A bearing zero-sample composite fault diagnosis method and diagnosis system
By combining multimodal data acquisition and fine-grained feature extraction with MSVDD and MultiLabel-GCN models, the problems of manual feature dependence and complex fault identification in zero-sample bearing fault diagnosis are solved, achieving accurate bearing fault diagnosis and adaptive fault attribute modeling, thus improving the accuracy of diagnosis and the robustness of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SICHUAN AGRI UNIV
- Filing Date
- 2026-07-07
- Publication Date
- 2026-08-04
AI Technical Summary
Existing zero-sample fault diagnosis methods for bearings rely on manual feature extraction, which makes it difficult to identify complex faults and results in mixed feature coupling, leading to a decline in diagnostic performance.
By combining multimodal data acquisition, fine-grained feature extraction, an improved support vector data description model (MSVDD) with a multi-label graph convolutional network (MultiLabel-GCN), and through adaptive feature knowledge base construction and attribute decoupling, we can achieve accurate differentiation between single faults and compound faults.
It improves the accuracy and stability of bearing fault diagnosis, enhances the system's generalization ability and applicability under varying operating conditions, and significantly improves the diagnostic accuracy and robustness of complex faults.
Smart Images
Figure CN122508301A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent mechanical operation and maintenance technology, and in particular to a method and system for diagnosing composite faults in bearings with zero samples. Background Technology
[0002] With the rapid development of intelligent manufacturing, the Industrial Internet of Things (IIoT), and big data technologies, the machinery and equipment industry is evolving towards high-end, intelligent, and large-scale production. Rolling bearings, as key supporting components of rotating machinery, are widely used in core industrial fields such as wind power, rail transportation, and precision machine tools. They endure harsh environments such as alternating loads, high temperatures, and dust for extended periods, making them prone to early failures such as fatigue spalling, wear, and cracks. Their health directly determines the operating efficiency and production safety of the entire equipment. Therefore, achieving accurate, efficient, and intelligent diagnosis of rolling bearing failures in industrial scenarios is crucial for promoting predictive maintenance of industrial equipment and improving production reliability and economic efficiency.
[0003] In recent years, zero-shot diagnostics (ZSD) has received increasing attention in the field of industrial equipment fault diagnosis due to its ability to identify faults of unknown categories using faults of known categories, even without any target fault samples. However, most ZSD methods still rely on manual knowledge to extract fault features, requiring the construction of different feature libraries for different operating conditions. Furthermore, under complex and long-term operating conditions, rotating bearings subjected to multiple sources of pressure and load can develop compound faults, which are much more difficult to accurately identify and isolate than single faults. Finally, the coupling effects between various features introduce confounding effects, reducing the performance of zero-shot fault diagnosis methods based on attribute prediction. Summary of the Invention
[0004] The purpose of this invention is to provide a method and system for diagnosing composite faults in bearings with zero samples, which solves the problems of existing bearing zero-sample fault diagnosis relying on manual extraction of fault features and requiring adjustment of the feature library according to operating conditions, while also making it difficult to identify composite faults in rotating bearings, and the hybrid effect of feature coupling reducing the performance of the diagnostic method.
[0005] The objective of this invention is achieved through the following technical solution: A method for diagnosing composite faults in bearings with zero samples includes the following steps: S1. Multimodal data acquisition and preliminary identification of abnormal data: Collect multi-source data of bearings and upload the abnormal data after initial screening. Convert the one-dimensional vibration data in the abnormal data into a two-dimensional time-frequency diagram through continuous wavelet transform. S2. Fine-grained feature extraction: Construct a fine-grained feature extractor FGFE, which extracts multi-scale, fine-grained fault features through wavelet packet convolution WPC layer and two pyramid convolution PyConv layers; S3. Distinguishing between single and compound faults: The improved Support Vector Data Description (MSVDD) model combined with the K-Nearest Neighbor (KNN) algorithm is used to perform secondary discrimination on misclassified samples. Based on the distance between the sample and the center of the hypersphere of each type of single fault, the fault is determined to be a single fault or a compound fault. S4. Adaptive Feature Knowledge Base Construction: Preprocess and extract fault features from known fault data, generate standard symmetric principal feature point attribute map SPFDA through symmetric point pattern SDP transformation, construct an adaptive feature knowledge base and complete spatial mapping and tensor quantization processing, and define attribute prototype vectors for each fault category. S5. Zero-shot model construction and attribute prediction: Based on known fault data, construct an attribute topology network, build a multi-label graph convolutional network (MultiLabel-GCN) model, embed attributes and concatenate features of abnormal fault features, and output predicted attribute probability vectors. S6. Similarity Matching and Fault Category Output: Perform cosine similarity matching between the predicted attribute probability vector and the attribute prototype vector of the unseen fault category, and combine the single fault or compound fault determination results to output the final fault category.
