Shafting vibration fault positioning method and system for complex working conditions

Abstract

The invention belongs to the technical field of fault positioning, and discloses a shafting vibration fault positioning method and system for a complex working condition, which realizes high-dimensional, global and nonlinear feature extraction and fault positioning of vibration signals under the complex working condition by introducing a graph neural network and a message transmission mechanism. The problems that a traditional method depends on a single amplitude characteristic, complex nonlinear vibration characteristics cannot be accurately captured and the like are remarkably solved, the precision and reliability of shafting vibration fault positioning are greatly improved, and the method has remarkable innovativeness and practical application value.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault location, and in particular relates to a method and system for locating shaft vibration faults in complex working conditions. Background Art

[0002] In modern industrial production, the shafting of mechanical equipment plays a critical role in transmission and support. Locating shafting vibration faults is particularly important under complex operating conditions. These conditions often involve factors such as variable loads, speeds, ambient temperatures, and harsh operating environments. These intertwined factors greatly increase the difficulty of diagnosing and locating shafting vibration faults.

[0003] With the increasing degree of industrial automation, the continuous operation of mechanical equipment is crucial to production efficiency and safety. Under complex operating conditions, if shaft vibration faults cannot be accurately and promptly located, they can lead to sudden equipment failures, resulting in significant economic losses and safety hazards. For example, in large wind turbines, vibration faults in the gearbox and generator shafts are difficult to locate. Once a fault occurs, downtime for repair is not only costly but also impacts the stability of the power grid.

[0004] Chinese patent CN118500667B discloses a method for quickly locating vibration faults in rotating machinery. The method obtains vibration data of a monitoring node of the rotating machinery, wherein the vibration data includes the amplitude of the monitoring node. The method determines whether a vibration abnormality occurs at the monitoring node based on the vibration data of the monitoring node. When a vibration abnormality occurs at the monitoring node, the method uses a first vibration abnormality diagnosis formula to locate the cause of the abnormality. The method can quickly locate the fault direction when a vibration abnormality occurs in the rotating machinery. However, the feature extraction dimension of the existing method is single, relying only on the amplitude data of the monitoring node (a single feature in the time domain), and is unable to capture the nonlinear vibration characteristics under complex working conditions. As a result, the fault location accuracy under complex working conditions deviates greatly from the actual value. To address the above problems, we propose a shaft system vibration fault location method and system for complex working conditions. Summary of the Invention

[0005] The purpose of the present invention is to overcome the problem that the existing methods have a single feature extraction dimension, rely only on the amplitude data of the monitoring nodes, cannot capture the nonlinear vibration characteristics under complex working conditions, and cause a large deviation between the fault location accuracy and the actual under complex working conditions, and provide a shaft system vibration fault location method and system for complex working conditions.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a method for locating shaft vibration faults in complex working conditions, comprising the following steps: Obtain historical working signals under target working conditions and build a vibration propagation database based on the historical working signals; A pre-built graph neural network model is used to extract feature vectors of historical working signals in the vibration propagation database. Based on the eigenvectors, a fault location-eigenvector mapping library based on the message passing mechanism is constructed. The fault source is traced back based on the message passing mechanism combined with the transfer path analysis. The inverse operation is used to solve the excitation source location. The physical constraint loss function is used to determine whether the excitation source location meets the loss threshold. If it meets the loss threshold, the fault location-eigenvector mapping library based on the message passing mechanism combined with the transfer path analysis is output. Acquire real-time working signals and use a graph neural network model to extract their feature vectors. Then, use a message passing mechanism combined with a fault location-feature vector mapping library based on transfer path analysis to perform correlation analysis on the feature vectors of the real-time working signals and determine the probability distribution of the fault location. According to the probability distribution of the fault location, a heat map is used to visualize the fault location in the three-dimensional shafting model.

[0007] A further improvement of the present invention is that the graph neural network model includes a spectral domain convolution GCN layer, an attention layer, a physical constraint module and an axis topology module. The axis topology module constructs an axis topology graph based on the signal feature vector. The nodes in the axis topology graph are component positions, and the connecting edges are signal feature vectors. The pooling layer of the graph neural network model is frozen and replaced by a migration convolutional network architecture.

[0008] A further improvement of the present invention is that the physical constraint loss function of the graph neural network model is expressed as:

[0009] in, Represent the weights of the data-driven loss term and the physical constraint loss term, respectively. represents the mean square error loss function, Indicates the transfer function gradient and the location of the excitation source for the inverse operation solution, represents the transfer function calculated based on the Jeffcott rotor model.

