A method, system and medium for diagnosing rolling bearing faults in belt conveyors
By constructing a sample association graph and an unsupervised graph neural network GraphSAGE model, the problems of scarce fault samples and neglect of topological information in the fault diagnosis of rolling bearings of belt conveyors are solved, and high-precision fault identification is achieved under conditions of few labels.
Patent Information
- Application Number
- CN202611439212.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2026-08-14
- Filing Date
- 2026-09-16
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies for diagnosing rolling bearing faults in belt conveyors suffer from a scarcity of fault samples and high labeling costs. Furthermore, traditional methods neglect the topological information between samples, resulting in limited recognition accuracy.
By acquiring historical vibration signals of rolling bearings under different fault conditions, a sample association graph is constructed. The K-nearest neighbor algorithm is used to calculate Euclidean distance to construct similarity relationships. The unsupervised graph neural network GraphSAGE model is used for unsupervised learning to extract fault features and perform identification.
It enables accurate identification of rolling bearing faults with fewer labels, improves the accuracy of fault feature extraction and identification, and reduces the dependence on labeling costs.
Smart Images

Figure CN123065759A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fault identification, and in particular to a method, system and medium for diagnosing faults in rolling bearings of belt conveyors. Background Technology
[0002] Belt conveyors are core transportation equipment in coal mines, and their operational status directly affects the overall production efficiency and safety of the mine. As a core power and transmission component, the reducer bearings operate under harsh conditions of heavy loads, strong impacts, and high dust levels, bearing complex time-varying loads. They are the weakest link and most prone to failure in the entire system. Therefore, timely detection and accurate fault diagnosis of the reducer bearings are crucial for ensuring safe and efficient operation of coal mining, avoiding unplanned downtime, and reducing maintenance costs.
[0003] In recent years, data-driven methods, represented by deep learning, have greatly promoted the development of intelligent diagnostic technology due to their powerful nonlinear mapping and automatic feature extraction capabilities. However, this progress has brought a new "paradigm constraint": model performance is highly dependent on large-scale, high-quality labeled training sets. In coal mine industrial sites where belt conveyors are located, the long-term continuous operation of equipment leads to a scarcity of fault samples and high labeling costs, making the "label scarcity" problem particularly prominent. This limitation of "label dependence" has become the primary bottleneck for current intelligent diagnostic technology to move from the laboratory to the industrial field. Faced with this severe constraint, developing unsupervised learning has become an inevitable choice to drive technological progress in this field.
[0004] However, a deeper, often overlooked limitation lies in the fact that most traditional supervised methods and existing unsupervised methods implicitly assume "independent and identically distributed samples." They treat each collected signal as an isolated data point, completely ignoring the potential intrinsic correlations between samples during varying operating conditions and complex fault evolution. This neglect of topological information between samples leads to insufficient information mining, which in turn fundamentally limits the ceiling of recognition accuracy. Summary of the Invention
[0005] This invention provides a method, system, and medium for diagnosing rolling bearing faults in belt conveyors, to solve the aforementioned problems in the prior art, namely, how to accurately identify the fault type of rolling bearings in belt conveyors. This invention provides a method for diagnosing rolling bearing faults in belt conveyors, which includes: The original vibration signals of the rolling bearings of the belt conveyor under different fault conditions are obtained. The original vibration signals are divided into multiple continuous and uniform vibration signal samples, and the self-semantic features in the vibration signal samples are extracted. Based on the Euclidean distance between the acquired vibration signal samples, a sample association graph reflecting the intrinsic similarity is constructed; wherein, the sample association graph uses each vibration signal sample as a node, semantic features as node attributes, and the similarity relationship between vibration signal samples as edges; Random paths are generated on the sample association graph, and node pairs with similar relationships are selected as positive samples from the random paths. Unconnected node pairs are randomly selected from the sample association graph as negative samples. Aggregated features corresponding to positive and negative samples are obtained. By concatenating the aggregated features, the embedding representation of the node pair is determined and input into the fully connected layer for link prediction of the node pair. Unsupervised learning is performed on the embedding representation of the node pair to train the constructed unsupervised graph neural network GraphSAGE model. A pre-trained unsupervised graph neural network GraphSAGE model is used to extract fault features from the sample association graph of the rolling bearing of the belt conveyor under test, determine the corresponding unsupervised deep fault features, and use a classifier to identify the unsupervised deep fault features to obtain the final fault type identification result.
[0006] Optionally, the step of constructing a sample association graph reflecting intrinsic similarity based on the Euclidean distance between the acquired vibration signal samples specifically includes: For each vibration signal sample To determine its relationship with other vibration signal samples Euclidean distance between ; According to Euclidean distance By combining the sample association edges E determined based on the similarity threshold SI, a sample association graph reflecting intrinsic similarity is obtained; wherein, the vibration signal samples are calculated using the following formula. Other vibration signal samples Euclidean distance between : ; Where n is the number of vibration signal samples.
[0007] Optionally, the KNN algorithm can be used to determine the Euclidean distance between the vibration signal samples.
