Monorail crane bearing fault diagnosis method based on adaptive graph convolutional network

Abstract

The invention relates to the technical field of coal mine underground intelligent operation and maintenance and fault diagnosis, and discloses a monorail crane bearing fault diagnosis method based on a self-adaptive graph convolutional network, which comprises the following steps: S1, sound signal acquisition and preprocessing: acquiring sound signals of a monorail crane bearing under different operation conditions, and preprocessing the sound signals to obtain sound signals of the monorail crane bearing under different operation conditions; to obtain signal segments for analysis; s2, time-frequency characteristics are generated; s3, constructing graph nodes; s4, multi-dimensional node feature extraction; s5, constructing a graph structure; s6, adaptive graph convolution feature extraction; and S7, carrying out fault identification and diagnosis output. A self-adaptive time-frequency parameter optimization mechanism based on an Archimedes optimization algorithm is introduced, and cooperative work of dynamic order self-adaptive graph convolution, a gating mechanism and a residual information retention mechanism is combined, so that under the condition of extremely strong noise, noise redundancy features can be effectively suppressed, the problem of feature over-smoothing is relieved, and the robustness of the system is improved. And the discrimination capability and the diagnosis reliability of the monorail crane bearing fault features are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and fault diagnosis technology in underground coal mines, specifically a fault diagnosis method for monorail crane bearings based on adaptive graph convolutional networks. Background Technology

[0002] Monorail systems are widely used rail transport equipment in mines, tunnels, and complex industrial environments. Their operational safety and reliability directly affect personnel safety and production efficiency. Bearings, as key rotating support components of monorails, endure complex loads and harsh environmental influences during long-term operation, and their operating status plays a crucial role in the stability of the monorail system. To ensure safe equipment operation, fault diagnosis of monorail bearings is a vital technical aspect of monorail equipment maintenance. Existing monorail bearing fault diagnosis methods typically rely on time-frequency analysis of acoustic or vibration signals, combined with feature extraction and classification models to achieve fault identification. In practical applications, short-time Fourier transform is often used for time-frequency analysis of bearing acoustic signals.

[0003] However, in current technology, short-time Fourier transform is widely used to obtain the time-frequency characteristics of sound signals. In practical applications, fixed time-frequency analysis parameter settings are often used, which makes it difficult to adapt to the non-stationary sound signal characteristics exhibited by monorail crane bearings under different operating conditions and noise environments. This can easily lead to limited time-frequency characteristic discrimination ability, thereby affecting the accuracy and stability of fault diagnosis results. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for diagnosing bearing faults in monorail cranes based on adaptive graph convolutional networks. This method solves the problem that fixed time-frequency analysis parameters in bearing fault diagnosis based on acoustic signals are difficult to adapt to the non-stationary characteristics of acoustic signals under different operating conditions, thus limiting the ability to identify fault features.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for diagnosing single-rail crane bearing faults based on adaptive graph convolutional networks, comprising: S1. Acoustic signal acquisition and preprocessing: Acquire acoustic signals of the monorail crane bearing under different operating conditions, and preprocess the acoustic signals to obtain signal segments for analysis. S2. Time-frequency feature generation: A parameter optimization algorithm is used to jointly optimize multiple parameters of the time-frequency analysis method, and the signal segment is transformed by time-frequency based on the optimized parameters to generate a time-frequency feature map. S3. Graph node construction: Divide the time-frequency feature map into regions, and use each region as a node in the graph structure to construct a node set. S4. Multi-dimensional node feature extraction: For the region corresponding to each node, extract multi-dimensional node feature vectors from the time-frequency feature map to characterize energy characteristics, change characteristics and transient characteristics. S5. Graph structure construction: Construct a graph structure based on the relationships between nodes and generate connection relationships for graph convolution operations. S6. Adaptive graph convolution feature extraction: The graph structure and node feature vectors are input into an adaptive graph convolution network. The neighborhood aggregation range of the graph convolution is dynamically adjusted by statistical analysis of the node features. A feature adjustment mechanism and a residual information preservation mechanism are introduced during the graph convolution process to obtain a graph-level feature representation for fault identification. S7. Fault identification and diagnosis output: Based on the graph-level feature representation, the operating status of the monorail crane bearing is classified, and the bearing fault diagnosis results are output.

[0006] Preferably, the time-frequency feature generation includes: A set of parameters to be optimized for time-frequency analysis is set for the preprocessed acoustic signal; The parameter optimization algorithm is used to iteratively search the set of parameters to be optimized, and the optimal parameter combination is determined by using the bearing fault classification accuracy as the evaluation index. The acoustic signal is subjected to time-frequency transformation using the optimal parameter combination to generate a corresponding time-frequency feature map.