[0010] A further improvement of the present invention is that when the axis topology module constructs the axis topology graph based on the signal eigenvector, the edge weights are adjusted by constructing an adaptive adjacency matrix between nodes. The adaptive adjacency matrix is ​​expressed as:

[0011] in, represents the activation function, represents a multilayer perceptron, Represents nodes respectively ,node exist The feature vector of the moment.

[0012] A further improvement of the present invention is that the specific method for tracing back the fault source based on the message transmission mechanism combined with transmission path analysis is as follows: Map the historical working signal feature vectors to the fault location-feature vector mapping library to simulate the propagation process of the vibration signal in the shafting structure; According to the weights of the nodes and connecting edges in the graph neural network model, the historical working signal feature vector is transmitted along the transmission path to build a mapping relationship between the fault location and the node; The mapping relationship between the loading fault location and the node is constructed, and the shaft system transfer function matrix is ​​constructed based on the transfer path analysis. The main and secondary paths of signal propagation are identified, and the path analysis results including the main and secondary paths are obtained.

[0013] A further improvement of the present invention is that after obtaining the path analysis results including the main path and the secondary path, the shaft system transfer function matrix is ​​flipped, so that the historical working signal eigenvector is propagated back to the potential fault location, the vibration response, the shaft system transfer function matrix and the physical constraints are loaded, the fault range is narrowed by solving the excitation source through pseudo-inverse, and the fault source with the narrowed range is output.

[0014] A further improvement of the present invention is that the fault range is narrowed by solving the excitation source through pseudo-inverse, and the specific method of outputting the fault source with a narrowed range is as follows:

[0015]

[0016] in, represents the axis transfer function matrix, represents the physical constraint function, is the number of constraints, Representation node The stiffness parameter, Indicates the source of the fault Associated Nodes the number of represents the vibration response, Indicates the fault type label, is the cross entropy loss function, Node The historical working signal feature vector and fault source Associated Nodes The signal feature vector.

[0017] A further improvement of the present invention is that the method for obtaining a real-time working signal and extracting a feature vector of the real-time working signal using a graph neural network model is as follows: Acquire a real-time working signal, adapt and normalize the format of the real-time working signal, and obtain a normalized set; The normalized set is analyzed using spectral domain convolution operation to obtain spectral domain analysis results. The Laplace matrix of the normalized set is calculated to implement normalized set signal filtering and feature extraction, and output the key feature set.

[0018] A further improvement of the present invention is that the specific method of performing correlation analysis on the feature vectors of real-time working signals by using a message transmission mechanism combined with a fault location-feature vector mapping library of transmission path analysis is as follows: Assign weights to key features in the key feature set, modify the key features, and fuse and enhance the modified key feature vectors to obtain a feature fusion vector; Combining the spectral domain analysis results and dilated causal convolution, the feature fusion vector is analyzed to capture the long-term dependency between the fault location and the feature vector.

[0019] In a second aspect, the present invention provides a shafting vibration fault location method system for complex working conditions, comprising: A historical signal acquisition module is used to obtain historical working signals under target working conditions and build a vibration propagation database based on the historical working signals; A vector extraction module is used to pre-build a graph neural network model and use it to extract feature vectors of historical working signals in the vibration propagation database; The fault-vector mapping module is used to construct a fault location-feature vector mapping library based on the message passing mechanism based on the feature vector. The fault source is traced back based on the message passing mechanism combined with the transfer path analysis. The inverse operation is used to solve the excitation source location. The physical constraint loss function is used to determine whether the excitation source location meets the loss threshold. If the loss threshold is met, the fault location-feature vector mapping library based on the message passing mechanism combined with the transfer path analysis is output. The fault point determination module is used to obtain real-time working signals, extract the feature vectors of the real-time working signals using a graph neural network model, and perform correlation analysis on the feature vectors of the real-time working signals using a message passing mechanism combined with a fault location-feature vector mapping library based on transfer path analysis to determine the probability distribution of the fault location. The visualization module is used to visualize the fault location points using a heat map in the three-dimensional shafting model based on the probability distribution of the fault location.