[0008] Optionally, the loss function of the unsupervised graph neural network GraphSAGE model adopts the form of binary cross-entropy.
[0009] Optionally, obtaining the binary cross-entropy specifically includes: The binary cross-entropy loss is obtained as follows: ; in, The number of selected node pairs, Indicates positive sample label, This represents the probability of a positive correlation.
[0010] Optionally, the classifier is a multilayer perceptron classifier.
[0011] Optionally, the unsupervised graph neural network GraphSAGE model has two network layers, the GCN has three layers, the activation function is ReLU, the optimizer is Adam, and the learning rate is 0.001.
[0012] Optionally, the fault type identification results include normal state, inner ring fault, outer ring fault and rolling element fault; the self-semantic features include maximum value, minimum value, average value, variance, root mean square and peak-to-peak value.
[0013] This invention provides a fault diagnosis system for rolling bearings of belt conveyors, comprising: The acquisition module is used to acquire the historical raw vibration signals of the rolling bearings of the belt conveyor under different fault conditions, divide the raw vibration signals into multiple continuous and uniform vibration signal samples, and extract the self-semantic features from the vibration signal samples. The sample association graph construction module is used to construct a sample association graph reflecting the intrinsic similarity based on the Euclidean distance between the acquired vibration signal samples; wherein, the sample association graph uses each vibration signal sample as a node, semantic features as node attributes, and the similarity relationship between vibration signal samples as edges; The model training module generates random paths on the sample association graph, selects node pairs with similar relationships as positive samples from the random paths, and randomly selects unconnected node pairs from the sample association graph as negative samples. It obtains the aggregated features corresponding to the positive and negative samples, concatenates the aggregated features to determine the embedding representation of the node pair, and inputs it into the fully connected layer to predict the link of the node pair, so as to perform unsupervised learning on the embedding representation of the node pair and train the constructed unsupervised graph neural network GraphSAGE model. The fault identification module uses a pre-trained unsupervised graph neural network GraphSAGE model to extract fault features from the sample association graph of the rolling bearing of the belt conveyor under test, determine the corresponding unsupervised deep fault features, and use a classifier to identify the unsupervised deep fault features to obtain the final fault type identification result.
[0014] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for diagnosing rolling bearing faults in belt conveyors.
[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a method for diagnosing rolling bearing faults in belt conveyors. By acquiring the historical original vibration signals of the rolling bearings of belt conveyors under different fault states, this method can reveal the state change patterns of the bearings at different operating stages, thereby achieving an accurate description and identification of the equipment's operating state and providing reliable data support for subsequent fault identification. Simultaneously, the method uses the K-Nearest Neighbors (KNN) algorithm to calculate the Euclidean distance between vibration signal samples and adaptively constructs a fault relationship graph reflecting intrinsic similarity. Through unsupervised training of a two-layer GraphSAGE model, this mechanism effectively mines the topological structure information between samples missed by traditional independent measurement point analysis, enabling the extraction of accurate fault features. Finally, the unsupervised learned feature vector is fed into a traditional machine learning classifier, enabling accurate identification of the bearing fault type in the reducer. Attached Figure Description
[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.
[0017] Figure 1 A flowchart of a method for diagnosing rolling bearing faults in a belt conveyor provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the construction of a sample association graph based on the K-nearest neighbor algorithm provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the GraphSAGE neighbor sampling and feature aggregation process provided in an embodiment of the present invention; Figure 4 This is a flowchart of the unsupervised GraphSAGE model training process provided in an embodiment of the present invention; Figure 5 A schematic diagram of the Case Western Reserve University bearing dataset acquisition device provided in an embodiment of the present invention; Figure 6 This is a confusion matrix diagram of GraphSAGE recognition results under different labels provided in embodiments of the present invention; Figure 7 This is a schematic diagram of feature visualization analysis under 5% labeled samples provided in an embodiment of the present invention. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0019] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.
[0020] Figure 1 This is a flowchart of a method for diagnosing rolling bearing faults in a belt conveyor provided by an embodiment of the present invention, such as... Figure 1 As shown in the figure, this embodiment illustrates a method for diagnosing rolling bearing faults in a belt conveyor, including: S1: Obtain the original vibration signals of the rolling bearings of the belt conveyor under different fault conditions, divide the original vibration signals into multiple continuous and uniform vibration signal samples, and extract the self-semantic features from the vibration signal samples.
[0021] For example, this invention takes the rolling bearing of a belt conveyor as the research object and extracts its time-domain features as the self-semantic features of the signal. Time-domain features can effectively characterize the dynamic changes of the bearing vibration signal in the time dimension, reflecting basic information about the signal in terms of amplitude, frequency, period, and duration. By extracting and analyzing these time-series features, the state change law of the bearing in different operating stages can be revealed, thereby achieving an accurate description and identification of the equipment's operating state. The specific time-domain feature parameters used and their calculation formulas are shown in Table 1.