[0007] Preferably, the graph node construction includes: The time-frequency feature map is segmented according to region division rules to obtain multiple regions; Each region is treated as a node in a graph structure, and the corresponding time-frequency data for that region is obtained. Assign a unique identifier to each node and build a node set.

[0008] Preferably, the node feature construction includes: For each node, calculate the characteristic quantity that can characterize the energy distribution of the corresponding regional data. Calculate characteristic quantities that reflect the magnitude or fluctuation of changes in the data of the region; The region data is used to calculate feature quantities that can characterize transient changes, and these feature quantities are combined to form a multidimensional feature vector of the node.

[0009] Preferably, the graph structure construction includes: The connection relationships between nodes are determined based on the spatial relationship, regional boundary relationship, or feature similarity between adjacent nodes; A connection matrix is ​​constructed based on the aforementioned connection relationships to characterize the connectivity between nodes; The connection matrix is ​​normalized to obtain the graph structure matrix used for graph convolution operations.

[0010] Preferably, the adaptive graph convolutional feature extraction includes: Perform global statistics on all node features to form a global feature vector that describes the overall graph characteristics; The global feature vector is input into the mapping module to generate dynamic control parameters for determining the order of the graph convolution polynomial; The polynomial order of the graph convolution is adaptively determined based on the dynamic control parameters, and the feature aggregation range of the graph convolution is adjusted.

[0011] Preferably, the adaptive graph convolution feature extraction further includes: Based on node features or intermediate features generated during graph convolution, adjustment parameters are generated to characterize the importance of node features of different orders. The contribution of node features of different orders or different channels is adaptively weighted according to the adjustment parameters.

[0012] Preferably, the initial features or low-order aggregated features of the nodes are preserved during the graph convolution process; The initial features or low-order aggregated features are fused with the graph convolution output features to reduce the oversmoothing of features caused by multi-layer graph convolution.

[0013] Preferably, the adaptive graph convolutional feature extraction also includes: Processing that dynamically determines the order of graph convolution polynomials based on global statistical features to adjust the range of feature aggregation; Adaptive weighting of convolutional features of different orders and fusion of initial features or low-order features are performed.

[0014] The fault identification and diagnosis output includes: The graph-level feature representation is input into the classification model for state determination; The fault type of the monorail crane bearing is determined based on the state discrimination result; Output the corresponding bearing fault diagnosis results.

[0015] This invention provides a method for fault diagnosis of monorail crane bearings based on adaptive graph convolutional networks. It has the following beneficial effects: 1. This invention introduces an adaptive time-frequency parameter optimization mechanism based on the Archimedes optimization algorithm, and combines it with dynamic order adaptive graph convolution, gating mechanism and residual information preservation mechanism to work together. Under extreme strong noise conditions, it can effectively suppress noise redundancy features, alleviate the problem of feature over-smoothing, and significantly improve the discrimination ability and diagnostic reliability of single-rail crane bearing fault features.

[0016] 2. This invention achieves effective differentiation of different bearing fault types under strong noise conditions by adaptively optimizing and modeling time-frequency features and constructing a multi-order graph convolution feature fusion representation. Compared with the traditional ChebGCN method and mainstream deep learning models such as CNN and LSTM, it shows significant performance advantages, verifying the effectiveness of the adaptive graph convolution network in bearing fault diagnosis under complex noise environments.

[0017] 3. This invention adopts a gating mechanism to adaptively adjust convolutional features of different order graphs and introduces a residual information preservation mechanism to stabilize the feature propagation process. This enables the model to converge to a higher diagnostic accuracy faster and with smaller fluctuations during training, effectively improving the problems of large training oscillations and unstable convergence in the comparison model and enhancing the stability of the algorithm in engineering applications. Attached Figure Description

[0018] Figure 1 This is a flowchart and network structure diagram of a single-rail crane bearing fault diagnosis method based on an adaptive graph convolutional network according to the present invention. Figure 2 This is a schematic diagram illustrating the construction of the time-frequency graph-to-graph data structure based on superpixel segmentation according to the present invention; Figure 3 This is a schematic diagram of the graph convolution feature extraction and gated residual fusion structure based on adaptive polynomial order of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see the appendix Figure 1 - Figure 3 This invention provides a method for diagnosing single-rail crane bearing faults based on adaptive graph convolutional networks, comprising: S1. Acoustic signal acquisition and preprocessing: Acquire acoustic signals of monorail crane bearings under different operating conditions, and preprocess the acoustic signals to obtain signal segments for analysis. Specifically, the acoustic signals of the monorail crane bearing under different operating conditions are collected. The acoustic signals can be collected by acoustic sensors placed near the bearing and converted into digital signals by a data acquisition device for storage. Before analyzing the acoustic signal, the acquired acoustic signal is preprocessed. The preprocessing includes removing the DC component, amplitude normalization, and signal segmentation to obtain multiple acoustic signal segments for subsequent analysis. This reduces noise interference and ensures the consistency of the signal segments on the time scale, providing a stable data foundation for subsequent time-frequency feature extraction and fault diagnosis. For example, in one specific embodiment, the continuously collected bearing operating sound signal is segmented according to a preset time length to obtain multiple sound signal segments, and the sound signal segments are input into subsequent processing steps for analysis.

