Hydro-generator stator shafting diagram neural network intelligent checking method and system

Abstract

The invention discloses a neural network intelligent checking method and system for a hydro-generator stator shafting diagram, and belongs to the technical field of hydro-generator stator shafting prediction and calibration. The problems that in the prior art, a traditional hydro-generator stator shafting centering checking method is insufficient in precision and high in manual dependence are solved. The data acquisition module acquires data in real time and transmits the data to the data processing module; converting the collected physical parameters into a graph structure, and enhancing the graph structure to obtain an enhanced graph structure; topological features of the shafting centering state are learned through a graph neural network module, and a final feature matrix is obtained and input to a check result output module; comparing the feature distribution of the real-time feature vector with the feature distribution of the standard feature vector through an intelligent checking module, and generating a shaft system centering deviation evaluation result; and triggering alarm or outputting correction suggestion parameters through the intelligent operation and maintenance module. The water turbine set shaft system centering detection efficiency is improved, and the method can be applied to generator monitoring.

Description

Technical Field

[0001] This invention relates to a neural network-based intelligent verification method and system for the stator shaft system diagram of a hydro-generator, belonging to the field of hydro-generator stator shaft system prediction and calibration technology. Background Technology

[0002] As the core equipment of a hydroelectric power generation system, the alignment accuracy of the stator shaft system of a hydro-generator directly affects the unit's operating efficiency and stability. Traditional shaft alignment verification methods mainly rely on manual inspection and experience-based judgment, which suffers from problems such as long inspection cycles and large subjective errors. With the development of automation technology, existing technologies have introduced shaft calibration systems based on laser ranging and vibration signal analysis. However, these systems still struggle to effectively handle the complex spatial relationship between the stator wiring leads and rotor components, resulting in limited verification accuracy.

[0003] Current automated verification systems mostly employ finite element analysis or traditional neural network algorithms, which suffer from the following technical shortcomings: 1) Traditional numerical analysis methods cannot dynamically characterize the topological connection characteristics between the stator lead-out terminals and adjacent components; 2) Conventional neural network models struggle to effectively extract the geometric correlation features between shaft components; 3) Existing systems do not fully consider the impact of the coupling effect of thermal deformation and mechanical stress on shaft alignment during unit operation. These shortcomings lead to discrepancies between verification results and actual operating conditions, potentially causing safety hazards such as bearing wear and excessive unit vibration. Overall, firstly, manual verification methods are inefficient and cannot meet the real-time monitoring needs of smart power plants; secondly, traditional algorithms suffer from modeling errors when dealing with complex spatial relationships at the stator lead-out terminals; and finally, existing intelligent verification systems lack adaptive adjustment mechanisms for dynamic operating parameters of the unit.

[0004] In summary, a neural network-based intelligent verification method and system for the stator shaft system diagram of a hydro-generator is needed. Summary of the Invention

[0005] A brief overview of the invention is given below to provide a basic understanding of certain aspects of it. It should be understood that this overview is not an exhaustive summary of the invention. It is not intended to identify key or essential parts of the invention, nor is it intended to limit the scope of the invention. Its purpose is merely to present certain concepts in a simplified form as a prelude to the more detailed description that follows.

[0006] In view of this, in order to solve the problems of insufficient accuracy and strong dependence on manual labor in the traditional hydro-generator stator shaft alignment verification method, the present invention provides a hydro-generator stator shaft alignment diagram neural network intelligent verification method and system.

[0007] Technical solution one is as follows: A neural network intelligent verification method for the stator shaft system diagram of a hydro-generator, comprising the following steps:

[0008] S1. The data acquisition module collects shaft vibration signals, displacement parameters, and temperature data in real time and transmits them to the data processing module;

[0009] S2. The collected physical parameters are converted into a graph structure through the data processing module, the graph structure is enhanced, and the enhanced graph structure is then fed into the graph neural network data structure.

[0010] S3. Through a graph neural network module containing a feature extraction layer based on an attention mechanism and a multilayer perceptron verification layer, the topological features of the axis alignment state are learned, and the final feature matrix is ​​input to the verification result output module.