[0022] Table 1 Time-domain characteristic parameters Feature parameters expression Feature parameters expression Maximum value average amplitude Minimum value Root mean square amplitude average value Peak factor variance Waveform factor Root Mean Square Impulsive factors Peak-to-peak value gap factor peak Kurtosis factor S2: Based on the Euclidean distance between the acquired vibration signal samples, a sample association graph reflecting the intrinsic similarity is constructed; wherein, the sample association graph takes each vibration signal sample as a node, the semantic features as node attributes, and the similarity relationship between vibration signal samples as edges.
[0023] For example, in the fault diagnosis of rolling bearings of belt conveyors, firstly, a fault sample is generated for each fault state by continuously and uniformly sampling the signal; secondly, for each fault sample, this section extracts its self-semantic features of the rolling bearing fault sample of the belt conveyor; finally, the signal sample is set as a node, and sample association edges are constructed based on the similarity between samples, thereby establishing the association information between labeled samples and unlabeled samples.
[0024] Specifically, firstly, this section uses the K-nearest neighbor algorithm to construct an association graph from signal samples, as shown in the diagram below. Figure 2 As shown, the sample correlation diagram contains l fault states and all signal samples. Where n is the total number of fault samples, based on the self-semantic features of the rolling bearing fault samples of the belt conveyor, the K-nearest neighbor algorithm is used to construct a sample association graph G(V,E), where the node set V represents all signal samples. Furthermore, the semantic features of each fault sample are used as node attributes. Secondly, the Euclidean distance Sim(sᵢ,s) between any two fault sample pairs (sᵢ,s) is calculated, where i=1,2,...,n and j=1,2,...,n. Based on the nearest neighbor parameter K (generally 4), the sample association edge E is obtained, and then the sample adjacency matrix A is constructed. If the Euclidean distance Sim(sᵢ, s) between sample sᵢ and sample s is one of the K closest samples to sample sᵢ, then there is an association edge between sample sᵢ and sample s; otherwise, there is no association edge.
[0025] Specifically, firstly, for each fault sample Calculate its comparison with other fault samples Euclidean distance between The calculation process is shown in the following formula: ; Secondly, based on the similarity threshold SI, the sample association edges E are obtained. Based on the calculated Euclidean distance, for each sample... Find the K nearest samples as its neighbors and establish the associated edges. Finally, by doing this for all samples, a complete sample association graph is constructed.
[0026] like Figure 2 As shown, the constructed graph not only contains the self-semantic information of the samples but also reveals the proximity relationships between samples. These relationships implied in the graph help the model learn the intrinsic connections between samples (labeled and unlabeled), thereby improving the accuracy of fault identification in scenarios with limited sample labels.
[0027] S3: Random paths are generated on the sample association graph, and node pairs with similar relationships are selected as positive samples from the random paths. Unconnected node pairs are randomly selected from the sample association graph as negative samples. Aggregated features corresponding to positive and negative samples are obtained. By concatenating the aggregated features, the embedding representation of the node pair is determined and input into the fully connected layer for link prediction of the node pair. Unsupervised learning is performed on the embedding representation of the node pair to train the constructed unsupervised graph neural network GraphSAGE model.
[0028] Generally, GraphSAGE is an inductive learning framework that uses node attribute information to generate node embedding features. The core idea of GraphSAGE is to learn an aggregation function that generates the embedding features of the target node by integrating the feature representations of neighboring nodes. This method consists of two steps: sampling neighboring nodes and aggregating sample node features. Figure 3 This demonstrates the specific flow of these two steps, and their implementation may include:
[0029] (1) Sample node neighbor sampling First, a fixed number of random samples are taken from neighboring nodes layer by layer, so as to... Figure 1 target node Taking a circle as an example, let its first-order neighbor N be defined. 1 (s i (square) and its second-order neighbor N 2 (s i The sampling quantities for the triangle are 10 and 5 respectively, and the number of first-order neighbors after sampling does not exceed 10, while the number of second-order neighbors does not exceed 10 × 5 = 50. Finally, around the target node s i The computation is confined to a two-layer subgraph with a limited number of nodes (consisting of the target node, first-order neighbors, and second-order neighbors), which effectively reduces computational overhead.
[0030] (2) Sample node feature aggregation After sampling to obtain the computational subgraph, feature aggregation is performed. The propagation direction of feature aggregation is opposite to that of neighbor sampling, that is, information is generated from the second-order neighbor N. 2 (s i Propagated layer by layer to the first-order neighbor N 1 (s i ), eventually converging at the target node s i The calculation process for feature aggregation is shown in the following formula.
[0031] (1) (2) (3) (4) Specifically, let s i Let n1 be the target node, and n2 be any first-order neighbor of n1. First, according to the formula... and formula Calculate the representation vector of the first-order neighbor n1 , where the formula The features of all n1 neighbors n2 are calculated by mean aggregation to obtain the aggregated vector. . formula Aggregate the eigenvectors of n1 itself with the aggregate vectors of its neighbors. The components are concatenated, and the final representation of n1 is generated using a ReLU nonlinear activation function. .