[0021] S2. Time-frequency feature generation: A parameter optimization algorithm is used to jointly optimize multiple parameters of the time-frequency analysis method, and the signal segment is transformed by time-frequency transformation based on the optimized parameters to generate a time-frequency feature map. Furthermore, the generation of the aforementioned time-frequency features includes: A set of parameters to be optimized for time-frequency analysis is set for the preprocessed acoustic signal; The parameter optimization algorithm is used to iteratively search the set of parameters to be optimized, and the optimal parameter combination is determined by using the bearing fault classification accuracy as the evaluation index. The acoustic signal is transformed by time-frequency conversion using the optimal parameter combination to generate the corresponding time-frequency feature map.

[0022] Specifically, in order to obtain the joint distribution characteristics of the acoustic signal in the time and frequency dimensions, time-frequency features are generated for the acoustic signal segments. Considering that the acoustic signal of the monorail crane bearing has obvious non-stationary characteristics under different operating conditions, and that fixed time-frequency analysis parameters are difficult to adapt to different acoustic signal characteristics, this invention uses a parameter optimization algorithm to jointly optimize multiple key parameters in the time-frequency analysis method to generate a time-frequency feature map with high discriminativeness. First, a set of parameters to be optimized for time-frequency analysis is set for the preprocessed acoustic signal. The set of parameters to be optimized includes parameters that affect the time-frequency resolution and feature expression effect during the time-frequency analysis process, such as the analysis window length, the number of overlapping samples, the number of time segments, and the number of frequency segments. By uniformly setting the above parameters, the subsequent optimization process can be searched within a reasonable parameter space, thereby avoiding the blurring of time-frequency features or the loss of key information due to improper parameter selection. Based on this, a parameter optimization algorithm is used to iteratively search the set of parameters to be optimized, and the optimal parameter combination is determined by using the bearing fault classification accuracy as the evaluation index. The optimal parameter optimization algorithm adopts the Archimedes optimization algorithm (AOA), which searches for parameters by simulating the buoyancy behavior of an object in a fluid, and uses the transfer factor to dynamically balance the global exploration and local development capabilities. By inputting the time-frequency feature map generated based on the current parameter combination into the fault classification model, the corresponding fault classification accuracy is calculated and used as the fitness function to guide the parameter search process to gradually tend towards the parameter combination that can improve the fault identification performance. Subsequently, the acoustic signal segment is subjected to time-frequency transformation using the optimal parameter combination to generate the corresponding time-frequency feature map. The time-frequency feature map is used to characterize the energy distribution features of the acoustic signal in the time and frequency planes, and serves as the basic input for region division, graph node construction, and node feature extraction in subsequent steps, thereby providing a high-quality time-frequency feature representation for subsequent graph-based fault diagnosis.

[0023] For example, in one specific embodiment, a candidate range is pre-defined for the window length, number of overlapping samples, number of time segments, and number of frequency segments of the acquired bearing operating sound signal segments; the Archimedes optimization algorithm is used to iteratively search within the candidate range, and the optimal parameter combination is selected using the fault classification accuracy as an evaluation index; then, the sound signal segments are subjected to time-frequency transformation based on the optimal parameter combination to generate the corresponding time-frequency feature map, and the time-frequency feature map is input into subsequent steps for processing.

[0024] S3. Graph node construction: Divide the time-frequency feature map into regions, and use each region as a node in the graph structure to construct a node set. Furthermore, graph node construction includes: The time-frequency feature map is segmented according to region division rules to obtain multiple regions; Each region is treated as a node in the graph structure, and the corresponding time-frequency data for that region is obtained. Assign a unique identifier to each node and build a node set.