[0011] S4. Based on the final feature matrix, the feature distribution of the real-time feature vector and the standard feature vector are compared through the intelligent verification module to generate the shaft alignment deviation evaluation result;

[0012] S5. Based on the deviation assessment results, trigger alarms or output correction suggestions through the intelligent operation and maintenance module.

[0013] Furthermore, step S1 includes the following steps:

[0014] S11. Install vibration sensors at the stator leads to measure shaft vibration signals in the 0-500Hz frequency range;

[0015] S12. Install non-contact displacement sensors and laser alignment instruments, and use the laser ranging principle to obtain the three-dimensional coordinate offset of the shaft components, i.e., the displacement parameters;

[0016] S13. Arrange the temperature sensing unit at the key contact surface of the shaft system to measure real-time temperature data;

[0017] S14. The shaft vibration signal, displacement parameters and temperature data obtained from the multi-source sensor group are transmitted to the feature extraction unit through the communication unit and stored through the storage unit.

[0018] Furthermore, in step S2, by configuring the feature extraction module of the vibration signal processing unit, temperature field analysis unit, and electromagnetic parameter detection unit, wavelet packet decomposition is performed on the shaft vibration signal, displacement parameters, and temperature data (i.e., the original signal) to obtain the wavelet packet decomposed time-frequency energy distribution information. The decomposition results are used to obtain node output feature vectors through feature mapping. , For node feature dimensions;

[0019] Wavelet packet decomposition results Represented as:

[0020]

[0021] in, For the first Layer Wavelet coefficients of nodes, For the mother wavelet function, This is the original vibration signal;

[0022] The graph structure construction module converts feature vectors into nodes and edges. Nodes represent shaft components, and edges represent the mechanical relationships between components. Node attributes include component mass. Stiffness and damping coefficient According to the connection stiffness between components Difference from actual displacement The edge weights are calculated. ;

[0023] Physical graph structure edge weights based on physical properties Represented as:

[0024]

[0025] in, This is the displacement sensitivity adjustment coefficient;

[0026] The graph structure building unit models the stator shaft system as a graph structure. Node set Includes mechanical components, edge sets Represents the connection relationships between components, edge weights It is calculated from the feature vectors of the nodes;

[0027] Feature map structure edge weights based on feature similarity Represented as:

[0028]

[0029] in, For nodes eigenvectors, For nodes eigenvectors, The Gaussian kernel width parameter;

[0030] The data augmentation unit outputs the augmented graph structure to the graph convolution unit.

[0031] Furthermore, step S3 includes the following steps:

[0032] S31. Based on the enhanced graph structure, by constructing graph convolution units with multiple layers of graph convolution operations, the convolutional node feature matrix is ​​output. And the node feature matrix after convolution Satisfies the matrix formula;

[0033] In S31, the formula for the multi-layer graph convolution matrix is ​​expressed as follows:

[0034]

[0035] in, , The total number of nodes. For feature dimension, To add self-connected adjacency matrices, It is an adjacency matrix. It is an N-order identity matrix. For degree matrix, For the first +1 layer node feature matrix For the first Layer node feature matrix, This is the degree matrix corresponding to the adjacency matrix. It is a trainable parameter matrix;

[0036] The graph convolutional unit adopts a three-layer GCN architecture:

[0037]

[0038] in, It is an adjacency matrix with self-connections. It is a degree matrix;

[0039] S32. A hybrid architecture is formed by combining the graph convolutional network layer of the graph convolutional unit with the graph attention layer of the attention mechanism unit, and the attention coefficients in the graph attention layer are used to... The node feature matrix after convolution The process is performed to obtain the final feature matrix, which is then input into the verification result output module via the decision unit.

[0040] In step S32, the attention coefficients are obtained by calculating the node importance weights through the attention mechanism units in the graph attention layer. ;

[0041] Attention coefficient Represented as:

[0042]

[0043] in, For attention vectors, For node features, It is the set of neighboring nodes.

[0044] Furthermore, in step S4, a dynamic benchmark model library is constructed through the report generation unit to store the standard feature vectors from step S3 under different operating conditions. ;

[0045] Real-time feature vectors are calculated by performing real-time feature comparison through visualization units. Compared with standard feature vectors cosine similarity :

[0046]

[0047] when At that time, the alarm unit triggers the three-level alarm strategy through the remote maintenance unit.