[0032] Similarly, the target node s i Embedded It can be derived from the formula and formula It was calculated that, after obtaining the representations of all first-order neighbors... Then, the formula Perform mean aggregation on it to obtain a vector. Finally, the formula Node s i Aggregate representation of its own features and its neighbors The nodes are spliced together and subjected to a nonlinear transformation to calculate the target node. The final embedding .
[0033] In summary, GraphSAGE learns the embedding representation of the target node by aggregating the self-semantic information of neighboring nodes. This process effectively integrates the node's own features with the local structural information of the graph.
[0034] For example, this invention models the constructed sample association graph using GraphSAGE to learn the potential fault characteristics of each fault sample. By aggregating the self-semantic information (temporal features) of neighboring samples layer by layer, the model can simultaneously capture the self-semantic information of nodes and the association information between samples. In the established GraphSAGE network, and These are the network parameters to be trained to obtain better fault feature representations of sample nodes. Since obtaining labeled samples is difficult and costly in real-world fault diagnosis scenarios, this section uses an unsupervised method to train GraphSAGE. The training process is as follows: Figure 4 As shown.
[0035] Specifically, this invention first performs a random walk on the sample association graph to generate random paths and selects node pairs as positive samples. Then, it randomly selects unconnected node pairs from the graph as negative samples, maintaining the same number as the positive samples. Next, the embedding representations of these node pairs (learned by GraphSAGE) are concatenated and input into a fully connected layer to predict the label of the node pairs (whether they are positive or negative samples). Thus, the model achieves unsupervised learning of node embeddings by distinguishing between positive and negative node pairs.
[0036] Overall, the constructed unsupervised GraphSAGE model consists of a two-layer GraphSAGE network and a node pair classification module. The GraphSAGE network learns node embeddings, while the node pair classification module utilizes edge existence information to construct the training objective. The loss function adopts the binary cross-entropy form, as shown in the following equation:
[0037] ; in, The number of selected node pairs, Indicates positive sample label, This represents the probability of a positive correlation.
[0038] S4: Using a pre-trained unsupervised graph neural network GraphSAGE model, fault features are extracted from the sample association graph of the rolling bearing of the belt conveyor under test, the corresponding unsupervised deep fault features are determined, and a classifier is used to identify the unsupervised deep fault features to obtain the final fault type identification result.
[0039] For example, this invention can obtain unsupervised deep fault features of fault samples by applying unsupervised GraphSAGE to the constructed sample association graph. Based on this, these unsupervised features, combined with a small amount of label information, are used to train a simple classifier (such as a multilayer perceptron classifier) for fault type identification. In summary, the fault diagnosis process for rolling bearings of belt conveyors based on unsupervised graph neural networks is as follows:
[0040] (1) First, the vibration acceleration signals of the rolling bearings of the belt conveyor under various fault conditions were collected. The original vibration signals were divided into multiple sub-samples of length 1024 using non-overlapping sampling;
[0041] (2) Extract the semantic features of the rolling bearing signals of the belt conveyor from the subsamples; (3) Treat the subsamples as nodes of the graph and calculate the Euclidean distance between all subsamples. For any subsample, select its K nearest neighbor samples as neighbors. Then, treat all signal samples as nodes and set their signal semantic features as node attributes; at the same time, connect each node and its K neighbor nodes to form edges, thereby constructing a sample association graph.
[0042] (4) Unsupervised GraphSAGE Model Construction: First, a two-layer GraphSAGE model is built on the constructed sample association graph, and node embedding learning is achieved through node sampling and feature aggregation. Second, a random walk is performed on the sample association graph, and node pairs are selected as positive samples. Then, unconnected node pairs are randomly selected from the graph as negative samples, with the number being the same as the number of positive samples. Finally, the embedded representation of the node pair is obtained by concatenating the aggregated features of each node pair sample, and then it is input into the fully connected layer to complete the link prediction of the node pair. Under the supervision of link labels, GraphSAGE can be trained in unsupervised mode. Finally, the trained model can extract the deep fault feature representation of each signal sample.
[0043] (5) Fault diagnosis of rolling bearings of belt conveyors: Based on the unsupervised deep fault features of samples obtained by the unsupervised GraphSAGE model, a simple classifier (such as a multilayer perceptron classifier) is trained with a small amount of label information to complete the identification of rolling bearing fault types of belt conveyors.
[0044] For example, after constructing the sample association graph, each vibration signal sample is treated as a graph node, its semantic features are used as node attributes, and similarity association edges between samples are established based on the KNN algorithm. Subsequently, a two-layer unsupervised GraphSAGE model is constructed on the sample association graph, learning node embedding representations through neighbor sampling and feature aggregation. The first layer mainly aggregates the feature information of the first-order neighbors of the target node, and the second layer further integrates higher-order neighborhood structure information, so that the representation of each sample node simultaneously includes its own vibration features and the association information of neighboring samples.
