Nuclear power equipment multidimensional data analysis method based on deep multi-signal fusion confrontation

Abstract

The invention belongs to the technical field of nuclear power operation safety, and particularly relates to a nuclear power equipment multi-dimensional data analysis method based on deep multi-signal fusion confrontation. Comprising the following steps: S1, constructing and setting a feature encoder; s2, constructing a preliminary data analysis module; s3, constructing a fine multi-dimensional data analysis module; s4, data acquisition; s5, data preprocessing; s6, model training; and S7, deploying the model. The nuclear power equipment multi-dimensional data analysis method based on deep multi-signal fusion adversarial has the beneficial effects that multi-dimensional data analysis is performed on multiple signals collected by a fusion multi-source sensor based on a convolutional neural network and a graph convolutional neural network by combining a multi-instance learning method and an adversarial learning method; the automation level of equipment fault diagnosis is realized, and the safety and stability of nuclear power plant operation are ensured.

Description

Technical Field

[0001] The present invention belongs to the field of nuclear power operation safety technology, and specifically relates to a multi-dimensional data analysis method for nuclear power equipment based on deep multi-signal fusion confrontation. Background Art

[0002] With the continuous growth of energy demand and the global promotion of clean energy, nuclear power, as an efficient and low-carbon energy supply method, is gaining favor in an increasing number of countries and regions. However, the safety of nuclear power generation remains a global concern. Consequently, many countries have introduced strict nuclear safety regulations and standards, requiring nuclear power plants to utilize state-of-the-art technologies and management methods to ensure safe and reliable operation. Against this backdrop, improving the intelligence level of nuclear power plants, particularly in equipment inspection and fault prevention, has become a key development direction for the industry.

[0003] Equipment fault diagnosis in nuclear power plants is a crucial component of ensuring the safe and efficient operation of nuclear power plants. The increasing complexity of nuclear power plant machinery makes it difficult for traditional manual fault diagnosis methods to quickly detect and address mechanical equipment failures in their early stages. These methods are subject to lags and are easily affected by human factors, making their efficiency and accuracy insufficient to meet increasingly stringent nuclear safety regulations. Therefore, an automated and intelligent method for mechanical equipment fault diagnosis is needed to improve the frequency and accuracy of nuclear power plant mechanical equipment fault diagnosis, reduce human error, and enhance the intelligence of fault diagnosis. Furthermore, with the development of artificial intelligence (AI) technology, the optimal application of AI technology to nuclear power plant operation and maintenance management has become a key trend in technological development. Summary of the Invention

[0004] The purpose of the present invention is to provide a multi-dimensional data analysis method for nuclear power equipment based on deep multi-signal fusion confrontation to solve the problems of low efficiency and strong hysteresis in fault diagnosis of traditional nuclear power plant equipment. By performing multi-dimensional data analysis on the fusion of mechanical equipment signals received by multi-source sensors, efficient and accurate fault diagnosis results are provided, which can improve the safety and efficiency of nuclear power plant operation, reduce the error rate and hysteresis of manual fault diagnosis, achieve rapid response and intelligent diagnosis support, thereby effectively improving the intelligent management level and overall operation and maintenance capabilities of nuclear power plants.

[0005] The technical solution of the present invention is as follows: A multi-dimensional data analysis method for nuclear power equipment based on deep multi-signal fusion confrontation, comprising the following steps:

[0006] S1: build setting feature encoder;

[0007] S2: construct preliminary data analysis module;

[0008] S3: Build a sophisticated multidimensional data analysis module;

[0009] S4: Data collection;

[0010] S5: Data preprocessing;

[0011] S6: Model training;

[0012] S7: Model deployment.

[0013] The S1 uses a convolutional neural network as a feature encoder to extract instance embeddings from signal data collected by multi-source sensors and aggregate all instance embeddings into packet embeddings for multidimensional data analysis and fault diagnosis;

[0014] The feature encoder takes raw data collected by multi-source sensors as input and uses the convolutional and pooling layers of CNNs to perform hierarchical feature extraction on the data. It mines fault features of different scales from the raw signals and ultimately outputs the instance embedding corresponding to each sensor data, achieving the conversion from raw data to feature representation.

[0015] The instance embeddings generated by each sensor are then further aggregated to form a package embedding.