[0048] Furthermore, step S5 includes the following steps:

[0049] S51. Construct an axis alignment knowledge base containing 327 entities and 5128 relationships using knowledge graph units. When the axis alignment deviation exceeds a threshold... When the alarm unit triggers the alarm through the remote maintenance unit, the report generation unit automatically generates a verification report, which includes dynamic imbalance, phase difference and correction recommendation parameters.

[0050] S52. The decision support unit calculates the deviation of node characteristics through the output layer of the shaft system verification decision module, and uses a multi-objective optimization algorithm to solve for the optimal maintenance scheme;

[0051] In S52, the deviation of node features Represented as:

[0052]

[0053] Where MAX represents maximum pooling of node features, and MLP represents a multilayer perceptron. The number of layers in the multilayer perceptron;

[0054] The optimal maintenance plan is expressed as:

[0055]

[0056] in, To maintain costs, This refers to the downtime. For risk coefficient, This represents the feasible solution space.

[0057] S53. The remote maintenance unit performs alarms and maintenance based on thresholds, their deviations, and the optimal maintenance plan;

[0058] In S53, the remote maintenance unit is configured to: set a threshold, and in the first-level alarm, the threshold corresponds to the deviation amount. Only logs are recorded; the deviation corresponding to the threshold in the level 2 alarm is recorded. This triggers an audible and visual alarm; the threshold value corresponds to the deviation in a level three alarm system. Interlocked shutdown protection;

[0059] S54. In response to the alarm, based on the node characteristic deviation output by the decision support unit in step S52, the parameter correction suggestion generation unit establishes a correction amount mapping model based on historical maintenance data and outputs the optimized alignment parameter suggestion to the remote maintenance unit.

[0060]

[0061] in, For the feature deviation vector, For ambient temperature, This is the current rotational speed.

[0062] Technical Solution 2 is as follows: A graph neural network intelligent verification system for the stator shaft system of a hydro-generator, used to execute the graph neural network intelligent verification method for the stator shaft system of a hydro-generator described in Technical Solution 1, comprising a data acquisition module, a data processing module, a graph neural network analysis module, a verification result output module, and an intelligent operation and maintenance module connected in sequence.

[0063] Furthermore, the data acquisition module includes a multi-source sensor group, a communication unit, and a storage unit connected in sequence. The multi-source sensor group includes a vibration sensor, a temperature sensor, a displacement sensor, and a laser alignment instrument connected in sequence.

[0064] The data processing module includes a feature extraction unit, a graph structure construction unit, and a data enhancement unit connected in sequence.

[0065] The graph neural network analysis module includes a graph convolutional unit, an attention mechanism unit, and a decision unit connected in sequence.

[0066] The verification result output module includes a visualization unit, an alarm unit, and a report generation unit;

[0067] The intelligent operation and maintenance module includes a knowledge graph unit, a decision support unit, and a remote maintenance unit connected in sequence.

[0068] The communication unit and storage unit are respectively connected to the feature extraction unit, the data augmentation unit is connected to the graph convolution unit, the visualization unit, the alarm unit and the report generation unit are respectively connected to the decision unit, the knowledge graph unit, the decision support unit and the remote maintenance unit are respectively connected to the report generation unit, and the alarm unit is connected to the remote maintenance unit.

[0069] The beneficial effects of this invention are as follows: This invention deploys a multi-modal sensor array at the stator lead-out end of the hydro-generator to collect real-time shaft alignment state parameters and electromagnetic field distribution characteristics; constructs a dynamic topological relationship model based on a graph neural network to establish a mapping between the spatial coordinates of the terminals and electromagnetic parameters; employs a multi-scale feature fusion algorithm to iteratively optimize and calculate shaft offset, achieving intelligent identification of shaft alignment deviation with millimeter-level accuracy; and develops a self-learning verification decision module to generate a comprehensive operation and maintenance strategy including vibration suppression schemes and insulation optimization suggestions. This system combines traditional mechanical calibration methods with deep learning technology, effectively solving the technical problems of low shaft alignment detection efficiency and error accumulation in large-size hydro-generator units, significantly improving equipment operational stability and reducing maintenance costs.