[0045] During model training, instead of directly using fault category labels, a node pair prediction task is constructed based on the sample association graph. Specifically, node pairs with similar relationships are selected as positive samples through random walks or existing association edges, while unconnected or weakly associated node pairs are randomly selected as negative samples. The embedding vectors obtained from GraphSAGE of the two nodes in a node pair are concatenated to form a node pair embedding representation, which is then input into a fully connected layer to determine whether a relationship exists between the node pairs. The model is trained using positive and negative nodes for the classification task and the binary cross-entropy loss function, thereby learning the latent structural relationships and fault feature distributions between samples under unsupervised conditions.
[0046] After training, the GraphSAGE feature extraction part is retained to output the unsupervised deep fault features of each vibration signal sample, and these features are input into the classifier to complete the fault type identification.
[0047] Employing a two-layer GraphSAGE architecture further enhances the ability to represent fault features. Compared to a single-layer structure, the two-layer structure not only aggregates first-order neighborhood information of directly similar samples but also introduces indirect association information from second-order neighborhoods, enabling the model to more fully explore the potential topological relationships between samples and enhancing the robustness and discriminativeness of features. Simultaneously, the two-layer structure achieves a good balance between expressive power and computational complexity, avoiding the problem of overly smoothed features caused by excessively deep networks, which is beneficial for improving the accuracy and generalization ability of rolling bearing fault identification under conditions with few labels.
[0048] For example, to verify the effectiveness of the method of the present invention, the experiment used the Case Western Reserve University (CWRU) bearing dataset and the non-drive end bearing data of a belt conveyor motor in a mine to evaluate the performance of the intelligent fault diagnosis method. The experimental setup for the CWRU bearing dataset is as follows: Figure 5 (a) and Figure 5 As shown in (b), the main components consist of an induction motor, a torque sensor, a dynamometer, and the bearings under test at the drive end and fan end. Vibration signals were collected using accelerometers mounted on the bearing housing. This study selected data collected from the drive end bearing under no-motor load (0HP) conditions, covering four health states: normal, inner ring fault, outer ring fault, and rolling element fault. Each fault type included three different damage diameters: 0.178 mm, 0.356 mm, and 0.533 mm. The data was divided into 10 categories: one normal state and nine different types and degrees of fault states. The dataset contained 1196 samples, with each category consisting of 119 or 120 samples, and each sample consisting of 1024 data points. A detailed description of the dataset is shown in Table 2. The belt conveyor motor bearing dataset includes normal, inner ring, and outer ring fault conditions, with 360 and 72 samples in the training and test sets, respectively, including 24 samples for each fault type.
[0049] Table 2. Description of CWRU Fault Data Fault diameter / mm Fault location Number of training / test samples Label — normal 96 / 24 0 0.178 Inner circle 96 / 23 1 0.356 Inner circle 96 / 24 2 0.533 Inner circle 96 / 24 3 0.178 Outer ring 96 / 23 4 0.356 Outer ring 96 / 23 5 0.533 Outer ring 96 / 23 6 0.178 Rolling body 96 / 24 7 0.356 Rolling body 96 / 24 8 0.533 Rolling body 96 / 24 9 For example, the GraphSAGE model is built on the PyTorch 2.0.1 deep learning framework, with Python interpreter version 3.9.17 and CUDA version 12.2. Some important parameter configurations for the experiment are shown in Table 3.
[0050] Table 3 Experimental Parameter Configuration Parameter name GCN layer Activation function Optimizer Learning rate Parameter value 3 Relu Adam 0.001 For example, to verify the effectiveness of the belt conveyor reducer fault diagnosis method based on unsupervised graph neural networks (GraphSAGE) proposed in this invention, five typical comparative algorithms were selected: Principal Component Analysis (PCA), Kernel Principal Component Analysis (KPCA), Autoencoder (AE), Variational Autoencoder (VAE), and Graph Autoencoder (GAE). These methods represent different technical approaches from linear dimensionality reduction and nonlinear mapping to deep feature learning and graph structure modeling, and can comprehensively reflect the performance differences of different algorithms in the belt conveyor reducer fault diagnosis task.
[0051] Principal Component Analysis (PCA) is a commonly used linear dimensionality reduction method. It extracts key eigenvalues by decomposing the eigenvalue covariance matrix, mapping high-dimensional vibration signals to a low-dimensional space to achieve feature compression and pattern recognition. PCA can extract the main energy characteristics of each measuring point to reflect state changes under different operating conditions. This method is simple to implement, computationally efficient, and suitable for rapid health status assessment and preliminary fault identification.
[0052] KPCA: Kernel Principal Component Analysis (KPCA) introduces a kernel function on top of PCA to nonlinearly map the original signal to a high-dimensional feature space, thereby revealing the potential nonlinear relationships in the vibration signal. KPCA can effectively distinguish nonlinear fault features such as gear cracks and bearing spalling, achieving more refined feature representation and providing more discriminative input features for subsequent model learning.