[0020] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a vibration propagation database by collecting and archiving historical operating signals under target operating conditions on a large scale. This database enables unified management and representation of multi-operating-condition, multi-node, and multi-dimensional data, enriching the data foundation and supporting subsequent deep feature extraction. Based on the shafting structure, the physical connections between nodes, and the vibration propagation path, this invention constructs a graph neural network model for graph-structured data input. Using GNNs, this model extracts high-dimensional, nonlinear features from complex network relationships, fully exploiting the dynamic connections and global structural information between nodes. Incorporating the GNN's message-passing mechanism, this method not only captures local information at each monitoring point but also enables the spatial flow and transmission of fault information through the network structure, tracing back the true excitation path of the fault source. The optimal excitation source location is selected through inverse operations and a physical constraint loss function, effectively constraining positioning errors. This invention establishes a mapping relationship between fault location and feature vector, forming a reusable and generalizable knowledge base, enabling efficient matching and probabilistic judgment of real-time signals and historical features. The present invention displays the positioning results as a heat map within a three-dimensional shafting model, visually displaying the probability distribution of the fault point, facilitating operational and maintenance decision-making and subsequent processing. In summary, the present invention realizes high-dimensional, global, nonlinear feature extraction and fault location of vibration signals under complex working conditions by introducing graph neural networks and message passing mechanisms. It significantly overcomes the problems of traditional reliance on single amplitude features and inability to accurately capture complex nonlinear vibration characteristics, greatly improves the accuracy and reliability of shaft vibration fault location, and has significant innovation and practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flow chart of the present invention; Figure 2 is a system diagram of the present invention; Figure 3 This is a schematic flow diagram of Example 1; Figure 4 A flowchart for implementing a method for tracing back the source of a fault through a message passing mechanism combined with transmission path analysis; Figure 5 A flowchart illustrating a method for extracting features from real-time working signals based on a graph neural network model. Figure 6 This is a schematic structural diagram of Example 4; Figure 7 This is a structural diagram of the obstacle visualization module. DETAILED DESCRIPTION

[0022] In order to further understand the content of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the embodiments are only for explaining the present invention and are not intended to limit it.

[0023] See also Figure 1, a shafting vibration fault location method for complex working conditions, comprising the following steps: S1, obtain the historical working signals under the target working conditions and build a vibration propagation database based on the historical working signals.

[0024] S2, pre-builds a graph neural network model and uses the graph neural network model to extract the feature vectors of historical working signals in the vibration propagation database.

[0025] S3, based on the feature vector, constructs a fault location-feature vector mapping library based on the message passing mechanism. Based on the message passing mechanism combined with the transfer path analysis, the fault source is traced back. The inverse operation is used to solve the excitation source position. The physical constraint loss function is used to determine whether the excitation source position meets the loss threshold. If it meets the loss threshold, the fault location-feature vector mapping library based on the message passing mechanism combined with the transfer path analysis is output.

[0026] S4, obtains real-time working signals, uses a graph neural network model to extract the feature vectors of the real-time working signals, and uses a message passing mechanism combined with a fault location-feature vector mapping library based on transmission path analysis to perform correlation analysis on the feature vectors of the real-time working signals to determine the probability distribution of the fault location.

[0027] S5, based on the probability distribution of the fault location, a heat map is used to visualize the fault location in the 3D shafting model.

[0028] See also Figure 2 , a shafting vibration fault location method system for complex working conditions, including: A historical signal acquisition module is used to obtain historical working signals under target working conditions and build a vibration propagation database based on the historical working signals; A vector extraction module is used to pre-build a graph neural network model and use it to extract feature vectors of historical working signals in the vibration propagation database; The fault-vector mapping module is used to construct a fault location-feature vector mapping library based on the message passing mechanism based on the feature vector. The fault source is traced back based on the message passing mechanism combined with the transfer path analysis. The inverse operation is used to solve the excitation source location. The physical constraint loss function is used to determine whether the excitation source location meets the loss threshold. If the loss threshold is met, the fault location-feature vector mapping library based on the message passing mechanism combined with the transfer path analysis is output. The fault point determination module is used to obtain real-time working signals, extract the feature vectors of the real-time working signals using a graph neural network model, and perform correlation analysis on the feature vectors of the real-time working signals using a message passing mechanism combined with a fault location-feature vector mapping library based on transfer path analysis to determine the probability distribution of the fault location. The visualization module is used to visualize the fault location points using a heat map in the three-dimensional shafting model based on the probability distribution of the fault location.