[0070] This invention covers dynamic monitoring of the stator structure of hydro-generators, shaft alignment status analysis, and the application of intelligent algorithms in the operation and maintenance of electromechanical equipment. Specifically, it includes the following technical directions: vibration signal feature extraction technology from the stator connection leads of hydro-generators; a shaft dynamic relationship modeling method based on graph neural networks; intelligent diagnosis technology for shaft alignment deviation under multi-physics coupling conditions; and a method for constructing an intelligent operation and maintenance system integrating equipment status data and topology. This patented technology also involves the fusion application of hydro-generator stator structure design in the field of electrical engineering and graph neural network algorithms in the field of computer science. The core of this invention lies in constructing a graph structure representation model of the spatial relationship of the stator connection leads, and realizing intelligent prediction and deviation verification of shaft alignment parameters through multi-dimensional feature fusion and deep learning. Attached Figure Description

[0071] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0072] Figure 1 This is a flowchart illustrating a neural network-based intelligent verification method for the stator shaft system diagram of a hydro-generator.

[0073] Figure 2 This is a schematic diagram of a neural network intelligent verification system for the stator shaft system of a hydro-generator.

[0074] Figure Descriptions: 100. Data Acquisition Module; 200. Data Processing Module; 300. Graph Neural Network Analysis Module; 400. Verification Result Output Module; 500. Intelligent Operation and Maintenance Module; 110. Multi-Source Sensor Group; 120. Communication Unit; 130. Storage Unit; 111. Vibration Sensor; 112. Temperature Sensor; 113. Displacement Sensor; 114. Laser Alignment Instrument; 210. Feature Extraction Unit; 220. Graph Structure Construction Unit; 230. Data Augmentation Unit; 310. Graph Convolution Unit; 320. Attention Mechanism Unit; 330. Decision Unit; 410. Visualization Unit; 420. Alarm Unit; 430. Report Generation Unit; 510. Knowledge Graph Unit; 520. Decision Support Unit; 530. Remote Maintenance Unit. Detailed Implementation

[0075] To make the technical solutions and advantages of the embodiments of the present invention clearer, the exemplary embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0076] Example 1: Reference Figures 1-2 This embodiment describes a neural network-based intelligent verification method for the stator shaft system diagram of a hydro-generator, which specifically includes the following steps:

[0077] S1. The data acquisition module collects shaft vibration signals, displacement parameters, and temperature data in real time and transmits them to the data processing module;

[0078] S2. The collected physical parameters are converted into a graph structure through the data processing module, the graph structure is enhanced, and the enhanced graph structure is then fed into the graph neural network data structure.

[0079] S3. Through a graph neural network module containing a feature extraction layer based on an attention mechanism and a multilayer perceptron verification layer, the topological features of the axis alignment state are learned, and the final feature matrix is ​​input to the verification result output module.

[0080] S4. Based on the final feature matrix, the feature distribution of the real-time feature vector and the standard feature vector are compared through the intelligent verification module to generate the shaft alignment deviation evaluation result;

[0081] S5. Based on the deviation assessment results, trigger alarms or output correction suggestions through the intelligent operation and maintenance module.

[0082] Furthermore, step S1 includes the following steps:

[0083] S11. Install vibration sensors at the stator leads to measure shaft vibration signals in the 0-500Hz frequency range;

[0084] S12. Install non-contact displacement sensors and laser alignment instruments, and use the laser ranging principle to obtain the three-dimensional coordinate offset of the shaft components, i.e., the displacement parameters;

[0085] S13. Arrange the temperature sensing unit at the key contact surface of the shaft system to measure real-time temperature data;

[0086] S14. The shaft vibration signal, displacement parameters and temperature data obtained from the multi-source sensor group are transmitted to the feature extraction unit through the communication unit and stored through the storage unit.