[0053] Autoencoders (AEs) are unsupervised feature learning methods based on neural networks. An encoder compresses the input signal into a latent feature representation, which is then reconstructed by a decoder, enabling automatic feature extraction. AEs can adaptively extract deep features representing health states and fault modes from multi-channel vibration signals, helping to reduce manual feature design and improve the stability and accuracy of feature representation.
[0054] Variational Autoencoders (VAEs) model the latent features of input signals by introducing probability distribution assumptions and utilize reconstruction errors to achieve anomaly identification. VAEs can learn the normal distribution characteristics of the system from healthy signal samples. When abnormal operating conditions such as bearing wear or gear breakage occur, the reconstruction error increases, thus enabling unsupervised detection of potential faults. This method has strong feature learning capabilities, providing a reliable foundation for intelligent fault diagnosis.
[0055] Graph Autoencoder (GAE) combines graph neural networks with autoencoder concepts. The encoder learns node embeddings, and the decoder reconstructs the relationships between nodes, enabling unsupervised feature extraction from graph-structured data. GAE can treat monitoring points such as input shafts, output shafts, gears, and bearings as graph nodes, with edges representing the coupling relationships between signals. It can comprehensively analyze the topological dependencies and dynamic characteristics between components, achieving system-level health status identification and providing crucial support for structured intelligent diagnosis of complex mechanical systems.
[0056] For example, to verify the effectiveness of the proposed unsupervised graph neural network (GraphSAGE) model in fault diagnosis under limited labeled sample conditions, multiple experiments were conducted on the Case Western Reserve University bearing dataset. The experiments used 5%, 15%, and 25% of labeled samples for training, respectively, while the remaining unlabeled samples were used for unsupervised feature learning. To reduce the impact of randomness, each experiment was repeated 10 times, and the average value was taken as the final diagnostic result. This invention compares and analyzes the proposed method with five mainstream algorithms (PCA, KPCA, AE, VAE, and GAE).
[0057] The results are shown in Table 4. Compared with other methods, the method proposed in this invention has the highest recognition accuracy, with an average improvement of 48.3%, 45.2%, 27.2%, 8.4%, and 4.1% compared with PCA, KPCA, AE, VAE, and GAE methods, respectively. As the proportion of labeled samples increases, the fault recognition accuracy of each model shows an upward trend. The method of this invention has good recognition effect with a label ratio of 5%, which proves the advantages of the method of this invention in unsupervised learning with few labels.
[0058] Table 4. Diagnostic results of each model under different scale labels Figure 6 This demonstrates the variation of the confusion matrix of the GraphSAGE model in bearing fault diagnosis tasks under different labeled sample ratios. From... Figure 6As can be seen, when only 5% of the samples are labeled, the model can already distinguish the ten fault categories quite well. The values on the main diagonal are significantly higher than in other areas, indicating that most samples are correctly classified, but a small number of misclassifications still exist. As the proportion of labeled samples increases, the model's classification accuracy improves significantly, and the number of misclassified samples gradually decreases. When the labeling ratio reaches 15%, most fault categories are correctly identified, indicating that the model can effectively learn the features of each category even with low sample conditions. When the labeled sample ratio reaches 25% or higher, the main diagonal of the confusion matrix is significantly highlighted, and the off-diagonal elements are close to zero, with the model achieving a recognition rate of 98.7% for the ten fault categories. The results show that the GraphSAGE model based on unsupervised feature learning and graph structure modeling can fully mine the topological association information in unlabeled data and maintain excellent diagnostic performance even with very few labeled samples.
[0059] Overall, as the annotation ratio increases, the model gradually transitions from local feature aggregation to global structural representation, demonstrating good stability, robustness, and generalization ability, providing effective support for intelligent and low-sample fault diagnosis of belt conveyor reducers.
[0060] To more intuitively display the feature distribution of the model under small sample sizes, T-SEN plots were used to complete the feature classification display of each model's output layer under 15% sample size; among them, Figure 7 (a) to Figure 7 (e) Displays the feature classifications using PCA, KPCA, AE, VAE, and GAE methods, respectively. Figure 7 As can be seen from (f), the GraphSAGE model has more compact features among the same fault classes and the greatest differences among different faults. Therefore, the model has the highest recognition accuracy with 15% labeled samples.
[0061] Multiple experiments were conducted on the belt conveyor bearing dataset. The experiments used 5%, 15%, and 25% of the labeled samples for training, respectively, while the remaining unlabeled samples were used for unsupervised feature learning. The results are shown in Table 5.
[0062] Table 5. Diagnostic results of each model under different scale labels The results are shown in Table 5. Under different few-label samples, the method proposed in this invention has the highest recognition accuracy compared with other methods. Compared with PCA, KPCA, AE, VAE and GAE methods, the model accuracy is improved by 43%, 34.7%, 25%, 13.8% and 9.7%, respectively, which proves the advantages of the method of this invention in few-label and unsupervised learning.
