A dual-drive fault diagnosis method for fusion reactor vacuum pump groups

Abstract

The present invention relates to the technical field of fault diagnosis of vacuum pump groups in the fusion field, and in particular to a dual-drive fault diagnosis method for fusion reactor vacuum pump groups. The technical solution includes the following steps: collecting vibration, current, temperature and sound operation data through multiple sensor units installed on the vacuum pump group, and constructing a multimodal data set in combination with historical maintenance logs; preprocessing the multimodal data set, including wavelet denoising and Z-score normalization, and using a time-frequency analysis method to convert the time series signal into a time-frequency graph; and constructing a knowledge-driven network and a data-driven network. The present invention combines the knowledge graph and fault data in the fusion field, and constructs a fault knowledge system with self-evolution capability by integrating historical equipment operation and maintenance data, domain expert knowledge and real-time monitoring information, and combines the multimodal attention mechanism to realize decision-making coordination of multi-source data, thereby improving the fault diagnosis accuracy and real-time warning capability of key equipment in fusion reactors.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis of vacuum pump groups in the fusion field, and in particular to a dual-drive fault diagnosis method for fusion reactor vacuum pump groups. Background Art

[0002] As the core equipment for maintaining a high vacuum environment, the operational reliability of a fusion reactor vacuum pump unit is directly related to the safety and stability of the fusion reactor. Current vacuum pump unit fault diagnosis relies primarily on single-sensor data monitoring and static threshold alarms, making it difficult to effectively handle dynamic correlation analysis and latent fault feature extraction from multi-source heterogeneous data under complex operating conditions. Traditional methods have three limitations:

[0003] First, the diagnostic method based on the mechanism model is difficult to adapt to the spatiotemporal correlation characteristics of the vacuum pump group's multimodal data (vibration, temperature, current, gas spectrum, etc.), resulting in early weak fault characteristics being easily masked by noise;

[0004] Second, existing knowledge graph technologies mostly use static knowledge representation, which cannot update the dynamically evolving failure modes and diagnostic rules during equipment degradation in real time;

[0005] Third, the lack of a cross-modal attention mechanism during multi-sensor data fusion leads to an imbalance in the weight distribution of key fault features in cross-domain collaborative diagnosis. In recent years, with the integrated development of deep learning and knowledge reasoning technology, how to build a multi-modal collaborative diagnosis framework with dynamic knowledge evolution capabilities to achieve accurate positioning and evolution trend prediction of complex faults in vacuum pump groups has become a technical bottleneck that urgently needs to be broken through in the field of intelligent operation and maintenance of nuclear fusion devices. Therefore, this application proposes a dual-drive fault diagnosis method for fusion reactor vacuum pump groups. Summary of the Invention

[0006] The purpose of the present invention is to address the problem in the background technology that it is impossible to accurately locate and predict the evolution trend of complex faults of vacuum pump groups, and to propose a dual-drive fault diagnosis method for fusion reactor vacuum pump groups.

[0007] The technical solution of the present invention is a dual-drive fault diagnosis method for a fusion reactor vacuum pump group, comprising the following steps:

[0008] Vibration, current, temperature, and sound operating data are collected through multiple sensor units installed on the vacuum pump group, and combined with historical maintenance logs to build a multimodal dataset;

[0009] Preprocessing the multimodal dataset includes wavelet denoising and Z-score normalization, and converting the time series signal into a time-frequency graph using a time-frequency analysis method;

[0010] Build knowledge-driven networks and data-driven networks;

[0011] Collaborative diagnosis: results based on knowledge-driven network output and confidence , rule matching degree , and the output of the data-driven network and confidence , conflict detection and collaborative decision-making are carried out through dynamic weighted rules, and the final diagnosis results are output.

[0012] Optionally, wavelet denoising uses the db4 fundamental wave with a decomposition level of 3;

[0013] The Z-score normalization formula is:

[0014] ,in, is the original data, is the mean, is the standard deviation.