[0016] The S2 preliminary data analysis module analyzes the contribution of each sensor data to fault diagnosis through instance average pooling, and quantifies the importance factor of the corresponding sensor. The preliminary data analysis module is based on multi-instance learning, fusing the extracted features with the calculated importance factor, integrating multi-source sensor information, and performing preliminary identification and judgment of equipment faults.

[0017] The S3 refined multidimensional data analysis module first performs dual-graph generation and learning. By constructing a feature base graph and using the feature matrix to generate a graph structure to mine data topological relationships, it also generates an importance base graph, combines the previous importance factors to construct a graph structure, and fuses topological features. Through a dual-graph learning strategy, it integrates the embedding of the two graphs to obtain a richer feature representation.

[0018] Secondly, a subgraph learning strategy is set up to select some nodes to generate subgraphs based on importance factors. The semantic consistency of the subgraphs is constrained by a loss function, which regularizes the training process and improves the robustness of the model to limited data.

[0019] Finally, dual-graph learning and sub-graph strategies are integrated, and the model parameters are updated in combination with the label smoothed cross entropy loss function, ultimately achieving refined multi-dimensional data analysis and outputting accurate diagnosis results.

[0020] The S4 installs vibration sensors and voiceprint sensors in the water pump room of the nuclear power plant to collect vibration signals and voiceprint signals of mechanical equipment in the nuclear power plant during operation, and stores them in the database for training of multidimensional data analysis models.

[0021] The S5 performs data cleaning on the data collected by the multi-source sensors to remove outliers, noise and missing data.

[0022] The S6 is based on a multi-instance learning method and an adversarial learning method, and uses pre-processed multi-dimensional data to train the model, thereby realizing multi-dimensional data analysis of nuclear power equipment based on deep multi-signal fusion adversarial.

[0023] The S7 integrates the model into the artificial intelligence algorithm component and deploys it as a whole on the inspection management platform to realize fault diagnosis of nuclear power plant mechanical equipment.

[0024] Including continuous iteration and optimization of S8.

[0025] The S8 continuously updates the model and optimizes performance based on the feedback of the algorithm diagnosis results, regularly collects diagnosis results and data, and continuously improves the accuracy of the model in multi-dimensional data analysis through retraining.

[0026] The beneficial effects of this invention are as follows: The present invention's multidimensional data analysis method for nuclear power equipment based on deep multi-signal fusion adversarial analysis combines multi-instance learning with adversarial learning methods, using convolutional neural networks and graph convolutional neural networks to perform multidimensional data analysis on multiple signals collected by fused multi-source sensors. This method achieves automated equipment fault diagnosis, ensuring the safety and stability of nuclear power plant operations. This method optimizes the fault detection process, reduces the frequency and error rate of manual operations, and enhances the ability to predict and respond to potential safety risks, thereby improving the operational efficiency and safety management effectiveness of nuclear power plants. The feature encoder is used to extract data features and generate embedded representations. The importance factor of each sensor in the preliminary data analysis is calculated based on multiple instance learning (MIL). Finally, the features and importance factors are fused to achieve preliminary signal fusion and data analysis based on MIL, completing the first stage of processing; a dual graph based on features and importance is generated to mine data topological relationships and fuse features; nodes are selected through subgraph learning strategies to generate subgraphs to ensure semantic consistency; finally, the above processing is integrated to achieve refined multidimensional data analysis, completing the second stage of refined data analysis; the method is deployed on an intelligent management platform, reducing the operational complexity of traditional manual inspections, realizing equipment fault diagnosis through automated data analysis, significantly reducing the human error rate, and improving the operational safety and efficiency of nuclear power plants; the method has excellent stability and accuracy, can effectively cope with daily maintenance tasks and emergency handling, ensure that nuclear power plants can maintain high standards of safety performance under various operating conditions, and enhance the emergency response capabilities of nuclear power plants in uncertain environments. DETAILED DESCRIPTION

[0027] The present invention is further described in detail below with reference to specific embodiments.

[0028] A multi-dimensional data analysis method for nuclear power equipment based on deep multi-signal fusion confrontation includes the following steps:

[0029] S1: Build Setup Feature Encoder

[0030] A convolutional neural network is used as the feature encoder to extract instance embeddings from signal data collected by multi-source sensors. All instance embeddings are aggregated into packet embeddings for multidimensional data analysis and fault diagnosis.