[0029] Example 1: See also Figure 3 This embodiment provides a shafting vibration fault location method for complex working conditions. Figure 3 A schematic diagram of a shafting vibration fault location method for complex working conditions is shown. The shafting vibration fault location method for complex working conditions specifically includes: Step S10: constructing a vibration propagation database based on historical working signals under complex working conditions, and extracting feature vectors of historical working signals using a pre-built graph neural network model; It should be noted that historical operating signals include basic vibration signal data, operating parameter data, and fault tag information. Basic vibration signal data may include acceleration, velocity, and displacement data, reflecting the vibration intensity and frequency changes of the shaft system during operation. Operating parameter data includes, but is not limited to, equipment speed, load, temperature, and pressure. Fault tag information includes, but is not limited to, fault type, fault occurrence time, and fault component location. Historical operating signals can be acquired via vibration sensors (such as accelerometers, velocity sensors, etc.), temperature sensors, pressure sensors, and speed sensors. The vibration propagation database can be stored in a relational database (such as MySQL, SQL Server, etc.) or a non-relational database (such as MongoDB, HBase, etc.). An appropriate database table structure or document structure should be designed based on the data characteristics and query requirements. It should be noted that in this embodiment, shaft system vibration refers to the mechanical vibration phenomenon occurring in the shaft system's supporting structure (such as bearings and bushings). It is the dynamic response of the shaft system during operation due to the stimulation of various internal and external factors.

[0030] Step S20: Load the historical working signal feature vectors, build a fault location-feature vector mapping library based on the message passing mechanism, trace the fault source in reverse through the message passing mechanism combined with the transfer path analysis, and solve the excitation source location through inverse calculation; Step S30: Using the physical constraint loss function to determine whether the excitation source position meets the loss threshold, if so, outputting a fault location-feature vector mapping library based on the message delivery mechanism combined with the transfer path analysis. It should be noted that the loss threshold can be 0.01-0.08. If the loss threshold is not met, the fault source is traced back through the message transmission mechanism combined with the transmission path analysis.

[0031] In step S40, a real-time working signal is obtained, and features of the real-time working signal are extracted based on a graph neural network model to output a real-time feature vector. The real-time feature vector is correlated and analyzed based on a fault location-feature vector mapping library to determine the probability distribution of the fault location. The fault location point is visualized using a heat map in the three-dimensional shafting model.

[0032] In this embodiment, a pre-built graph neural network model is used to extract the feature vectors of historical and real-time working signals, which can comprehensively mine the multi-dimensional feature information in the vibration signal, including time domain, frequency domain and other features, thereby more accurately reflecting the vibration state of the shaft system and providing a richer information basis for fault location. At the same time, combined with the physical constraint loss function to judge and optimize the position of the excitation source, it can be more in line with actual complex working conditions and effectively improve the fault location accuracy.

[0033] It should be noted that the pre-built graph neural network model includes an input layer, a graph convolution layer, a pooling layer, and an output layer. The graph convolution layer includes a spectral domain convolution GCN layer, an attention layer, a physical constraint module, and an axis topology module. The axis topology module constructs an axis topology graph based on the signal feature vector. The nodes in the axis topology graph are component positions, and the connecting edges are signal feature vectors. The input layer receives rich raw data. The spectral domain convolution GCN layer, the attention layer, the physical constraint module, and the axis topology module in the graph convolution layer work together to deeply explore the key features in the data from different angles and levels. For example, the spectral convolutional network (GCN) layer captures frequency-domain correlations in vibration signals, the attention layer focuses on key features, the physical constraint module ensures that features conform to physical laws, and the shaft topology module constructs a topological map reflecting the vibration propagation relationships between components. These features, combined, provide a more comprehensive and accurate representation of the shaft vibration state. The pooling layer of the graph neural network model is frozen and replaced with a transferred convolutional network architecture. This transfer convolutional network architecture in the graph neural network model utilizes dilated causal convolution to capture the long-range dependencies between the fault location and the feature vector. Replacing the pooling layer with the transferred convolutional network architecture and utilizing dilated causal convolution expands the receptive field of the convolution operation and captures the long-range dependencies between the fault location and the feature vector. Under complex operating conditions, shaft vibration fault features may be scattered across a long signal sequence or a large spatial range. Dilated causal convolution integrates these scattered, fault-related features, facilitating the discovery of potential fault modes and fault location. This overcomes the issues of feature information loss or premature dimensionality reduction that can occur with traditional pooling layers. Furthermore, the physical constraint module constrains and corrects the feature extraction process based on the physical laws of shaft vibration, ensuring that the extracted feature vectors conform to the actual physical conditions. This makes the model's output more credible and interpretable, allowing maintenance personnel and engineers to better understand how the model locates faults based on physical principles, rather than just getting a difficult-to-interpret black box result. This improves the model's application value and acceptance in actual industrial scenarios.