[0063] In summary, in the fault diagnosis of rolling bearings in belt conveyors, PCA and KPCA mainly rely on the dimensionality-reduced feature distribution for classification. While they can reflect the differences in fault characteristics to some extent, their ability to express complex nonlinear relationships is limited. AE and VAE automatically extract deep features of signals through autoencoders, which can better describe the feature distribution under different working conditions. VAE, in particular, performs well in anomaly detection, but fails to utilize the structural information between measurement points. GAE introduces graph structure learning on this basis, enabling the characterization of correlations between node features and significantly improving diagnostic accuracy. However, the GraphSAGE model based on unsupervised graph neural networks proposed in this invention further improves upon this, revealing the topological similarity of vibration signals from multiple measurement points through KNN graph construction, and achieving feature self-learning and structural correlation modeling under unsupervised conditions using neighborhood sampling and feature aggregation strategies. This method fully utilizes the potential structural information in unlabeled data, enabling the extraction of more discriminative graph embedding features, thus achieving high-precision bearing fault identification even with few samples. Experimental results show that the GraphSAGE model has excellent stability and generalization performance, providing an efficient and reliable new approach for intelligent and unsupervised fault diagnosis of belt conveyors.
[0064] This invention addresses the challenges of scarce fault sample labels and the neglect of topological relationships between samples in traditional diagnostic methods for scraper conveyor reducers under complex operating conditions. It proposes a novel fault feature extraction and diagnosis method based on unsupervised graph neural networks (GraphSAGE). Extensive comparative experiments on the Case Western Reserve University bearing public dataset and a self-built belt conveyor motor bearing dataset yield the following main conclusions:
[0065] (1) This study breaks through the dual dependence of traditional intelligent diagnostic methods on massive labels and the assumption of "sample independence". It innovatively decouples feature extraction from fault classification. First, it uses the K-Nearest Neighbors (KNN) algorithm to calculate the Euclidean distance between vibration signal samples, adaptively constructing a fault relationship graph reflecting intrinsic similarity. Then, under label-free conditions, it drives the GraphSAGE model to undergo unsupervised training through proxy tasks such as link prediction. This mechanism effectively mines the topological information between samples missed by traditional independent measurement point analysis, providing a new paradigm for unsupervised intelligent diagnosis.
[0066] (2) The method exhibits excellent diagnostic accuracy and strong cross-dataset generalization ability under conditions of few labeled samples. Experimental results show that the extracted graph embedding features have extremely high discriminative power under the multilayer perceptron classifier. Compared with five mainstream baseline models such as PCA, KPCA, AE, VAE and GAE, the method of this invention performs best under different proportions of labeled samples. In particular, under the extremely low resource condition of only 5%, the diagnostic accuracy of the model on the CWRU dataset and the self-built dataset is still as high as 91.5% and 87.5%, respectively; while under 25% labeled samples, the accuracy of the two increases to 98.7% and 100%, respectively. This quantitative result significantly reduces the dependence of model development on the cost of manual annotation and has extremely strong robustness.
[0067] (3) Unsupervised graph embedding features exhibit excellent clustering distribution and engineering interpretability. Visual analysis of the confusion matrix and t-SNE features under different label ratios confirms that the GraphSAGE model proposed in this invention can effectively penetrate complex working condition noise, enabling different fault categories to present an ideal clustering state of "compact within classes and separated between classes" in the feature space. This not only intuitively verifies the model's powerful high-dimensional feature deconstruction capability but also provides highly reliable technical support for intelligent condition monitoring and daily low-sample operation and maintenance of complex mechanical equipment such as belt conveyors in coal mines.
[0068] The above are one or more embodiments of the belt conveyor rolling bearing fault diagnosis method provided in this specification. Based on the same idea, this specification also provides a corresponding belt conveyor rolling bearing fault diagnosis system, including: The acquisition module is used to acquire the historical raw vibration signals of the rolling bearings of the belt conveyor under different fault conditions, divide the raw vibration signals into multiple continuous and uniform vibration signal samples, and extract the self-semantic features from the vibration signal samples. The sample association graph construction module is used to construct a sample association graph reflecting the intrinsic similarity based on the Euclidean distance between the acquired vibration signal samples; wherein, the sample association graph uses each vibration signal sample as a node, semantic features as node attributes, and the similarity relationship between vibration signal samples as edges; The model training module generates random paths on the sample association graph, selects node pairs with similar relationships as positive samples from the random paths, and randomly selects unconnected node pairs from the sample association graph as negative samples. It obtains the aggregated features corresponding to the positive and negative samples, concatenates the aggregated features to determine the embedding representation of the node pair, and inputs it into the fully connected layer to predict the link of the node pair, so as to perform unsupervised learning on the embedding representation of the node pair and train the constructed unsupervised graph neural network GraphSAGE model. The fault identification module uses a pre-trained unsupervised graph neural network GraphSAGE model to extract fault features from the sample association graph of the rolling bearing of the belt conveyor under test, determine the corresponding unsupervised deep fault features, and use a classifier to identify the unsupervised deep fault features to obtain the final fault type identification result.
[0069] Specific limitations regarding the fault diagnosis system for rolling bearings of belt conveyors can be found in the limitations of the fault diagnosis method for rolling bearings of belt conveyors mentioned above, and will not be repeated here. Each module in the aforementioned fault diagnosis system for rolling bearings of belt conveyors can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.