[0031] The feature encoder takes the raw data collected by multi-source sensors as input and uses components such as convolutional layers and pooling layers of CNNs to perform hierarchical feature extraction on the data, mining fault features of different scales from the raw signals. It finally outputs the instance embedding corresponding to each sensor data, realizing the conversion from raw data to feature representation.

[0032] The instance embeddings generated by each sensor are then aggregated to form a package embedding. This process provides a unified feature representation for subsequent preliminary fault diagnosis based on multiple instance learning (MIL). It enables the model to integrate multi-source sensor information and calculate the contribution (i.e., importance factor) of each sensor in fault diagnosis, thus supporting the first stage of the two-stage diagnosis framework.

[0033] S2: Building a preliminary data analysis module

[0034] Instance average pooling (IAP) is introduced to calculate the contribution of each sensor to fault diagnosis, and then calculate the importance factor, improve the gradient update of instance embedding, and improve the accuracy of data analysis results.

[0035] The preliminary data analysis module uses instance average pooling (IAP) to analyze the contribution of each sensor data to fault diagnosis and quantify the importance factor of the corresponding sensor. This factor is used to optimize the gradient update mechanism of instance embedding, making the model focus more on sensor features that are critical to diagnosis and enhancing the diagnostic effectiveness of the feature representation.

[0036] The module is based on multiple instance learning (MIL), which fuses the extracted features with the calculated importance factors, integrates multi-source sensor information, and performs preliminary identification and judgment of equipment faults, providing basic results for the subsequent second-stage fine diagnosis.

[0037] S3: Building a sophisticated multidimensional data analysis module

[0038] We propose dual-graph generation, including feature-based and importance-based graph generation. We fuse the embeddings of two graphs through a dual-graph learning strategy to generate refined multi-dimensional data analysis results. We also design a loss function to achieve fault information fusion based on data analysis results.

[0039] The refined multidimensional data analysis module first performs dual-graph generation and learning. By constructing a feature base graph and using the feature matrix to generate a graph structure, the module mines data topological relationships. Simultaneously, the module generates an importance base graph, combining the previously identified importance factors to construct a graph structure and integrate topological features. Through a dual-graph learning strategy, the module integrates the embeddings of the two graphs to obtain a richer feature representation.

[0040] Secondly, a subgraph learning strategy is set up to select some nodes to generate subgraphs based on the importance factor. The semantic consistency of the subgraphs is constrained by the loss function, the training process is regularized, the robustness of the model to limited data is improved, and label bias and training bias problems are avoided.

[0041] Finally, dual-graph learning and sub-graph strategies are integrated, and loss functions such as label smoothed cross entropy are used to update model parameters, ultimately achieving refined multi-dimensional data analysis and outputting accurate diagnostic results.

[0042] S4: Data Collection

[0043] Vibration sensors and voiceprint sensors are installed in the pump room of a nuclear power plant to collect vibration signals and voiceprint signals during the operation of mechanical equipment in the nuclear power plant, and store them in a database for training multidimensional data analysis models.

[0044] S5: Data Preprocessing

[0045] Data collected by multi-source sensors is cleaned to remove outliers, noise and missing data. Since different sensors have different sampling rates, data needs to be synchronized for effective analysis.

[0046] S6: Model Training

[0047] Based on the multi-instance learning method and adversarial learning method, the model is trained using preprocessed multidimensional data to realize multidimensional data analysis of nuclear power equipment based on deep multi-signal fusion adversarial.

[0048] Multiple-Instance Learning (MIL) is a weakly supervised learning method. Its core concept is that the input is a "bag" rather than a single sample. Each bag contains multiple "instances." Only the bags have labels; the instances themselves are unlabeled. MIL achieves classification or regression for unknown instances by learning the relationship between bag-level labels and instance features.

[0049] Adversarial learning is a machine learning paradigm that trains models through a "competition-collaboration" mechanism. Its core concept is to pit two neural networks against each other, improving their respective capabilities through competition. Adversarial learning strategies can enhance data, generate synthetic fault signals, and alleviate the small sample size problem (for example, by simulating rare fault types).

[0050] S7: Model deployment

[0051] The model is integrated into the artificial intelligence algorithm component and deployed as a whole on the inspection management platform to realize fault diagnosis of nuclear power plant mechanical equipment.

[0052] S8: Continuous Iteration and Optimization

[0053] Based on the feedback from the algorithm's diagnostic results, the model is continuously updated and performance optimized, diagnostic results and data are collected regularly, and the accuracy of the model in multi-dimensional data analysis is continuously improved through retraining.