[0006] Furthermore, in step S1, the multi-source data of the bearing includes bearing vibration data, image data, acoustic emission data, pressure data, rotational speed data, current data, and temperature data. The initial screening calculates the total vibration value of the bearing using the NI cRIO embedded controller. When the total vibration value exceeds a set threshold, it is determined to be abnormal data and uploaded.
[0007] Furthermore, in step S2, FGFE includes a wavelet packet convolutional layer and two PyConv layers, and the output of the WPC layer is:
[0008] Where Z is the input feature map, For convolution operations with different kernel sizes, Concat is a feature concatenation operation along the channel dimension; The PyConv layer feature fusion output is:
[0009] in, For the first Feature maps output by each convolutional branch It is the transpose of the feature map height. This is the transpose of the feature map width. This represents the total number of channels before merging.
[0010] The loss function for FGFE training cross-entropy is:
[0011] in, For real labels, For predicting labels, N is the number of single fault types, and V is the total number of samples participating in training.
[0012] Furthermore, in step S3, the hypersphere radius of the MSVDD model is:
[0013] Where R is the radius of the optimal hypersphere. For the first in the dataset One sample, For the first in the dataset One sample, For the first Sample Lagrange multipliers, For the first Sample Lagrange multipliers, Let be any support vector on the boundary of the hypersphere; The fault detection formula for the MSVDD model is:
[0014] in, To determine the distance from the test sample to the center of the hypersphere of the k-th type of fault, R k Let M be the radius of the k-th type of hypersphere, and M be the total number of known single fault categories; The MSVDD model calculates the Euclidean distance between a misclassified sample and its K nearest neighbors in the feature space, retrieves the K nearest neighbor samples, counts the fault category labels corresponding to the nearest neighbor samples, and assigns the category with the highest frequency to the misclassified sample through a majority voting mechanism.
[0015] Furthermore, in step S4, preprocessing and fault feature extraction include: X1, dividing each group of known fault data of length n into a samples, and introducing samples of length n... The sliding window performs overlapping sampling of known fault data, expanding the effective data length of each group of known fault data. This multiplies the original input features; X2. Perform global average pooling on the input features to obtain channel features, use convolution to obtain spatial features, and calculate the attention matrix:
[0016] Where S is the attention matrix, X is the Sigmoid activation function. D X is the channel feature vector. K For spatial feature vectors; X3. The attention matrix is weighted and fused with the original input features, and residual connections are introduced to generate the final output: ,
[0017] Among them, S X X is the weighted feature map, and X is the original input feature map. The weight parameters are the input feature map. represents the weight parameters of the attention matrix, and Y represents the final output.
[0018] Furthermore, in step S5, the adjacency matrix is embedded in the attribute layer of the MultiLabel-GCN model, and the asymmetric focus loss, negative probability pruning loss, and dynamic weight adjustment are introduced into the total loss function. The expression for the MultiLabel-GCN model is as follows:
[0019] in, Let I be the initial attribute embedding matrix, I be the independent initial vector for each attribute, A be the number of attributes, and q be the initial attribute embedding matrix. emb For the embedding dimension, For the first Convolution output before layer activation For the first The attribute embedding matrix of the layer, M norm For normalized attribute coupling topology, For the first Layer attribute embedding matrix, For learnable weight matrix, Let L be the dropout rate, and L be the final attribute node embedding matrix. For dimensions.
[0020] Furthermore, the adjacency matrix is constructed by training independent classifiers for each attribute in the feature knowledge base and calculating residuals, and then using a Bayesian network structure to learn and extract attribute dependencies.