[0087] Furthermore, in step S2, by configuring the feature extraction module of the vibration signal processing unit, temperature field analysis unit, and electromagnetic parameter detection unit, wavelet packet decomposition is performed on the shaft vibration signal, displacement parameters, and temperature data (i.e., the original signal) to obtain the wavelet packet decomposed time-frequency energy distribution information. The decomposition results are used to obtain node output feature vectors through feature mapping. , For node feature dimensions;

[0088] Wavelet packet decomposition results Represented as:

[0089]

[0090] in, For the first Layer Wavelet coefficients of nodes, For the mother wavelet function, This is the original vibration signal;

[0091] The graph structure construction module converts feature vectors into nodes and edges. Nodes represent shaft components, and edges represent the mechanical relationships between components. Node attributes include component mass. Stiffness and damping coefficient According to the connection stiffness between components Difference from actual displacement The edge weights are calculated. ;

[0092] Physical graph structure edge weights based on physical properties (stiffness, displacement) Represented as:

[0093]

[0094] in, This is the displacement sensitivity adjustment coefficient;

[0095] The graph structure building unit models the stator shaft system as a graph structure. Node set Includes mechanical components such as shaft segments, bearings, and couplings, and is part of a conglomerate. Represents the connection relationships between components, edge weights It is calculated from the feature vectors of the nodes;

[0096] Feature graph structure edge weights based on feature similarity (learned feature vectors) Represented as:

[0097]

[0098] in, For nodes eigenvectors, For nodes eigenvectors, The Gaussian kernel width parameter;

[0099] The data augmentation unit outputs the augmented graph structure to the graph convolution unit.

[0100] Specifically, the data augmentation unit 230 applies three data augmentation methods: random rotation, additive noise, and time warp, with the augmentation coefficient set to 0.2.

[0101] Furthermore, step S3 includes the following steps:

[0102] S31. Based on the enhanced graph structure, by constructing graph convolution units with multiple layers of graph convolution operations, the convolutional node feature matrix is ​​output. And the node feature matrix after convolution Satisfies the matrix formula;

[0103] In S31, the formula for the multi-layer graph convolution matrix is ​​expressed as follows:

[0104]

[0105] in, , The total number of nodes. For feature dimension, To add self-connected adjacency matrices, It is an adjacency matrix. It is an N-order identity matrix. For degree matrix, For the first +1 layer node feature matrix For the first Layer node feature matrix, This is the degree matrix corresponding to the adjacency matrix. It is a trainable parameter matrix;

[0106] The graph convolutional unit adopts a three-layer GCN architecture:

[0107]

[0108] in, It is an adjacency matrix with self-connections. It is a degree matrix;

[0109] The graph convolutional unit 310 adopts a three-layer GCN architecture:

[0110]

[0111] in, It is an adjacency matrix with self-connections. For degree matrix, This is a trainable parameter matrix.

[0112] S32. A hybrid architecture is formed by combining the graph convolutional network layer of the graph convolutional unit with the graph attention layer of the attention mechanism unit, and the attention coefficients in the graph attention layer are used to... The node feature matrix after convolution The process is performed to obtain the final feature matrix, which is then input into the verification result output module via the decision unit.

[0113] In step S32, the attention coefficients are obtained by calculating the node importance weights through the attention mechanism units in the graph attention layer. ;

[0114] Attention coefficient Represented as:

[0115]

[0116] in, For attention vectors, For node features, It is the set of neighboring nodes.

[0117] Furthermore, in step S4, a dynamic benchmark model library is constructed through the report generation unit to store the standard feature vectors from step S3 under different operating conditions. ;

[0118] Real-time feature vectors are calculated by performing real-time feature comparison through visualization units. Compared with standard feature vectors cosine similarity :

[0119]

[0120] when At that time, the alarm unit triggers the three-level alarm strategy through the remote maintenance unit.

[0121] Furthermore, step S5 includes the following steps:

[0122] S51. Construct an axis alignment knowledge base containing 327 entities and 5128 relationships using knowledge graph units. When the axis alignment deviation exceeds a threshold... When the alarm unit triggers an alarm (level 3 warning) through the remote maintenance unit, the report generation unit automatically generates a verification report that conforms to the IEC-60034 standard, which includes dynamic imbalance, phase difference and correction recommendation parameters;

[0123] S52. The decision support unit calculates the deviation of node characteristics through the output layer of the shaft system verification decision module, and uses a multi-objective optimization algorithm to solve for the optimal maintenance scheme;

[0124] In S52, the deviation of node features Represented as:

[0125]

[0126] Where MAX represents maximum pooling of node features, and MLP represents a multilayer perceptron. The number of layers in the multilayer perceptron;

[0127] The optimal maintenance plan is expressed as:

[0128]

[0129] in, To maintain costs, This refers to the downtime. For risk coefficient, This represents the feasible solution space.

