediction. Scheme 3 uses only global equipment anomaly prediction values ​​for prediction. The effectiveness test results of the equipment anomaly prediction schemes are shown in Table 3.

[0109] Table 3. Results of the effectiveness test of the equipment anomaly prediction scheme

[0110] Test Plan Prediction accuracy Option 1 90.5％ Option 2 88.3％ Option 3 88.9％

[0111] As can be seen from the results in Table 3, the anomaly prediction method proposed in this invention, using Scheme 1, yields better anomaly prediction accuracy test results than other schemes. This indicates that it is necessary to combine local anomaly prediction and global anomaly prediction of equipment, which can significantly improve the effectiveness of anomaly prediction and adaptive maintenance of power generation equipment.

[0112] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A power generation equipment anomaly prediction and adaptive maintenance system based on fusion deep learning, characterized in that, include: The multimodal acquisition unit is used to acquire multimodal local data and multimodal global data of the power generation equipment; Multimodal local data refers to the multimodal state data of equipment components; Multimodal global data includes: equipment operating parameter data, environmental data, power consumption data, and maintenance log data; multimodal status data includes: temperature, sound, fluid, voltage, pressure, and image data; equipment operating parameter data includes: generator temperature, equipment load, and generator voltage; environmental data includes ambient temperature and ambient humidity. The first neural network processing unit is used to input the multimodal local data into a knowledge-embedded hybrid neural network model to obtain local predicted values ​​of equipment anomalies. For different modalities, a corresponding neural network model is selected: a one-dimensional convolutional neural network is used for sound data, a two-dimensional convolutional neural network for image data, a long short-term memory network for temperature, pressure, and voltage data, and an autoencoder network for oil data. A knowledge graph of the power generation equipment is constructed. The knowledge graph is input into a pre-trained knowledge graph embedding model to obtain entity embedding vectors. Entity relationships in the knowledge graph are extracted from equipment manuals and maintenance records. Entity relationships include: equipment components, anomaly types, connections between components, and causal relationships between components and faults. The knowledge-embedded hybrid neural network model includes an input layer, a modal feature extraction layer, a knowledge embedding fusion layer, and a feature fusion and prediction layer. The input layer is used to receive preprocessed multimodal local data; the modality feature extraction layer is used to process the preprocessed multimodal local data using multiple parallel neural network models to generate feature vectors of different modalities; the knowledge embedding and fusion layer is used to fuse entity embedding vectors with the feature vectors of different modalities; the feature fusion and prediction layer is used to integrate the fused feature vectors of different modalities using neural network layers to output device anomaly local prediction values. The second neural network processing unit is used to input the multimodal global data into an encoder-decoder network based on cross-modal attention to obtain a global prediction value of device anomaly. The reinforcement learning processing unit is used to optimize the equipment maintenance strategy using a reinforcement learning model and to perform hyperparameter optimization on the reinforcement learning model using a genetic algorithm to obtain an updated equipment maintenance strategy; the reward function is constructed using power generation, equipment health status, maintenance cost, and downtime. The output unit is used to determine a comprehensive equipment anomaly prediction value based on the weighted result of the local equipment anomaly prediction value and the global equipment anomaly prediction value, and output the anomaly prediction result corresponding to the comprehensive equipment anomaly prediction value and the updated equipment maintenance strategy.

2. The power generation equipment anomaly prediction and adaptive maintenance system based on fused deep learning according to claim 1, characterized in that, The multimodal local data refers to the multimodal state data of the equipment components; the multimodal global data includes: equipment operating parameter data, environmental data, power consumption data, and maintenance log data.

3. The power generation equipment anomaly prediction and adaptive maintenance system based on fused deep learning according to claim 1, characterized in that, The encoder-decoder network based on cross-modal attention, implemented by the second neural network processing unit, is configured to compute the global prediction value of the device anomaly in the following manner: The encoder-decoder network based on cross-modal attention includes an encoder and a decoder; The encoder includes: Multiple modality-independent encoding subnetworks are used to each process a global data modality to generate an initial vector representation; The cross-modal attention mechanism module is used to calculate and apply attention weights, so that the vector representations of each modality can be adjusted and weighted according to information from other modalities to capture the inter-modal dependencies. The context aggregation module is used to aggregate the modal representations after cross-modal attention adjustment into a context vector, which contains comprehensive state information of the multimodal global data; The decoder receives the context vector and maps the context vector using a multilayer perceptron combined with an activation function to output the global prediction value of the device anomaly.

4. The power generation equipment anomaly prediction and adaptive maintenance system based on fused deep learning according to claim 1, characterized in that, The process by which the reinforcement learning processing unit obtains an updated equipment maintenance strategy by combining reinforcement learning and genetic algorithms includes: Define the state space, action space, and reward function in the environment model; DQN was chosen as the reinforcement learning agent model and interacted with a simulated environment to learn the optimal policy. The key hyperparameters of the reinforcement learning agent model are optimized using a genetic algorithm. This process includes: encoding hyperparameter combinations into individuals in the genetic algorithm; determining the fitness of the individuals by evaluating the learning performance of the reinforcement learning agent model under the corresponding hyperparameter configuration; and iteratively evolving the hyperparameter combinations by performing selection, crossover, and mutation operations on the individual population to search for the optimal hyperparameter configuration. The reinforcement learning agent model is trained using the optimal hyperparameter configuration to obtain the updated equipment maintenance strategy.

5. The power generation equipment anomaly prediction and adaptive maintenance system based on fused deep learning according to claim 4, characterized in that, The process of training the reinforcement learning agent model using the optimal hyperparameter configuration includes: initializing the DQN model parameters using the optimal hyperparameter configuration; obtaining the current state, selecting and executing actions using the ε parameter in the optimal hyperparameter configuration and according to the ε-greedy policy, and observing the reward and the next state from the environment; storing the experience in the experience replay buffer and randomly sampling a batch of experience from it; calculating the target Q-value and the loss function; updating the DQN parameters using the learning rate α in the optimal hyperparameter configuration; periodically updating the target network; and obtaining the final reinforcement learning agent model after training converges or reaches a predetermined condition.

Images