[0015] Optionally, the knowledge-driven network is based on the knowledge graph in the fusion field, forms fault knowledge triples through structured expert rules, historical maintenance logs and equipment topology relationships, and calculates the knowledge graph scoring function to analyze the equipment operating status; the data-driven network uses a multimodal ResNet50 network combined with a multi-head attention mechanism to train and infer the time-frequency graph, and outputs data-driven equipment status results.

[0016] Optionally, the construction of the knowledge graph includes the following steps:

[0017] Expert Rule Set , Historical maintenance log set and device topology relationship set Merge into knowledge sets ,Right now

[0018] , each piece of knowledge is represented as a triple ,in For entities, For relationships, Fault facts are represented by a fault knowledge triple ,in It is the vacuum pump head entity. For relationships, is the tail entity;

[0019] The knowledge graph scoring function for the fusion reactor vacuum pump group is defined as:

[0020] ,in, is the score of the vacuum pump unit failure fact, express Norm or norm.

[0021] Optionally, in the data-driven network:

[0022] The calculation formula of the multi-head attention mechanism is:

[0023] , in, is the query vector, is the key vector, is a value vector, is the transformation formula of the multi-head attention mechanism of the vacuum pump group, is the first independent attention head, is the second independent attention head, is the nth independent attention head, for The weight matrix, for The weight matrix, for The weight matrix of

[0024] The training parameters include: the number of training rounds of the multimodal ResNet50 network in the data-driven network is 100 rounds, the learning rate is 0.001, the batch size is 32, and the weight decay is 0.005.

[0025] Optionally, the dynamic weighting rule of the collaborative diagnosis is:

[0026] ,

[0027] Among them, 0 means no fault, and other label numbers correspond to the corresponding fault types.

[0028] Optionally, the multimodal data set includes vibration signals, current signals, temperature signals and sound signals, and is converted into a time-frequency diagram through an EEMD time-frequency analysis method.

[0029] Optionally, the output results of the knowledge-driven network and the data-driven network interact in real time through a shared interface for automatic updating of the fault knowledge system.

[0030] Compared with the prior art, this application has at least one of the following beneficial technical effects:

[0031] By fusing multimodal sensor data with the knowledge graph in the field of fusion, and combining it with the dynamic attention mechanism to optimize feature weight distribution, the ability to extract weak fault features is significantly enhanced, achieving accurate positioning of early faults and real-time warning.

[0032] Based on the continuous updating of historical equipment operation and maintenance data, expert rules and real-time monitoring information, a self-evolving fault knowledge system is constructed, which can adapt to the degradation evolution of vacuum pump groups under complex working conditions and improve the adaptability of long-term diagnosis.

[0033] Through the knowledge-driven and data-driven dual-driven collaborative mechanism, combined with dynamic weighted rules to resolve result conflicts, and integrating domain experience and data model advantages, the reliability and robustness of diagnostic results are improved.

[0034] The multi-head attention mechanism is introduced to effectively integrate the spatiotemporal correlation characteristics of multi-source heterogeneous data, avoiding the imbalance problem of key feature weights caused by modal differences in traditional methods.

[0035] Designed for the special electromagnetic environment of fusion reactors, the system overcomes the limitations of single sensors or static models through the complementarity of structured knowledge graphs and deep learning, ensuring the safe and stable operation of the vacuum pump group under extreme conditions.

[0036] This invention combines the knowledge graph and fault data in the fusion field, and constructs a fault knowledge system with self-evolution capabilities by integrating historical equipment operation and maintenance data, domain expert knowledge and real-time monitoring information. It also combines the multimodal attention mechanism to achieve decision-making collaboration of multi-source data, thereby improving the fault diagnosis accuracy and real-time warning capabilities of key fusion reactor equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 The figure is a flow chart of a dual-drive fault diagnosis method for a fusion reactor vacuum pump group. DETAILED DESCRIPTION

[0038] The technical solution of the present invention is further described below with reference to the accompanying drawings and specific embodiments.