[0021] Furthermore, the expression for the total loss function is:
[0022] Where e is the attribute index, B e r is the model's predicted probability for the attribute with index e. e For a true binary label, 1 indicates that the attribute exists. It is an asymmetric focal parameter, which is [value] for positive samples. When negative samples are It is a minimal constant that prevents logarithmic calculations from becoming invalid. The negative probability pruning loss is... This is a dynamic weight adjustment item. This represents the number of negative samples in the current batch. This represents the number of positive samples in the current batch. For the total loss function, It is balance The contribution weight hyperparameter, It is the basic weighting coefficient.
[0023] Furthermore, in step S6, the expression for cosine similarity is:
[0024] in, The attribute probability vector representing the prediction. C represents the prototype vector of the defined attributes of the t-th unseen class, and C represents the total number of attributes. Represents the Euclidean length of the vector; If the fault is a single fault, the MultiLabel-GCN model calculates the cosine similarity between the predicted attribute probability vector and the attribute prototype vector, and assigns the sample to the unseen fault category with the highest similarity, thus obtaining the single fault category prediction. If the fault is a composite fault, the MultiLabel-GCN model will calculate the cosine similarity between the predicted attribute probability vector and the attribute prototype vector, and combine the unseen fault categories with higher similarity to form a composite fault category prediction.
[0025] The present invention also discloses a bearing zero-sample composite fault diagnosis system. The diagnosis system adopts the above-mentioned bearing zero-sample composite fault diagnosis method and includes an intelligent perception layer, an intelligent core layer and an intelligent operation and maintenance layer. The intelligent sensing layer performs multimodal real-time acquisition and preliminary fault diagnosis of bearing data, and uploads abnormal data to the intelligent core layer. The intelligent core layer performs fault identification and diagnosis on abnormal data of the bearing. The intelligent operation and maintenance layer assesses the status of the equipment, generates operation and maintenance strategies, and manages resource scheduling based on the fault categories diagnosed by the intelligent core layer.
[0026] The present invention has the following advantages: 1. By combining multimodal perception, fine-grained feature extraction, zero-sample graph convolutional inference and composite fault classification, accurate diagnosis of rolling bearings is achieved without unknown fault samples. This effectively solves the problems of traditional methods relying on manual features, difficulty in identifying composite faults, and large interference from attribute coupling, and significantly improves the accuracy and stability of fault diagnosis.
[0027] 2. By adopting an adaptive feature knowledge base and a MultiLabel-GCN network for attribute decoupling and topology relationship learning, fault attribute modeling and unseen fault identification can be completed adaptively without the need to pre-build an artificial feature base, which greatly improves the system's generalization ability and applicability under varying operating conditions.
[0028] 3. By combining MSVDD with KNN for secondary discrimination, we can accurately distinguish between single faults and compound faults. With the optimization of the total loss function, we can alleviate the problem of attribute imbalance, reduce misjudgment caused by fault feature coupling, and significantly improve the diagnostic accuracy and system robustness of compound faults and early weak faults. Attached Figure Description
[0029] Figure 1 This is a schematic diagram of the diagnostic process of the present invention.
[0030] Figure 2 This is a schematic diagram of a multi-label graph convolutional network structure.
[0031] Figure 3 This is a schematic diagram of the diagnostic system. Detailed Implementation
[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0033] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0034] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.
[0035] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0036] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. In addition, the terms "first," "second," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.
[0037] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0038] Example 1
[0039] refer to Figure 1 As shown in Figure 2, a method for diagnosing composite faults in bearings with zero samples includes the following steps: S1. Multimodal Data Acquisition and Preliminary Anomaly Identification: The system acquires multimodal, full-dimensional signals of the rolling bearing's operating status using vibration sensors, image sensors, acoustic emission sensors, pressure sensors, speed sensors, current sensors, and temperature sensors. The acquired bearing vibration data, image data, acoustic emission data, pressure data, speed data, current data, and temperature data are then uploaded to an NI cRIO embedded controller. This controller is a baseboard integrating an embedded real-time controller and an FPGA chip, featuring expandable interfaces and suitable for multimodal sensor data acquisition and industrial control. It can perform preliminary fault diagnosis and network upload of the acquired multimodal sensor data.