[0034] Example 2: In this embodiment, a graph neural network model is provided. A spectral domain convolution GCN layer, an attention layer, a physical constraint module, and an axis topology module are introduced into the graph neural network model. The axis topology module can provide physical structure priors, while the spectral domain convolution GCN and the physical constraint module jointly ensure that feature learning conforms to the laws of dynamics, thereby forming a closed loop of "data-driven feature extraction-physical model verification". The introduction of the physical constraint module stabilizes the output of the model, enabling it to maintain relatively stable performance under different input signals and working conditions. At the same time, the coordinated cooperation of various modules makes the model more robust to data noise and outliers, reduces model performance fluctuations caused by data quality issues, and improves the stability and reliability of the model.

[0035] In this embodiment, the physical constraint loss function of the graph neural network model is expressed as: (1) in, Represent the weights of the data-driven loss term and the physical constraint loss term, respectively. represents the mean square error loss function, Indicates the transfer function gradient and the location of the excitation source for the inverse operation solution, represents the transfer function calculated based on the Jeffcott rotor model; When the axis topology module constructs the axis topology graph based on the signal eigenvector, it adjusts the edge weights by constructing an adaptive adjacency matrix between nodes. The adaptive adjacency matrix is ​​expressed as: (2) in, represents the activation function, represents a multilayer perceptron, Represents nodes respectively ,node exist The feature vector of the moment.

[0036] Example 3: This embodiment provides a method for tracing back the source of a fault by combining a message delivery mechanism with delivery path analysis. Figure 4 The following is a flow chart showing a method for tracing back the source of a fault by combining a message passing mechanism with a transfer path analysis. The method for tracing back the source of a fault by combining a message passing mechanism with a transfer path analysis specifically includes: Step S101: Obtain historical operating signal feature vectors and map them to a fault location-feature vector mapping library. By mapping historical operating signal feature vectors to the fault location-feature vector mapping library and leveraging the message passing mechanism of a graph neural network model, the propagation of vibration signals within the shafting graph structure can be accurately simulated. This provides a deeper understanding of how vibration signals propagate and interact between various components of the shafting system, providing an accurate propagation model foundation for subsequent fault location. Based on the message passing mechanism of a graph neural network model, the propagation of vibration signals within the shafting graph structure can be simulated. Step S102: The historical working signal feature vector is transferred along the transfer path according to the weights of the nodes and connecting edges in the graph structure. This helps determine which nodes in the axis system are associated with a specific fault location, thereby more accurately narrowing the fault scope and providing more specific guidance for subsequent path analysis and fault source tracing. At the same time, during the transfer process, feature nodes that are critical to fault location can be screened based on factors such as weight. These key nodes often contain rich information related to the fault characteristics. Through further analysis of them, it is possible to more deeply explore the fault characteristics, improve the accuracy of fault location, reduce unnecessary node analysis, and establish a mapping relationship between the fault location and the node; Step S103 , loading the mapping relationship between the fault location and the node, constructing the shaft system transfer function matrix based on the transfer path analysis, identifying the main path and the secondary path of the signal propagation, and obtaining the path analysis result including the main path and the secondary path.

[0037] In this embodiment, an axis transfer function matrix is ​​constructed based on transfer path analysis to identify the primary and secondary paths of signal propagation, thereby extracting paths with higher contribution and ensuring that computing and processing resources are concentrated on the critical paths, thereby improving diagnostic efficiency.

[0038] Step S104: Obtain the path analysis results, flip the shaft transfer function matrix, propagate the historical working signal eigenvector back to the potential fault location, load the vibration response, shaft transfer function matrix, and physical constraints, narrow the fault range by solving the excitation source through pseudo-inverse, and output the narrowed fault source.

[0039] In this embodiment, the shaft transfer function matrix is ​​flipped so that the historical working signal eigenvector propagates back toward the potential fault location. By combining the vibration response, the shaft transfer function matrix, and physical constraints, the excitation source is solved by pseudo-inverse method, which can effectively narrow the scope of the fault source and thus improve the reliability and credibility of the positioning results.

[0040] In this embodiment, when the fault range is narrowed by solving the excitation source through pseudo-inverse, the shafting transfer function matrix is ​​expressed as: (3) (4) in, represents the axis transfer function matrix, represents the physical constraint function, is the number of constraint items. In this embodiment, it can be 2-5. Representation node The stiffness parameter, Indicates the source of the fault Associated Nodes the number of represents the vibration response, Indicates the fault type label, is the cross entropy loss function, Node The historical working signal feature vector and fault source Associated Nodes The signal feature vector.