[0039] Example

[0040] like Figure 1 As shown, the present invention proposes a dual-drive fault diagnosis method for a fusion reactor vacuum pump group, which is described in detail below.

[0041] Various sensor units are installed on the vacuum pump assembly to capture vibration, current, temperature, and acoustic operating data. Combined with daily maintenance logs, this data is assembled into a multimodal dataset for the fusion reactor vacuum pump assembly. Data preprocessing denoises and normalizes the multimodal dataset. Wavelet denoising is used for denoising, with the base wave chosen as db4 and a decomposition level of three. Z-score normalization is used for standardization, and time-frequency analysis is used for feature extraction, converting the time series signal into a time-frequency graph. Data features and log records are fed into a knowledge-driven network, which uses knowledge graphs and expert rules for analysis and judgment to determine the current equipment operating status. Simultaneously, data features are fed into a data-driven network, which uses deep neural networks and attention mechanisms for training and inference to obtain data-driven equipment operating status results. The knowledge-driven and data-driven results are collaboratively analyzed to achieve result sharing, conflict detection, dynamic weighting, and collaborative decision-making, ultimately yielding the final output.

[0042] Knowledge-driven: By structuring empirical knowledge, expert rules, specific failure modes, historical maintenance logs and records, and equipment topology relationships are converted into knowledge graphs, including failure modes, causal relationships, etc. The expert rule set for the fusion reactor vacuum pump group is , the historical maintenance log set is , the device topology relationship set is Import all knowledge sets to form a knowledge set of multiple facts .

[0043] ,

[0044] For each piece of knowledge ,in, For entities, For relationships, is the fault fact. The fact is represented by a fault knowledge triple ,in It is the vacuum pump head entity. For relationships, The tail entity.

[0045] For example, the vacuum pump group includes bearing faults, etc., where the vacuum pump group is the head entity, belongs to the tail relationship, and the bearing fault is the tail entity.

[0046] To this end, the knowledge graph scoring function of the fusion reactor vacuum pump group is defined:

[0047] ,in, is the score of the vacuum pump unit failure fact, is the L1 norm or the L2 norm.

[0048] Data driven: The vibration, current, temperature, and sound time series signals are converted into EEMD time-frequency graphs and input into the multimodal ResNet50 network for training to obtain a multi-source neural network model. The number of training rounds is 100, the learning rate is 0.001, the batch size is set to 32, and the weight decay is set to 0.005. Considering the multi-channel input data, a multi-head attention mechanism is inserted into the multi-source neural network. is the query vector of the neural network feature of the vacuum pump group, is the key vector, is a value vector. The calculation formula of the vacuum pump group feature attention algorithm is

[0049] , in, is the transformation formula of the multi-head attention mechanism of the vacuum pump group, is the first independent attention head, is the second independent attention head, is the nth independent attention head, for The weight matrix, for The weight matrix, for The weight matrix of .

[0050] Collaborative diagnosis: For the knowledge-driven output result y1, the confidence level is c1 and the knowledge rule matching level is r1. For the data-driven output result y2, the confidence level is c2. To resolve the result conflict problem, the following rules are designed based on result sharing and dynamic weighting.

[0051] , where the output result is 0, indicating no fault, and other label numbers correspond to the corresponding fault types.

[0052] Through this rule, the final output result is calculated.

[0053] This invention fuses multimodal sensor data (vibration, current, temperature, and sound) with a fusion domain knowledge graph, and employs a dynamic attention mechanism to optimize feature weight distribution. This significantly enhances the ability to extract subtle fault signatures, enabling precise location and real-time early warning of faults. Based on the continuous updating of historical equipment operation and maintenance data, expert rules, and real-time monitoring information, a self-evolving fault knowledge system is constructed. This system can adapt to the degradation evolution of vacuum pump groups under complex operating conditions, improving the adaptability of long-term diagnostics.