[0130] S53. The remote maintenance unit performs alarms and maintenance based on thresholds, their deviations, and the optimal maintenance plan;

[0131] In S53, the remote maintenance unit is configured to: set a threshold, and in the first-level alarm, the threshold corresponds to the deviation amount. Only logs are recorded; the deviation corresponding to the threshold in the level 2 alarm is recorded. This triggers an audible and visual alarm; the threshold value corresponds to the deviation in a level three alarm system. Interlocked shutdown protection;

[0132] S54. In response to the alarm, based on the node characteristic deviation output by the decision support unit in step S52, the parameter correction suggestion generation unit establishes a correction amount mapping model based on historical maintenance data and outputs the optimized alignment parameter suggestion to the remote maintenance unit.

[0133]

[0134] in, For the feature deviation vector, For ambient temperature, This is the current rotational speed.

[0135] Specifically, the training phase of the modified quantity mapping model uses the Adam optimizer, with an initial learning rate of 0.001, 500 training cycles, a batch size of 32, and training data containing 21,740 samples under 12 typical working conditions.

[0136] Example 2: Reference Figure 2 This embodiment describes a hydro-generator stator shaft system graph neural network intelligent verification system, used to execute the hydro-generator stator shaft system graph neural network intelligent verification method described in Embodiment 1. The system includes a data acquisition module 100, a data processing module 200, a graph neural network analysis module 300, a verification result output module 400, and an intelligent operation and maintenance module 500 connected in sequence.

[0137] Furthermore, the data acquisition module 100 includes a multi-source sensor group 110, a communication unit 120, and a storage unit 130 connected in sequence. The multi-source sensor group 110 includes a vibration sensor 111, a temperature sensor 112, a displacement sensor 113, and a laser alignment instrument 114 arranged in sequence at the stator wiring lead-out end.

[0138] The data processing module 200 includes a feature extraction unit 210, a graph structure construction unit 220, and a data augmentation unit 230 connected in sequence.

[0139] The graph neural network analysis module 300 includes a graph convolution unit 310, an attention mechanism unit 320, and a decision unit 330 connected in sequence.

[0140] The verification result output module 400 includes a visualization unit 410, an alarm unit 420, and a report generation unit 430;

[0141] The intelligent operation and maintenance module 500 includes a knowledge graph unit 510, a decision support unit 520, and a remote maintenance unit 530 connected in sequence.

[0142] The communication unit 120 and storage unit 130 are respectively connected to the feature extraction unit 210, the data augmentation unit 230 is connected to the graph convolution unit 310, the visualization unit 410, the alarm unit 420 and the report generation unit 430 are respectively connected to the decision unit 330, the knowledge graph unit 510, the decision support unit 520 and the remote maintenance unit 530 are respectively connected to the report generation unit 430, and the alarm unit 420 is connected to the remote maintenance unit 530.

[0143] Specifically, the modules interact with each other via an industrial bus, and the system is deployed on an edge computing device and a cloud server collaborative platform;

[0144] Communication unit 120 adopts a dual redundancy design of RS-485 and CAN bus, with a transmission rate of not less than 10Mbps;

[0145] Storage unit 130 is equipped with 64GB of DDR4 memory and a 1TB NVMe solid-state drive.

[0146] Although the invention has been described with reference to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments are conceivable within the scope of the invention described herein. Furthermore, it should be noted that the language used in this specification has been chosen primarily for readability and instructional purposes, and not for the purpose of interpreting or limiting the subject matter of the invention. Therefore, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the invention is illustrative and not restrictive, and the scope of the invention is defined by the appended claims.