[0040] The controller first calculates the total vibration value of the vibration signal. When the total vibration value exceeds the preset threshold, the current data is determined to be abnormal data and uploaded to the intelligent core layer of the system for in-depth diagnosis. To facilitate subsequent feature extraction, the one-dimensional vibration signal is mapped into a two-dimensional time-frequency image through continuous wavelet transform (CWT), so that the fault presents more obvious separability features in the time-frequency domain.
[0041] S2. Fine-grained feature extraction: In order to capture the subtle differences between weak faults and complex faults, this embodiment builds a fine-grained feature extractor FGFE to extract multi-scale, fine-grained fault features, which includes a wavelet packet convolution WPC layer and two pyramid convolution PyConv layers. The WPC layer uses convolutional kernels of different scales to extract multi-band fault information in parallel, enhancing the ability to detect early and weak faults. The output of the WPC layer is:
[0042] Where Z is the input feature map, For convolution operations with different kernel sizes, Concat is a feature concatenation operation along the channel dimension; The PyConv layer feature fusion output is:
[0043] in, For the first Feature maps output by each convolutional branch It is the transpose of the feature map height. This is the transpose of the feature map width. This represents the total number of channels before merging.
[0044] The loss function for FGFE training cross-entropy is:
[0045] in, For real labels, For predicting labels, N is the number of single fault types, and V is the total number of samples participating in training.
[0046] The above structure enables the extraction of multi-scale, fine-grained, and highly robust fault features from time-frequency images, providing a high-quality feature foundation for subsequent fault differentiation.
[0047] S3. Distinguishing between single and compound faults: An improved Support Vector Data Description (MSVDD) model combined with the K-Nearest Neighbor (KNN) algorithm is used to perform secondary discrimination on misclassified samples. Based on the distance between the sample and the center of the hypersphere of each type of single fault, the fault is determined to be either a single fault or a compound fault. Specifically... The hypersphere radius of the MSVDD model is:
[0048] Where R is the radius of the optimal hypersphere. For the first in the dataset One sample, For the first in the dataset One sample, For the first Sample Lagrange multipliers, For the first Sample Lagrange multipliers, Let be any support vector on the boundary of the hypersphere; The fault detection formula for the MSVDD model is:
[0049] in, To determine the distance from the test sample to the center of the hypersphere of the k-th type of fault, R k Let M be the radius of the k-th type of hypersphere, and M be the total number of known single fault categories. MSVDD is trained using pure single fault samples to construct a compact hypersphere for each single fault feature. The constructed hypersphere can accurately obtain the boundaries of each type of single fault, providing a reliable single fault benchmark for subsequent compound fault discrimination. If the test sample falls into any hypersphere, it is judged as a single fault; if the sample falls outside all hyperspheres, it is judged as a compound fault.
[0050] Since MSVDD is prone to misclassification near the boundary in high-dimensional feature space, this embodiment introduces the K-nearest neighbor algorithm to perform secondary discrimination on boundary samples: The MSVDD model calculates the Euclidean distance between the misclassified sample and its K nearest neighbors in the feature space, retrieves the K nearest neighbors, counts the fault category labels corresponding to the nearest neighbors, and assigns the category with the highest frequency to the misclassified sample through a majority voting mechanism, thereby correcting the classification result and improving the accuracy of distinguishing between single faults or compound faults.
[0051] S4. Adaptive Feature Knowledge Base Construction: To break away from the reliance on human experience in traditional zero-sample diagnosis, this embodiment adopts an adaptive approach to construct a fault feature knowledge base.
[0052] First, the known fault data is preprocessed and fault features are extracted, which specifically includes: X1. Divide each group of known fault data of length n into a samples, and introduce a sample of length n. The sliding window performs overlapping sampling of known fault data, expanding the effective data length of each group of known fault data. This multiplies the original input features, expands the effective sample length, and improves data utilization. X2. Perform global average pooling on the input features to obtain channel features, use convolution to obtain spatial features, and calculate the attention matrix:
[0053] Where S is the attention matrix, X is the Sigmoid activation function. D X is the channel feature vector. K For spatial feature vectors; X3. The attention matrix is weighted and fused with the original input features, and residual connections are introduced to generate the final output:
[0054] Among them, S X X is the weighted feature map, and X is the original input feature map. The weight parameters are the input feature map. represents the weight parameters of the attention matrix, and Y represents the final output (extracted features).