[0025] Specifically, the time-frequency feature map after AOA optimization is processed to construct graph nodes. In order to convert the continuous time-frequency feature map into a discrete structure representation suitable for graph convolutional network processing, the time-frequency feature map is divided into regions, and the divided regions are used as nodes in the graph structure to construct a node set. The time-frequency feature map optimized by AOA is divided into multiple uniform superpixel blocks. These superpixel blocks are used to aggregate adjacent time-frequency points in the time-frequency feature map according to certain rules, so that the time-frequency features within the same superpixel block have relatively consistent distribution characteristics. By dividing the time-frequency feature map into uniform superpixel blocks, the local structural information of the time-frequency features can be preserved while reducing the data scale and noise sensitivity in the subsequent feature modeling process. After the superpixel blocks are divided, each superpixel block is treated as a node in the graph structure, and the corresponding time-frequency data is obtained. Each node is used to represent the overall time-frequency feature information within its corresponding superpixel block, thereby realizing the transformation from a continuous time-frequency feature map to a discrete graph node representation, providing basic input for subsequent node feature extraction and graph convolution feature learning; Furthermore, a unique identifier is assigned to each node, and all nodes are aggregated to construct a node set. The node set is used to describe the overall composition of all superpixel nodes in the time-frequency feature map and serves as the basis for subsequent graph structure construction and node relationship modeling.

[0026] For example, in one specific embodiment, the time-frequency feature map after parameter optimization is divided into multiple uniformly sized superpixel blocks according to a preset superpixel partitioning method, and each superpixel block corresponds to a graph node; then, a unique number is assigned to each graph node to form a node set, and the node set is input into subsequent steps for node feature extraction and graph convolution feature learning.

[0027] S4. Multi-dimensional node feature extraction: For the region corresponding to each node, extract multi-dimensional node feature vectors from the time-frequency feature map to characterize energy characteristics, change characteristics and transient characteristics. Furthermore, node feature construction includes: For each node, calculate the characteristic quantity that can characterize the energy distribution of the corresponding regional data. Calculate characteristic quantities for regional data that can reflect its magnitude of change or fluctuation; The feature quantities that can characterize transient changes are calculated for regional data, and these feature quantities are combined to form a multidimensional feature vector of the node.

[0028] Specifically, based on the completion of the graph node construction, for the region corresponding to each node, a multi-dimensional node feature vector is extracted from the time-frequency feature map to characterize the energy characteristics, change characteristics and transient characteristics. By integrating multiple acoustic physical features, the node features can reflect the energy distribution characteristics and dynamic change characteristics of the bearing signal in the local time-frequency region from different perspectives, thereby improving the ability of the subsequent fault diagnosis model to distinguish different operating states. First, calculate the characteristic quantity that can characterize the energy distribution for the region data corresponding to each node, that is, for the first node... The energy characteristics of the superpixel regions corresponding to each node are characterized by the normalized mean spectral amplitude, denoted as . The calculation method is as follows: ; in, Indicates the first The region corresponding to the node is the Each time-frequency amplitude, This indicates the number of time-frequency sampling points contained in the region. The above energy characteristics can reflect the overall energy level of the bearing signal in the corresponding time-frequency region. Simultaneously, characteristic quantities that can reflect the magnitude or fluctuation of regional data are calculated, specifically for the first... The fluctuation of time-frequency amplitude within the region corresponding to each node is characterized by the normalized standard deviation of the spectral amplitude, denoted as . The calculation method is as follows: ; By introducing characteristic quantities of changing features, the stability of energy changes in bearing signals within local time-frequency regions can be reflected, which helps to distinguish between stable operating conditions and fault conditions with abnormal fluctuations. Furthermore, characteristic quantities that characterize transient changes are calculated for regional data. The transient impact characteristics are characterized by calculating the range of time-frequency amplitude gradients within the region. These transient characteristic quantities are denoted as... The calculation method is as follows: ; Here, gradient represents the gradient information of time-frequency amplitude within the region. By introducing transient feature quantities, the impact acoustic characteristics generated by bearing failure can be effectively captured.

[0029] After completing the calculations of the above energy characteristics, change characteristics, and transient characteristics, , and The nodes are fused together to construct their three-dimensional feature vectors: ; The three-dimensional node feature vector is used to comprehensively characterize the acoustic physical properties within the corresponding node region and serves as the node input feature for the subsequent graph convolution feature extraction process.

[0030] For example, in a specific embodiment, the time-frequency feature map optimized by AOA is divided into multiple node regions by superpixel division; for any one of the node regions, the normalized mean, standard deviation and gradient range of the time-frequency amplitude in the region are calculated respectively, and the three types of features are combined to form the three-dimensional feature vector of the corresponding node, which is used for subsequent graph structure feature learning and bearing fault diagnosis.

[0031] S5. Graph structure construction: Construct a graph structure based on the relationships between nodes and generate connection relationships for graph convolution operations. Furthermore, graph structure construction includes: The connection relationships between nodes are determined based on the spatial relationship, regional boundary relationship, or feature similarity between adjacent nodes; A connection matrix is ​​constructed based on the connection relationships to represent the connectivity between nodes; The connection matrix is ​​normalized to obtain the graph structure matrix used for graph convolution operations.