[0041] In this embodiment, when tracing back the fault source through the message passing mechanism combined with the transfer path analysis, the historical working signal characteristic vector is transmitted along the transfer path according to the weights of the nodes and connecting edges in the graph structure. This helps to determine which nodes in the shaft system are associated with a specific fault location, thereby more accurately narrowing the fault range and providing more specific direction for subsequent path analysis and fault source tracing. The shaft system transfer function matrix is ​​flipped to make the historical working signal characteristic vector propagate back to the potential fault location. In combination with the vibration response, the shaft system transfer function matrix and the physical constraints, the excitation source is solved by pseudo-inverse, which can effectively narrow the fault source range, thereby improving the reliability and credibility of the positioning results.

[0042] Example 4: This embodiment provides a method for extracting features from real-time working signals based on a graph neural network model. Figure 5 A schematic diagram of a method for extracting features from real-time working signals based on a graph neural network model is shown. The method for extracting features from real-time working signals based on a graph neural network model specifically includes: Step S201: Loading a real-time working signal, and the input layer adapting and formatting the real-time working signal to obtain a normalized set; In this embodiment, when the input layer adapts and normalizes the format of the real-time working signal, dynamic adaptive normalization (such as Z-score + sliding window normalization) can be used to eliminate the interference caused by sensor range differences and sudden changes in working conditions.

[0043] In step S202, a normalized set is obtained and analyzed using spectral domain convolution. The improved spectral domain convolution (GCN) can focus on fault-sensitive frequency bands, thereby accurately obtaining spectral domain analysis results. The Laplacian matrix of the normalized set is calculated to implement normalized set signal filtering and feature extraction, and output a key feature set.

[0044] It should be noted that filtering the normalized set signal by calculating the Laplacian matrix can effectively remove noise components from the signal and enhance the signal-to-noise ratio. This helps improve the accuracy of feature extraction and enables subsequent analysis to more clearly identify characteristic information related to the fault.

[0045] In step S203, the attention layer dynamically assigns weights to key features in the key feature set, and the physical constraint module modifies the key features. The physical constraint module embeds the Jeffcott rotor equation to force the features to conform to the laws of shaft system dynamics. The shaft system topology module fuses and enhances the key feature vectors to obtain a fused feature vector. The physical constraint module modifies the key features to ensure that the feature extraction results conform to the physical laws of shaft system vibration. This avoids misjudgments caused by the model extracting unreasonable features and improves the credibility and reliability of the features. Step S204: Load the feature fusion vector, migrate the convolutional network architecture, combine the spectral domain analysis results, and perform dilated causal convolution to capture the long-term dependency between the fault location and the feature vector.

[0046] In this embodiment, by capturing the long-range dependencies between fault location and feature vectors, a more comprehensive understanding of the distribution patterns of fault features in both temporal and spatial dimensions is achieved. This is crucial for identifying potential faults that gradually emerge during long-term operation or in complex shafting structures, thereby improving early warning capabilities for faults. Furthermore, under complex operating conditions, the propagation path of vibration signals can be complex. The analysis of long-range dependencies helps the model better track the source of faults, narrow the scope of the fault, and provide more precise guidance for subsequent repair and maintenance work.

[0047] On the other hand, this embodiment also provides a shafting vibration fault location system for complex working conditions. Figure 6 The structure diagram of the shafting vibration fault location system for complex working conditions is shown. The shafting vibration fault location system for complex working conditions specifically includes: The vector extraction module 100 builds a vibration propagation database based on historical working signals under complex working conditions and uses a pre-built graph neural network model to extract feature vectors of historical working signals; The fault-vector mapping module 200 loads historical working signal feature vectors and constructs a fault location-feature vector mapping library based on a message passing mechanism. The module then uses the message passing mechanism combined with transfer path analysis to reversely trace the fault source and solve the excitation source location through inverse calculations. The module then uses a physical constraint loss function to determine whether the excitation source location meets a loss threshold. If so, the module then outputs the fault location-feature vector mapping library based on the message passing mechanism combined with transfer path analysis. The fault point visualization module 300 is used to obtain real-time working signals, extract features from the real-time working signals based on a graph neural network model, output real-time feature vectors, perform correlation analysis on the real-time feature vectors based on a fault location-feature vector mapping library, determine the probability distribution of the fault location, and use a heat map to visualize the fault location in the three-dimensional shafting model.