Claims

1. A neural network-based intelligent verification method for the stator shaft system diagram of a hydro-generator, characterized in that, Includes the following steps: S1. The data acquisition module collects shaft vibration signals, displacement parameters, and temperature data in real time and transmits them to the data processing module; S2. The collected physical parameters are converted into a graph structure through the data processing module, the graph structure is enhanced, and the enhanced graph structure is then converted into a graph neural network data structure. S3. Through a graph neural network module containing a feature extraction layer based on an attention mechanism and a multilayer perceptron verification layer, the topological features of the axis alignment state are learned, and the final feature matrix is ​​input to the verification result output module. S4. Based on the final feature matrix, the feature distribution of the real-time feature vector and the standard feature vector are compared through the intelligent verification module to generate the shaft alignment deviation evaluation result; S5. Based on the deviation assessment results, trigger alarms or output correction suggestions through the intelligent operation and maintenance module.

2. The neural network intelligent verification method for the stator shaft system diagram of a hydro-generator according to claim 1, characterized in that, S1 includes the following steps: S11. Install vibration sensors at the stator lead-out terminals to measure shaft vibration signals in the frequency range of 0-500Hz; S12. Install non-contact displacement sensors and laser alignment instruments, and use the laser ranging principle to obtain the three-dimensional coordinate offset of the shaft components, i.e., the displacement parameters; S13. Arrange the temperature sensing unit at the key contact surface of the shaft system to measure real-time temperature data; S14. The shaft vibration signal, displacement parameters and temperature data obtained from the multi-source sensor group are transmitted to the feature extraction unit through the communication unit and stored through the storage unit.

3. The neural network intelligent verification method for the stator shaft system diagram of a hydro-generator according to claim 2, characterized in that, In step S2, by configuring a feature extraction module that includes a vibration signal processing unit, a temperature field analysis unit, and an electromagnetic parameter detection unit, wavelet packet decomposition is performed on the shaft vibration signal, displacement parameters, and temperature data (i.e., the original signal) to obtain the wavelet packet decomposed time-frequency energy distribution information. The decomposition results are used to obtain the node output feature vectors through feature mapping. , For node feature dimensions; Wavelet packet decomposition results Represented as: in, For the first Layer Wavelet coefficients of nodes, For the mother wavelet function, This is the original vibration signal; The graph structure construction module converts feature vectors into nodes and edges. Nodes represent shaft components, and edges represent the mechanical relationships between components. Node attributes include component mass. Stiffness and damping coefficient According to the connection stiffness between components Difference from actual displacement The edge weights are calculated. ; Physical graph structure edge weights based on physical properties Represented as: in, This is the displacement sensitivity adjustment coefficient; The graph structure building unit models the stator shaft system as a graph structure. Node set Includes mechanical components, edge sets Represents the connection relationships between components, edge weights It is calculated from the feature vectors of the nodes; Feature map structure edge weights based on feature similarity Represented as: in, For nodes eigenvectors, For nodes eigenvectors, The Gaussian kernel width parameter; The data augmentation unit outputs the augmented graph structure to the graph convolution unit.

4. The neural network intelligent verification method for the stator shaft system diagram of a hydro-generator according to claim 3, characterized in that, S3 includes the following steps: S31. Based on the enhanced graph structure, by constructing graph convolution units with multiple layers of graph convolution operations, the convolutional node feature matrix is ​​output. And the node feature matrix after convolution Satisfies the matrix formula; In S31, the formula for the multi-layer graph convolution matrix is ​​expressed as follows: in, , The total number of nodes. For feature dimension, To add self-connected adjacency matrices, It is an adjacency matrix. It is an N-order identity matrix. For degree matrix, For the first +1 layer node feature matrix For the first Layer node feature matrix This is the degree matrix corresponding to the adjacency matrix. It is a trainable parameter matrix; The graph convolutional unit adopts a three-layer GCN architecture: in, It is an adjacency matrix with self-connections. It is a degree matrix; S32. A hybrid architecture is formed by combining the graph convolutional network layer of the graph convolutional unit with the graph attention layer of the attention mechanism unit, and the attention coefficients in the graph attention layer are used to... The node feature matrix after convolution The process is performed to obtain the final feature matrix, which is then input into the verification result output module via the decision unit. In step S32, the attention coefficients are obtained by calculating the node importance weights through the attention mechanism units in the graph attention layer. ; Attention coefficient Represented as: in, For attention vectors, For node features, It is the set of neighboring nodes.