Claims

1. A multi-dimensional data analysis method for nuclear power equipment based on deep multi-signal fusion confrontation, characterized in that: The steps include: S1: build setting feature encoder; S2: construct preliminary data analysis module; S3: Build a sophisticated multidimensional data analysis module; S4: Data collection; S5: Data preprocessing; S6: Model training; S7: Model deployment.

2. The method for multidimensional data analysis of nuclear power equipment based on deep multi-signal fusion confrontation according to claim 1, characterized in that: The S1 uses a convolutional neural network as a feature encoder to extract instance embeddings from signal data collected by multi-source sensors and aggregate all instance embeddings into packet embeddings for multidimensional data analysis and fault diagnosis; The feature encoder takes raw data collected by multi-source sensors as input and uses the convolutional and pooling layers of CNNs to perform hierarchical feature extraction on the data. It mines fault features of different scales from the raw signals and ultimately outputs the instance embedding corresponding to each sensor data, achieving the conversion from raw data to feature representation. The instance embeddings generated by each sensor are then further aggregated to form a package embedding.

3. The method for multidimensional data analysis of nuclear power equipment based on deep multi-signal fusion confrontation according to claim 1, characterized in that: The S2 preliminary data analysis module analyzes the contribution of each sensor data to fault diagnosis through instance average pooling, and quantifies the importance factor of the corresponding sensor. The preliminary data analysis module is based on multi-instance learning, fusing the extracted features with the calculated importance factor, integrating multi-source sensor information, and performing preliminary identification and judgment of equipment faults.

4. The method for multidimensional data analysis of nuclear power equipment based on deep multi-signal fusion confrontation according to claim 1, characterized in that: The S3 refined multidimensional data analysis module first performs dual-graph generation and learning. By constructing a feature base graph and using the feature matrix to generate a graph structure to mine data topological relationships, it also generates an importance base graph, combines the previous importance factors to construct a graph structure, and fuses topological features. Through a dual-graph learning strategy, it integrates the embedding of the two graphs to obtain a richer feature representation. Secondly, a subgraph learning strategy is set up to select some nodes to generate subgraphs based on importance factors. The semantic consistency of the subgraphs is constrained by a loss function, which regularizes the training process and improves the robustness of the model to limited data. Finally, dual-graph learning and sub-graph strategies are integrated, and the model parameters are updated in combination with the label smoothed cross entropy loss function, ultimately achieving refined multi-dimensional data analysis and outputting accurate diagnosis results.

5. The method for multidimensional data analysis of nuclear power equipment based on deep multi-signal fusion confrontation according to claim 1, characterized in that: The S4 installs vibration sensors and voiceprint sensors in the water pump room of the nuclear power plant to collect vibration signals and voiceprint signals of mechanical equipment in the nuclear power plant during operation, and stores them in the database for training of multidimensional data analysis models.

6. The method for multidimensional data analysis of nuclear power equipment based on deep multi-signal fusion confrontation according to claim 1, characterized in that: The S5 performs data cleaning on the data collected by the multi-source sensors to remove outliers, noise and missing data.

7. The method for multidimensional data analysis of nuclear power equipment based on deep multi-signal fusion confrontation according to claim 1, characterized in that: The S6 is based on a multi-instance learning method and an adversarial learning method, and uses pre-processed multi-dimensional data to train the model, thereby realizing multi-dimensional data analysis of nuclear power equipment based on deep multi-signal fusion adversarial.

8. The method for multidimensional data analysis of nuclear power equipment based on deep multi-signal fusion confrontation according to claim 1, characterized in that: The S7 integrates the model into the artificial intelligence algorithm component and deploys it as a whole on the inspection management platform to realize fault diagnosis of nuclear power plant mechanical equipment.

9. The method for multidimensional data analysis of nuclear power equipment based on deep multi-signal fusion confrontation according to claim 1, characterized in that: Including continuous iteration and optimization of S8.

10. The method for multi-dimensional data analysis of nuclear power equipment based on deep multi-signal fusion confrontation according to claim 9, characterized in that: The S8 continuously updates the model and optimizes performance based on the feedback of the algorithm diagnosis results, regularly collects diagnosis results and data, and continuously improves the accuracy of the model in multi-dimensional data analysis through retraining.