[0055] The extracted features are transformed into a standard symmetric principal feature point attribute map (SPFDA) by the Symmetric Point Pattern (SDP) transformation. All standard attributes of all fault categories are integrated to form an adaptive feature knowledge base of global static prior attributes. Spatial mapping and tensor quantization are then performed to define and generate attribute prototype vectors for each fault category, which serve as semantic anchors for zero-shot inference.
[0056] S5. Zero-Shot Model Construction and Attribute Prediction: To address the problem of mixed coupling between fault attributes, this embodiment constructs a MultiLabel-GCN model. First, an attribute topology network is built based on known fault data to construct the MultiLabel-GCN model. Compared to the traditional GCN algorithm, it introduces an adjacency matrix M representing attribute dependencies, eliminating the confusion between attributes and enabling the extraction of direct attribute coupling. Simultaneously, through a loss function, asymmetric focus loss, negative probability pruning, and dynamic weight adjustment are further integrated to effectively alleviate the severe attribute imbalance problem. The structure is as follows: Figure 2 As shown, the adjacency matrix M is embedded into the attribute L layer by the GCN, and then concatenated with the features of the samples and input into the classification head to predict the attributes.
[0057] The expression for the MultiLabel-GCN model is as follows:
[0058] in, Let I be the initial attribute embedding matrix, I be the independent initial vector for each attribute, A be the number of attributes, and q be the initial attribute embedding matrix. emb For the embedding dimension, For the first Convolution output before layer activation For the first The attribute embedding matrix of the layer, M norm For normalized attribute coupling topology, For the first Layer attribute embedding matrix, For learnable weight matrix, Let L be the dropout rate, and L be the final attribute node embedding matrix. The adjacency matrix M is defined by the dimension of the property knowledge base. During the training phase, independent predictors are trained for each attribute in the property knowledge base, single-attribute residuals are calculated, and large direct dependencies between attributes are extracted through a Bayesian network.
[0059] During the testing and diagnostic phase, the unknown fault abnormal multimodal signal x n Construct the main feature matrix Meanwhile, an independent predictor is trained for each attribute, the single-attribute residuals are calculated to form an error matrix, the consistency and dependency of attributes are analyzed, and after decoupling, the attribute topology network is extracted.
[0060] For the attribute embedding part, each standard feature in the feature knowledge base is defined as a category semantic attribute, which serves as the comparison anchor for subsequent zero-shot inference and is used to initialize the initial attribute embedding matrix of the convolutional network. Meanwhile, GCN performs multi-level neighbor information aggregation on attribute nodes through two layers of graph convolution, with each layer followed by the ReLU loss function and random deactivation Dropout. The first layer of DenseGCN takes the initial attribute embedding matrix as input. Output attribute node embedding matrix The second layer DenseGCN input attribute node embedding matrix Output attribute node embedding matrix Ultimately, a vector is embedded for each attribute structure, effectively integrating the coupling information between attributes. The feature concatenation part is implemented by the classification head module, which combines the main feature matrix... After expansion, the attribute structure embedding vector obtained from GCN is concatenated with the concatenated tensor along the feature dimension to form a concatenated tensor. This tensor is used for classification between different fault types in the attribute embedding part. After propagation through the total loss function in the classification of different fault types, the concatenated tensor is used by the classification head of each attribute to predict the attribute probability vector.
[0061] The expression for the total loss function is:
[0062] Where e is the attribute index, B e r is the model's predicted probability for the attribute with index e. e For a true binary label, 1 indicates that the attribute exists. It is an asymmetric focal parameter, which is [value] for positive samples. When negative samples are It is a minimal constant that prevents logarithmic calculations from becoming invalid. The negative probability pruning loss is... This is a dynamic weight adjustment item. This represents the number of negative samples in the current batch. This represents the number of positive samples in the current batch. For the total loss function, It is balance The contribution weight hyperparameter, It is the basic weighting coefficient.
[0063] S6. Similarity Matching and Fault Category Output: Perform cosine similarity matching between the predicted attribute probability vector and the attribute prototype vector of the unseen fault category, and combine the single fault or compound fault determination results to output the final fault category.