5. The neural network intelligent verification method for the stator shaft system diagram of a hydro-generator according to claim 4, characterized in that, In step S4, a dynamic benchmark model library is constructed through the report generation unit to store the standard feature vectors from step S3 under different operating conditions. ; Real-time feature vectors are calculated by performing real-time feature comparison through visualization units. Compared with standard feature vectors cosine similarity : when At that time, the alarm unit triggers the three-level alarm strategy through the remote maintenance unit.

6. The neural network intelligent verification method for the stator shaft system diagram of a hydro-generator according to claim 5, characterized in that, S5 includes the following steps: S51. Construct an axis alignment knowledge base containing 327 entities and 5128 relationships using knowledge graph units. When the axis alignment deviation exceeds a threshold... When the alarm unit triggers the alarm through the remote maintenance unit, the report generation unit automatically generates a verification report, which includes dynamic imbalance, phase difference and correction recommendation parameters. S52. The decision support unit calculates the deviation of node characteristics through the output layer of the axis system verification decision module, and uses a multi-objective optimization algorithm to solve for the optimal maintenance scheme; In S52, the deviation of node features Represented as: Where MAX represents maximum pooling of node features, and MLP represents a multilayer perceptron. The number of layers in the multilayer perceptron; The optimal maintenance plan is expressed as: in, To maintain costs, This refers to the downtime. For risk coefficient, This represents the feasible solution space. S53. The remote maintenance unit performs alarms and maintenance based on thresholds, their deviations, and the optimal maintenance plan; In S53, the remote maintenance unit is configured to: set a threshold, and in the first-level alarm, the threshold corresponds to the deviation amount. Only logs are recorded; the deviation corresponding to the threshold in the level 2 alarm is recorded. This triggers an audible and visual alarm; the threshold value corresponds to the deviation in a level three alarm system. Interlocked shutdown protection; S54. In response to the alarm, based on the node characteristic deviation output by the decision support unit in step S52, the parameter correction suggestion generation unit establishes a correction amount mapping model based on historical maintenance data and outputs the optimized alignment parameter suggestion to the remote maintenance unit. in, For the feature deviation vector, For ambient temperature, This is the current rotational speed.

7. A neural network intelligent verification system for the stator shaft system diagram of a hydro-generator, characterized in that, The method for performing the intelligent verification method of the stator shaft system of a hydro-generator according to any one of claims 1-6 includes a data acquisition module (100), a data processing module (200), a graph neural network analysis module (300), a verification result output module (400), and an intelligent operation and maintenance module (500) connected in sequence.

8. The neural network intelligent verification system for the stator shaft system diagram of a hydro-generator according to claim 7, characterized in that, The data acquisition module (100) includes a multi-source sensor group (110), a communication unit (120) and a storage unit (130) connected in sequence. The multi-source sensor group (110) includes a vibration sensor (111), a temperature sensor (112), a displacement sensor (113) and a laser alignment instrument (114) connected in sequence. The data processing module (200) includes a feature extraction unit (210), a graph structure construction unit (220), and a data enhancement unit (230) connected in sequence. The graph neural network analysis module (300) includes a graph convolution unit (310), an attention mechanism unit (320), and a decision unit (330) connected in sequence. The verification result output module (400) includes a visualization unit (410), an alarm unit (420), and a report generation unit (430). The intelligent operation and maintenance module (500) includes a knowledge graph unit (510), a decision support unit (520), and a remote maintenance unit (530) connected in sequence. The communication unit (120) and storage unit (130) are respectively connected to the feature extraction unit (210), the data augmentation unit (230) is connected to the graph convolution unit (310), the visualization unit (410), the alarm unit (420) and the report generation unit (430) are respectively connected to the decision unit (330), the knowledge graph unit (510), the decision support unit (520) and the remote maintenance unit (530) are respectively connected to the report generation unit (430), and the alarm unit (420) is connected to the remote maintenance unit (530).

Images