[0032] Specifically, after completing the construction of the node set and the extraction of node features, a graph structure is constructed based on the relationships between nodes, and connection relationships for graph convolution operations are generated. By introducing a graph structure to model the relationships between nodes, the subsequent graph convolutional network can fully utilize the relationship information between nodes based on the node features, realizing the propagation and aggregation of information in the graph structure; First, the connection relationship between nodes is determined based on the spatial relationship, regional boundary relationship or feature similarity between adjacent nodes. The spatial relationship between adjacent nodes is used to characterize the relative positional relationship of nodes in the time-frequency feature map. The regional boundary relationship is used to describe whether there is a directly adjacent boundary between different superpixel nodes. The feature similarity is used to reflect the degree of similarity between different nodes at the acoustic feature level. By comprehensively considering one or more of the above relationships, a connection relationship reflecting the association characteristics between nodes can be constructed. After determining the connection relationships between nodes, a connection relationship matrix is ​​constructed based on the connection relationships to characterize the connectivity between nodes. The connection relationship matrix records whether there are connections between nodes and their corresponding relationships, thereby expressing the topological structure information between nodes in matrix form and providing structural input for subsequent graph convolution operations. Furthermore, the connection relation matrix is ​​normalized to obtain the graph structure matrix for graph convolution operations. By normalizing the connection relation matrix, the influence of different nodes in the graph structure can be balanced, avoiding adverse effects on the graph convolution feature learning process due to excessive differences in the number of node connections, thereby improving the stability and robustness of the graph convolution network when processing node features.

[0033] For example, in one specific embodiment, for multiple nodes obtained by superpixel partitioning, the connection relationship between nodes is determined according to the adjacency relationship of nodes in the time-frequency feature map. Then, the corresponding connection relationship matrix is ​​constructed and the matrix is ​​normalized to obtain the graph structure matrix for graph convolution operation. The graph structure matrix and node features are then input into the subsequent graph convolution network for feature learning.

[0034] S6. Adaptive graph convolution feature extraction: Input the graph structure and node feature vectors into the adaptive graph convolution network. By statistically analyzing the node features, the neighborhood aggregation range of the graph convolution is dynamically adjusted. In the graph convolution process, a feature adjustment mechanism and a residual information preservation mechanism are introduced to obtain graph-level feature representations for fault identification. Furthermore, adaptive graph convolutional feature extraction includes: Perform global statistics on all node features to form a global feature vector that describes the overall graph characteristics; The global feature vector is input into the mapping module to generate dynamic control parameters for determining the order of the graph convolution polynomial; The polynomial order of graph convolution is adaptively determined based on dynamic control parameters, thereby adjusting the feature aggregation range of graph convolution.

[0035] Specifically, based on the completion of node feature extraction and graph structure construction, the graph structure and node features are input into an adaptive graph convolutional network for feature extraction. To overcome the problem that the convolution order in traditional graph convolutional networks is fixed and difficult to adapt to different data distribution characteristics, a dynamic control mechanism based on global statistical information is introduced to adaptively adjust the polynomial order of the graph convolution, thereby realizing the dynamic adjustment of the feature aggregation range. First, global statistics are performed on the features of all nodes to form a global feature vector describing the overall graph characteristics. In one specific implementation, suppose the graph has a total of There are nodes, and the three-dimensional feature vector corresponding to each node is... The global feature vector is obtained by averaging the feature vectors of all nodes. The calculation method is as follows: ; in, Indicates the first The feature vector corresponding to each node This represents the total number of nodes in the graph. Through global statistical operations, the average energy distribution and variation characteristics of the current time-frequency feature map can be characterized from an overall perspective, providing a basis for the adaptive adjustment of subsequent graph convolution parameters. After obtaining the global feature vector, it is input into the mapping module to generate dynamic control parameters for determining the order of the graph convolution polynomial. The mapping module uses a multilayer perceptron structure to map the global feature vector and combines activation functions and rounding operations to output the dynamic polynomial order. The calculation method is as follows: ; in, This represents the mapping function of a multilayer perceptron. This represents the activation function. This represents the rounding operation, which uses a mapping method to dynamically adjust the polynomial order of the graph convolution according to the overall feature distribution of the current sample, rather than using a fixed manually set value. After determining the order of the dynamic polynomial, the order of the polynomial in graph convolution is adaptively determined based on the dynamic control parameters, and the feature aggregation range in graph convolution operation is adjusted. Specifically, a larger polynomial order corresponds to a larger neighborhood aggregation range, which is used to capture a wider range of node association information, while a smaller polynomial order is used to focus on local node features. In this way, adaptive selection of the feature aggregation scale can be achieved under different working conditions and different acoustic signal characteristics, thereby improving the flexibility and effectiveness of graph convolution feature extraction.