[0048] This embodiment provides a fault point visualization module 300, Figure 7 FIG. 3 shows a schematic diagram of the structure of the fault point visualization module 300 , which specifically includes: A real-time feature extraction unit 310 is used to obtain a real-time working signal, perform feature extraction on the real-time working signal based on a graph neural network model, and output a real-time feature vector; The fault correlation analysis unit 320 performs real-time feature vector correlation analysis based on the fault location-feature vector mapping library to determine the probability distribution of the fault location; The visualization unit 330 is used to visualize the fault location in the three-dimensional shafting model using a heat map.

[0049] In this embodiment, the heat map is drawn based on the probability distribution data of the fault location. In the three-dimensional shafting model, each component location has a corresponding failure probability value. Based on these probability values, the visualization unit 330 assigns a corresponding color to each component location according to a preset color mapping rule. The color mapping rule typically uses a gradient from light to dark colors, such as blue, green, yellow, orange, and red, from low probability to high probability. The fault point visualization module 300, through the coordinated cooperation of the real-time feature extraction unit 310 and the fault correlation analysis unit 320, can quickly combine the characteristics of the real-time operating signal with historical failure experience to determine the probability distribution of the fault location. Combined with the heat map display of the visualization unit 330, operators can immediately and intuitively understand the failure risk status of each component in the shafting and quickly locate the location most likely to fail, significantly shortening fault diagnosis time and improving equipment maintenance efficiency.

[0050] In summary, the present invention provides a method and system for locating shaft vibration faults under complex working conditions. In this embodiment, a pre-built graph neural network model is used to extract feature vectors of historical and real-time working signals, which can comprehensively mine the multi-dimensional feature information in the vibration signal, including time domain, frequency domain and other features, thereby more accurately reflecting the vibration state of the shaft system and providing a richer information basis for fault location. At the same time, the physical constraint loss function is combined to judge and optimize the position of the excitation source, which can be more in line with actual complex working conditions and effectively improve the fault location accuracy.

[0051] In this embodiment, a graph neural network model is provided. A spectral domain convolution GCN layer, an attention layer, a physical constraint module, and an axis topology module are introduced into the graph neural network model. The axis topology module can provide physical structure priors, while the spectral domain convolution GCN and the physical constraint module jointly ensure that feature learning conforms to the laws of dynamics, thereby forming a closed loop of "data-driven feature extraction-physical model verification". The introduction of the physical constraint module stabilizes the output of the model, enabling it to maintain relatively stable performance under different input signals and working conditions. At the same time, the coordinated cooperation of various modules makes the model more robust to data noise and outliers, reduces model performance fluctuations caused by data quality issues, and improves the stability and reliability of the model.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A shafting vibration fault location method for complex working conditions, characterized in that: The following steps are involved: Obtain historical working signals under target working conditions and build a vibration propagation database based on the historical working signals; A pre-built graph neural network model is used to extract feature vectors of historical working signals in the vibration propagation database. Based on the eigenvectors, a fault location-eigenvector mapping library based on the message passing mechanism is constructed. The fault source is traced back based on the message passing mechanism combined with the transfer path analysis. The inverse operation is used to solve the excitation source location. The physical constraint loss function is used to determine whether the excitation source location meets the loss threshold. If it meets the loss threshold, the fault location-eigenvector mapping library based on the message passing mechanism combined with the transfer path analysis is output. Acquire real-time working signals and use a graph neural network model to extract their feature vectors. Then, use a message passing mechanism combined with a fault location-feature vector mapping library based on transfer path analysis to perform correlation analysis on the feature vectors of the real-time working signals and determine the probability distribution of the fault location. According to the probability distribution of the fault location, a heat map is used to visualize the fault location in the three-dimensional shafting model.

2. The shafting vibration fault location method for complex working conditions according to claim 1 is characterized in that: The graph neural network model includes a spectral domain convolution GCN layer, an attention layer, a physical constraint module and an axis topology module. The axis topology module constructs an axis topology graph based on the signal feature vector. The nodes in the axis topology graph are component positions, and the connecting edges are signal feature vectors. The pooling layer of the graph neural network model is frozen and replaced by a migration convolutional network architecture.