[0054] It is worth noting that through the dual-driven collaborative mechanism of knowledge-driven and data-driven, combined with dynamic weighted rules to resolve result conflicts, the advantages of field experience and data models are integrated to improve the reliability and robustness of diagnostic results. The multi-head attention mechanism is introduced to effectively integrate the spatiotemporal correlation characteristics of multi-source heterogeneous data, avoiding the problem of imbalance in key feature weights caused by modal differences in traditional methods. Designed for the special electromagnetic environment of fusion reactors, the limitations of single sensors or static models are overcome through the complementarity of structured knowledge graphs and deep learning, ensuring the safe and stable operation of vacuum pump groups under extreme conditions. Providing a complete technical framework from data acquisition, knowledge modeling to collaborative decision-making, it provides feasible innovative solutions for the intelligent operation and maintenance of fusion devices, reducing maintenance costs and failure risks.

[0055] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art may make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A dual-drive fault diagnosis method for a fusion reactor vacuum pump group, characterized in that: The following steps are involved: Vibration, current, temperature, and sound operating data are collected through multiple sensor units installed on the vacuum pump group, and combined with historical maintenance logs to build a multimodal dataset; Preprocessing the multimodal dataset includes wavelet denoising and Z-score normalization, and converting the time series signal into a time-frequency graph using a time-frequency analysis method; Build knowledge-driven networks and data-driven networks; Collaborative diagnosis: results based on knowledge-driven network output and confidence , rule matching degree , and the output of the data-driven network and confidence , conflict detection and collaborative decision-making are carried out through dynamic weighted rules, and the final diagnosis results are output; The knowledge-driven network is based on the fusion field knowledge graph, forms fault knowledge triples through structured expert rules, historical maintenance logs and equipment topology relationships, and calculates the knowledge graph scoring function to analyze the equipment operation status; The data-driven network uses a multimodal ResNet50 network combined with a multi-head attention mechanism to train and infer the time-frequency graph and output data-driven device status results; The construction of the knowledge graph includes the following steps: Expert Rule Set , Historical maintenance log set and device topology relationship set Merge into knowledge sets ,Right now Each piece of knowledge is represented as a triple ,in, For entities, For relationships, Fault facts are represented by a fault knowledge triple ,in It is the vacuum pump head entity. For relationships, is the tail entity; The knowledge graph scoring function for the fusion reactor vacuum pump group is defined as: in, is the score of the vacuum pump unit failure fact, express Norm or norm; The dynamic weighting rule of the collaborative diagnosis is: Among them, 0 means no fault, and other label numbers correspond to the corresponding fault types.

2. A dual-drive fault diagnosis method for a fusion reactor vacuum pump group according to claim 1, characterized in that: Wavelet denoising uses the db4 fundamental wave with a decomposition level of 3; the Z-score normalization formula is: in, is the original data, is the mean, is the standard deviation.

3. A dual-drive fault diagnosis method for a fusion reactor vacuum pump group according to claim 1, characterized in that: In the data-driven network: The calculation formula of the multi-head attention mechanism is: in, is the query vector, is the key vector, is a value vector, is the transformation formula of the multi-head attention mechanism of the vacuum pump group, is the first independent attention head, is the second independent attention head, For the independent attention heads, for The weight matrix, for The weight matrix, for The weight matrix of The training parameters include: the number of training rounds of the multimodal ResNet50 network in the data-driven network is 100 rounds, the learning rate is 0.001, the batch size is 32, and the weight decay is 0.

005.

4. A dual-drive fault diagnosis method for a fusion reactor vacuum pump group according to claim 1, characterized in that: The multimodal data set includes vibration signals, current signals, temperature signals and sound signals, and is converted into a time-frequency diagram through the EEMD time-frequency analysis method.

5. The dual-drive fault diagnosis method for a fusion reactor vacuum pump group according to claim 1, characterized in that: The output results of the knowledge-driven network and the data-driven network interact in real time through a shared interface, and are used for self-evolution and updating of the fault knowledge system.

Images