[0036] For example, when the node features reflected by the global feature vector change significantly, a larger polynomial order is output through the mapping module to enhance the aggregation capability of long-distance correlation information between nodes. When the node features reflected by the global feature vector are relatively stable, a smaller polynomial order is output to avoid feature smoothing caused by excessive aggregation, thereby improving the discrimination performance of subsequent fault diagnosis.

[0037] Furthermore, adaptive graph convolutional feature extraction also includes: Based on node features or intermediate features generated during graph convolution, adjustment parameters are generated to characterize the importance of node features of different orders. The contribution of node features of different orders or different channels is adaptively weighted based on the adjustment parameters.

[0038] Specifically, based on the adaptive determination of the graph convolution polynomial order based on global statistical features, in order to further distinguish the importance of node features of different orders in fault identification, a feature adjustment mechanism is introduced to adaptively weight the contribution of node features of different orders or node features of different channels, thereby enhancing the ability to express key features. First, based on node features or intermediate features generated during graph convolution, adjustment parameters are generated to characterize the importance of node features of different orders. That is, the intermediate features obtained during graph convolution are input into the gating module, and the corresponding adjustment weight vector is generated through the mapping function. The adjustment weights are used to characterize the importance of graph convolution features of different orders under the current sample conditions. The calculation method is as follows: ; Where H represents the intermediate node features or the combination of features of different orders obtained during graph convolution. and Let represent the weight parameters and bias parameters in the gating module, respectively; σ(·) represents the activation function; and α represents the generated adjustment parameter vector. Thus, through gating mapping, the adjustment parameters can be adaptively adjusted according to the changes in node features. After obtaining the adjustment parameters, adaptive weighting is performed on the contribution of node features of different orders or different channels based on the adjustment parameters. The convolutional features of different orders of graphs are weighted with the corresponding adjustment parameters respectively, and the weighted features are fused. The calculation method is as follows: ; in, This represents the node features obtained from the k-th order graph convolution. The adjustment weights corresponding to the k-th order features are represented by K, where K represents the order of the graph convolution polynomial determined by the current adaptive method. By weighting, the k-th order features that contribute more to fault identification are given higher weights, while the k-th order features with weaker discriminative ability are suppressed, thereby improving the effectiveness of the overall feature representation. For example, when low-order features of graph convolution can effectively reflect local bearing fault information, the adjustment parameters generated by the gating module will assign greater weights to these low-order features. Conversely, when high-order features are more advantageous for capturing long-distance node association information, the weights of the high-order features will be increased accordingly. This achieves adaptive adjustment of the contribution of node features of different orders, improving the accuracy and stability of subsequent fault diagnosis.

[0039] Furthermore, the initial features or low-order aggregated features of nodes are preserved during graph convolution; The initial features or low-order aggregated features are fused with the output features of graph convolution to reduce the oversmoothing of features caused by multi-layer graph convolution.

[0040] Specifically, based on the completion of graph convolution feature extraction based on dynamic polynomial order and adaptive weighting of node features of different orders, in order to avoid the problem of feature oversmoothing caused by the convergence of node features due to multiple neighborhood aggregations, a residual information preservation mechanism is introduced in the graph convolution process to balance the global expressive power of high-order features and the local discriminative power of low-order features. Specifically, in the graph convolution feature extraction process, the initial features or low-order aggregated features of the nodes are preserved, and the features are linearly transformed to form residual information. In one specific implementation, the residual information is calculated as follows: ; in, This represents the initial features or low-order aggregated features of a node. This represents the linear transformation weight matrix used for residual mapping. Through this linear transformation, the initial or low-order features are made consistent with the graph convolution output features in both dimension and scale, thus facilitating subsequent fusion. After calculating the residual information, the residual information is fused with the graph convolution output features to form the network's final output features. In one specific implementation, the network output features... The calculation method is as follows: ; in, Represented by the Graph Laplace matrix The polynomial operator of order 1 Indicates the first The weight matrix corresponding to the order graph convolution. The first term generated by the feature adjustment mechanism Adjusting weights based on first-order features This represents the order of the graph convolution polynomial determined through a dynamic control mechanism. Residual information is represented. By fusing residual information with the output features of multi-order graph convolution, the network can aggregate high-order neighborhood information while retaining initial features or low-order local features that are important for bearing fault identification. For example, when the number of graph convolutional layers is large or the aggregation range is large, multiple neighborhood aggregations can easily lead to excessive smoothing of node features. By introducing the above residual information preservation mechanism, the initial features or low-order features of the nodes are directly introduced into the final output, so that different nodes still maintain sufficient distinguishability in the feature space, thereby improving the accuracy and robustness of bearing fault diagnosis.