3. The shafting vibration fault location method for complex working conditions according to claim 1, characterized in that: The physical constraint loss function of the graph neural network model is expressed as: in, Represent the weights of the data-driven loss term and the physical constraint loss term, respectively. represents the mean square error loss function, represents the transfer function gradient and the location of the excitation source solved by the inverse operation, represents the transfer function calculated based on the Jeffcott rotor model.

4. The shafting vibration fault location method for complex working conditions according to claim 2, characterized in that: When the axis topology module constructs the axis topology graph based on the signal eigenvector, it adjusts the edge weights by constructing an adaptive adjacency matrix between nodes. The adaptive adjacency matrix is ​​expressed as: in, represents the activation function, represents a multilayer perceptron, Represents nodes respectively ,node exist The feature vector of the moment.

5. The shafting vibration fault location method for complex working conditions according to claim 1, characterized in that: The specific method of tracing back the fault source based on the message passing mechanism combined with transmission path analysis is as follows: Map the historical working signal feature vectors to the fault location-feature vector mapping library to simulate the propagation process of the vibration signal in the shafting structure; According to the weights of the nodes and connecting edges in the graph neural network model, the historical working signal feature vector is transmitted along the transmission path to build a mapping relationship between the fault location and the node; The mapping relationship between the loading fault location and the node is constructed, and the shaft system transfer function matrix is ​​constructed based on the transfer path analysis. The main and secondary paths of signal propagation are identified, and the path analysis results including the main and secondary paths are obtained.

6. The shafting vibration fault location method for complex working conditions according to claim 5, characterized in that: After obtaining the path analysis results including the primary and secondary paths, the shafting transfer function matrix is ​​flipped to make the historical working signal eigenvector propagate back to the potential fault location. The vibration response, shafting transfer function matrix, and physical constraints are loaded. The fault range is narrowed by solving the excitation source through pseudo-inverse solution, and the fault source with a narrowed range is output.

7. The shafting vibration fault location method for complex working conditions according to claim 6, characterized in that: The specific method of narrowing the fault range by solving the excitation source through pseudo-inverse solution and outputting the fault source with narrowed range is as follows: in, represents the axis transfer function matrix, represents the physical constraint function, is the number of constraints, Representation node The stiffness parameter, Indicates the source of the fault Associated Nodes the number of represents the vibration response, Indicates the fault type label, is the cross entropy loss function, Node The historical working signal feature vector and fault source Associated Nodes The signal feature vector.

8. The shafting vibration fault location method for complex working conditions according to claim 1, characterized in that: The method for obtaining real-time working signals and extracting feature vectors of real-time working signals using a graph neural network model is as follows: Acquire a real-time working signal, adapt and normalize the format of the real-time working signal, and obtain a normalized set; The normalized set is analyzed using spectral domain convolution operation to obtain spectral domain analysis results. The Laplace matrix of the normalized set is calculated to implement normalized set signal filtering and feature extraction, and output the key feature set.

9. The shafting vibration fault location method for complex working conditions according to claim 1, characterized in that: The specific method for performing correlation analysis on the feature vectors of real-time working signals using the message passing mechanism combined with the fault location-feature vector mapping library of the transfer path analysis is as follows: Assign weights to key features in the key feature set, modify the key features, and fuse and enhance the modified key feature vectors to obtain a feature fusion vector; Combining the spectral domain analysis results and dilated causal convolution, the feature fusion vector is analyzed to capture the long-term dependency between the fault location and the feature vector.

10. A shafting vibration fault location method system for complex working conditions, characterized in that: include: A historical signal acquisition module is used to obtain historical working signals under target working conditions and build a vibration propagation database based on the historical working signals; A vector extraction module is used to pre-build a graph neural network model and use it to extract feature vectors of historical working signals in the vibration propagation database; The fault-vector mapping module is used to construct a fault location-feature vector mapping library based on the message passing mechanism based on the feature vector. The fault source is traced back based on the message passing mechanism combined with the transfer path analysis. The inverse operation is used to solve the excitation source location. The physical constraint loss function is used to determine whether the excitation source location meets the loss threshold. If the loss threshold is met, the fault location-feature vector mapping library based on the message passing mechanism combined with the transfer path analysis is output. The fault point determination module is used to obtain real-time working signals, extract the feature vectors of the real-time working signals using a graph neural network model, and perform correlation analysis on the feature vectors of the real-time working signals using a message passing mechanism combined with a fault location-feature vector mapping library based on transfer path analysis to determine the probability distribution of the fault location. The visualization module is used to visualize the fault location points using a heat map in the three-dimensional shafting model based on the probability distribution of the fault location.

Images