[0070] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described method for diagnosing rolling bearing faults in belt conveyors.
[0071] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.
Claims
1. A method for diagnosing rolling bearing faults in a belt conveyor, characterized in that, include: The original vibration signals of the rolling bearings of the belt conveyor under different fault conditions are obtained. The original vibration signals are divided into multiple continuous and uniform vibration signal samples, and the self-semantic features in the vibration signal samples are extracted. Based on the Euclidean distance between the acquired vibration signal samples, a sample association graph reflecting the intrinsic similarity is constructed; wherein, the sample association graph uses each vibration signal sample as a node, semantic features as node attributes, and the similarity relationship between vibration signal samples as edges; Random paths are generated on the sample association graph, and node pairs with similar relationships are selected as positive samples from the random paths. Unconnected node pairs are randomly selected from the sample association graph as negative samples. Aggregated features corresponding to positive and negative samples are obtained. By concatenating the aggregated features, the embedding representation of the node pair is determined and input into the fully connected layer for link prediction of the node pair. Unsupervised learning is performed on the embedding representation of the node pair to train the constructed unsupervised graph neural network GraphSAGE model. A pre-trained unsupervised graph neural network GraphSAGE model is used to extract fault features from the sample association graph of the rolling bearing of the belt conveyor under test, determine the corresponding unsupervised deep fault features, and use a classifier to identify the unsupervised deep fault features to obtain the final fault type identification result.
2. The method for diagnosing rolling bearing faults in a belt conveyor as described in claim 1, characterized in that, The method of constructing a sample association graph reflecting intrinsic similarity based on the Euclidean distance between the acquired vibration signal samples specifically includes: For each vibration signal sample To determine its relationship with other vibration signal samples Euclidean distance between ; According to Euclidean distance By combining the sample association edges E determined based on the similarity threshold SI, a sample association graph reflecting intrinsic similarity is obtained; wherein, the vibration signal samples are calculated using the following formula. Other vibration signal samples Euclidean distance between : ; Where n is the number of vibration signal samples.
3. The method for diagnosing rolling bearing faults in a belt conveyor as described in claim 1, characterized in that, The KNN algorithm was used to determine the Euclidean distance between the vibration signal samples.
4. The method for diagnosing rolling bearing faults in a belt conveyor as described in claim 1, characterized in that, The loss function of the unsupervised graph neural network GraphSAGE model adopts the form of binary cross-entropy.
5. The method for diagnosing rolling bearing faults in a belt conveyor as described in claim 4, characterized in that, The acquisition of the binary cross-entropy specifically includes: The binary cross-entropy loss is obtained as follows: ; in, The number of selected node pairs, Indicates positive sample label, This represents the probability of a positive correlation.
6. The method for diagnosing rolling bearing faults in a belt conveyor as described in claim 1, characterized in that, The classifier is a multilayer perceptron classifier.
7. The method for diagnosing rolling bearing faults in a belt conveyor as described in claim 1, characterized in that, The unsupervised graph neural network GraphSAGE model has two network layers, the GCN has three layers, the activation function is ReLU, the optimizer is Adam, and the learning rate is 0.
001.
8. The method for diagnosing rolling bearing faults in a belt conveyor as described in claim 1, characterized in that, The fault type identification results include normal state, inner ring fault, outer ring fault and rolling element fault; the self-semantic features include maximum value, minimum value, average value, variance, root mean square and peak-to-peak value.
9. A fault diagnosis system for rolling bearings of a belt conveyor, characterized in that, include: The acquisition module is used to acquire the historical raw vibration signals of the rolling bearings of the belt conveyor under different fault conditions, divide the raw vibration signals into multiple continuous and uniform vibration signal samples, and extract the self-semantic features from the vibration signal samples. The sample association graph construction module is used to construct a sample association graph reflecting the intrinsic similarity based on the Euclidean distance between the acquired vibration signal samples; wherein, the sample association graph uses each vibration signal sample as a node, semantic features as node attributes, and the similarity relationship between vibration signal samples as edges; The model training module generates random paths on the sample association graph, selects node pairs with similar relationships as positive samples from the random paths, and randomly selects unconnected node pairs from the sample association graph as negative samples. It obtains the aggregated features corresponding to the positive and negative samples, concatenates the aggregated features to determine the embedding representation of the node pair, and inputs it into the fully connected layer to predict the link of the node pair, so as to perform unsupervised learning on the embedding representation of the node pair and train the constructed unsupervised graph neural network GraphSAGE model. The fault identification module uses a pre-trained unsupervised graph neural network GraphSAGE model to extract fault features from the sample association graph of the rolling bearing of the belt conveyor under test, determine the corresponding unsupervised deep fault features, and use a classifier to identify the unsupervised deep fault features to obtain the final fault type identification result.
10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method for diagnosing rolling bearing faults in belt conveyors as described in any one of claims 1-8.