[0041] Furthermore, adaptive graph convolutional feature extraction also includes: Processing that dynamically determines the order of graph convolution polynomials based on global statistical features to adjust the range of feature aggregation; Adaptive weighting of convolutional features of different orders and fusion of initial features or low-order features are performed.

[0042] Specifically, adaptive graph convolution feature extraction includes a process that dynamically determines the order of the graph convolution polynomial based on global statistical features to adjust the feature aggregation range, and a process that adaptively weights graph convolution features of different orders and fuses initial features or low-order features. Through the synergistic effect of the above multiple mechanisms, the graph convolution feature extraction process can be adaptively adjusted according to the overall characteristics and local feature distribution of the input acoustic signal, thereby improving the effectiveness of fault feature expression. Specifically, in the graph convolution feature extraction process, firstly, global statistics are performed on node features to obtain global feature information describing the overall graph characteristics. Then, based on the global feature information, the polynomial order of the graph convolution is dynamically determined to adaptively adjust the feature aggregation range in the graph convolution operation. Through this processing, the graph convolutional network can flexibly select an appropriate neighborhood aggregation scale under different sample conditions, avoiding the feature underfitting or overfitting problems caused by using a fixed aggregation range. Building upon this, a feature adjustment mechanism is introduced to adaptively weight graph convolution features of different orders, targeting those obtained through different polynomial orders. The weighting process determines the contribution of features of different orders based on node features or intermediate features generated during graph convolution, giving greater weight to features of higher order that have higher discriminative power for fault identification, thereby enhancing the influence of key features in the final feature representation. Furthermore, after multi-level feature weighting, the weighted graph convolutional features are fused with the initial features or low-level aggregated features of the nodes, introducing a residual information preservation mechanism. Through this fusion process, the final features, while integrating high-level node association information, retain initial or low-level information that is important for the local fault features of the bearing, thereby mitigating the feature oversmoothing problem caused by multi-level graph convolution operations and improving the discriminative power of node features and graph-level features.

[0043] S7. Fault Identification and Diagnosis Output: Based on graph-level feature representation, the operating status of the monorail crane bearing is classified, and the bearing fault diagnosis results are output.

[0044] Furthermore, the fault identification and diagnosis outputs include: Input graph-level feature representations into a classification model for state determination; The fault type of the monorail crane bearing is determined based on the state discrimination results; Output the corresponding bearing fault diagnosis results.

[0045] Specifically, after completing the adaptive graph convolution feature extraction, the operating status of the monorail crane bearing is identified based on the obtained graph-level feature representation, and the corresponding bearing fault diagnosis results are output. The graph-level feature representation comprehensively reflects the local acoustic features of the bearing acoustic signal in the time and frequency domain as well as the structural correlation information between nodes, and can effectively characterize the overall characteristics of the bearing under different operating conditions. Specifically, the graph-level feature representation is input into the classification model for state discrimination. The classification model can adopt the existing classification model structure to analyze the input graph-level feature representation and output the corresponding operating state category. Through this classification process, the normal operating state of the bearing and different fault states can be distinguished. After completing the status identification, the fault type of the monorail crane bearing is determined based on the status identification results. The fault type can correspond to different bearing operating states, which can be used to reflect whether there is an abnormality in the bearing and the specific category of the abnormality, thereby providing a basis for subsequent operation and maintenance. Finally, the corresponding bearing fault diagnosis results are output. The diagnosis results can be used to indicate the current operating status of the monorail crane bearing, assisting maintenance personnel in timely understanding the bearing health status and achieving effective monitoring and diagnosis of bearing faults.

[0046] For example, in one specific embodiment, such as Figure 2 In the middle, the left side is the time-frequency graph after superpixel segmentation. Each superpixel block is used as a node in the graph data structure. Each node contains 3-dimensional features and is connected to the four surrounding nodes (up, down, left, and right).

[0047] like Figure 3 In this process, graph data is convolved through graph Fourier transform, and the order of the graph convolution polynomial is dynamically determined based on the average node features to adjust the feature aggregation range. Furthermore, gating and residual modules are integrated to optimize the information flow, and finally new features are generated.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for fault diagnosis of monorail crane bearings based on adaptive graph convolutional networks, characterized in that, include: S1. Acoustic signal acquisition and preprocessing: Acquire acoustic signals of the monorail crane bearing under different operating conditions, and preprocess the acoustic signals to obtain signal segments for analysis. S2. Time-frequency feature generation: A parameter optimization algorithm is used to jointly optimize multiple parameters of the time-frequency analysis method, and the signal segment is transformed by time-frequency based on the optimized parameters to generate a time-frequency feature map. S3. Graph node construction: Divide the time-frequency feature map into regions, and use each region as a node in the graph structure to construct a node set. S4. Multi-dimensional node feature extraction: For the region corresponding to each node, extract multi-dimensional node feature vectors from the time-frequency feature map to characterize energy characteristics, change characteristics and transient characteristics. S5. Graph structure construction: Construct a graph structure based on the relationships between nodes and generate connection relationships for graph convolution operations. S6. Adaptive graph convolution feature extraction: The graph structure and node feature vectors are input into an adaptive graph convolution network. The neighborhood aggregation range of the graph convolution is dynamically adjusted by statistical analysis of the node features. A feature adjustment mechanism and a residual information preservation mechanism are introduced during the graph convolution process to obtain a graph-level feature representation for fault identification. S7. Fault identification and diagnosis output: Based on the graph-level feature representation, the operating status of the monorail crane bearing is classified, and the bearing fault diagnosis results are output.

2. The method for diagnosing single-rail crane bearing faults based on adaptive graph convolutional networks according to claim 1, characterized in that, The time-frequency feature generation includes: A set of parameters to be optimized for time-frequency analysis is set for the preprocessed acoustic signal; The parameter optimization algorithm is used to iteratively search the set of parameters to be optimized, and the optimal parameter combination is determined by using the bearing fault classification accuracy as the evaluation index. The acoustic signal is subjected to time-frequency transformation using the optimal parameter combination to generate a corresponding time-frequency feature map.

3. The method for diagnosing single-rail crane bearing faults based on adaptive graph convolutional networks according to claim 1, characterized in that, The graph node construction includes: The time-frequency feature map is segmented according to a region division rule to obtain multiple regions; Each region is treated as a node in a graph structure, and the corresponding time-frequency data for that region is obtained. Assign a unique identifier to each node and build a node set.

4. The method for diagnosing single-rail crane bearing faults based on adaptive graph convolutional networks according to claim 1, characterized in that, The node feature construction includes: For each node, calculate the characteristic quantity that can characterize the energy distribution of the corresponding regional data. Calculate characteristic quantities that reflect the magnitude or fluctuation of changes in the data of the region; The region data is used to calculate feature quantities that can characterize transient changes, and these feature quantities are combined to form a multidimensional feature vector of the node.

5. The method for diagnosing single-rail crane bearing faults based on adaptive graph convolutional networks according to claim 1, characterized in that, The graph structure construction includes: The connection relationships between nodes are determined based on the spatial relationship, regional boundary relationship, or feature similarity between adjacent nodes; A connection matrix is ​​constructed based on the aforementioned connection relationships to characterize the connectivity between nodes; The connection matrix is ​​normalized to obtain the graph structure matrix used for graph convolution operations.

6. The method for diagnosing single-rail crane bearing faults based on adaptive graph convolutional networks according to claim 1, characterized in that, The adaptive graph convolutional feature extraction includes: Perform global statistics on all node features to form a global feature vector that describes the overall graph characteristics; The global feature vector is input into the mapping module to generate dynamic control parameters for determining the order of the graph convolution polynomial; The polynomial order of the graph convolution is adaptively determined based on the dynamic control parameters, and the feature aggregation range of the graph convolution is adjusted.

7. The method for diagnosing single-rail crane bearing faults based on adaptive graph convolutional networks according to claim 6, characterized in that, The adaptive graph convolutional feature extraction also includes: Based on node features or intermediate features generated during graph convolution, adjustment parameters are generated to characterize the importance of node features of different orders. The contribution of node features of different orders or different channels is adaptively weighted according to the adjustment parameters.

8. The method for diagnosing single-rail crane bearing faults based on adaptive graph convolutional networks according to claim 7, characterized in that, The graph convolution process preserves the initial features or low-order aggregated features of the nodes; The initial features or low-order aggregated features are fused with the graph convolution output features to reduce the oversmoothing of features caused by multi-layer graph convolution.

9. A method for diagnosing single-rail crane bearing faults based on adaptive graph convolutional networks according to claim 7 or 8, characterized in that, The adaptive graph convolutional feature extraction also includes: Processing that dynamically determines the order of graph convolution polynomials based on global statistical features to adjust the range of feature aggregation; Adaptive weighting of convolutional features of different orders and fusion of initial features or low-order features are performed.

10. The method for diagnosing single-rail crane bearing faults based on adaptive graph convolutional networks according to claim 1, characterized in that, The fault identification and diagnosis output includes: The graph-level feature representation is input into the classification model for state determination; The fault type of the monorail crane bearing is determined based on the state discrimination result; Output the corresponding bearing fault diagnosis results.

Images