[0064] The expression for cosine similarity is:
[0065] in, The attribute probability vector representing the prediction. C represents the prototype vector of the defined attributes of the t-th unseen class, and C represents the total number of attributes. Represents the Euclidean length of the vector; If the fault is a single fault, the MultiLabel-GCN model calculates the cosine similarity between the predicted attribute probability vector and the attribute prototype vector, and assigns the sample to the unseen fault category with the highest similarity, thus obtaining the single fault category prediction. If the fault is a composite fault, the MultiLabel-GCN model will calculate the cosine similarity between the predicted attribute probability vector and the attribute prototype vector, and combine the unseen fault categories with higher similarity to form a composite fault category prediction.
[0066] Compared to traditional zero-shot diagnostic methods, the prediction results can adaptively generate a feature knowledge base, no longer relying on experience settings. At the same time, through the MSVDD strategy, it can separate the diagnosis of single faults and compound faults, which has a wider range of application scenarios than methods that can only handle one type of fault.
[0067] Finally, the MultiLabel-GCN model eliminates the obfuscation effect of input data and achieves direct attribute coupling extraction. Based on the prediction results, the condition of the rolling bearing is diagnosed and intelligent strategies are provided. Simultaneously, related equipment and spare parts inventory are adjusted to achieve dynamic intelligent management of maintenance and replacement. Dynamic optimization of personnel and equipment scheduling and allocation ensures production efficiency and avoids safety issues.
[0068] By employing the two core algorithms described above to achieve phased collaboration, accurate prediction of new and compound faults can be achieved without any unknown fault training samples. After abnormal vibration data is input, fine-grained features are first extracted via FGFE, and then the data enters the MSVDD module for the first stage of fault type discrimination. Simultaneously, the output of the MSVDD module provides crucial information for the MultiLabel-GCN inference strategy: when a sample is determined to be a single fault, it directly proceeds to the subsequent zero-sample classification process; when determined to be a compound fault, the compound fault handling mechanism is triggered. In this application, the prediction of compound faults is not a simple repetition of the single-fault process. Instead, the predicted attribute probability vector output by GCN is used to calculate the cosine similarity with the attribute prototypes of unseen categories. Multiple fault categories with the highest similarity are then combined for inference, forming the final diagnosis of the compound fault. This divide-and-conquer architecture, which decouples joint attributes, enables the system to first decompose complex composite fault problems into the identification of multiple single fault components. Then, through the confusion effect between attributes decoupled by GCN, it can accurately identify multiple fault modes that coexist in composite faults, thereby improving the ability to diagnose composite faults from single faults.
[0069] Example 2
[0070] like Figure 1-3 As shown, this embodiment provides a bearing zero-sample composite fault diagnosis system, which uses the diagnosis method of Embodiment 1 to diagnose bearing zero-sample composite faults. The diagnosis system is deployed on the NI cRIO embedded controller platform and specifically includes three parts: an intelligent sensing layer, an intelligent core layer, and an intelligent operation and maintenance layer, realizing the integration of data acquisition, fault diagnosis, and intelligent operation and maintenance. The intelligent sensing layer consists of a multimodal sensing acquisition unit and an NI cRIO embedded controller. The multimodal sensing acquisition unit acquires vibration, image, acoustic emission, pressure, speed, current, and temperature signals of the rolling bearing in real time. The NI cRIO embedded controller performs preliminary preprocessing on the above multi-source data, completes the abnormal data screening by vibration threshold judgment, and uploads the abnormal data to the intelligent core layer to provide reliable data support for subsequent in-depth diagnosis.
[0071] The intelligent core layer serves as the core of the system's diagnostics, deploying an FGFE fine-grained feature extractor, an MSVDD fault differentiation module, and a MultiLabel-GCN zero-shot diagnostic module. FGFE extracts fine-grained fault features; MSVDD and KNN are then used for joint discrimination to classify faults into single or compound faults; finally, MultiLabel-GCN performs attribute decoupling, attribute prediction, and cosine similarity matching to output the final fault category, achieving accurate diagnosis of single and compound faults under zero-shot conditions.
[0072] The intelligent operation and maintenance layer includes an assessment unit, an intelligent strategy unit, and a scheduling and management unit. Based on the fault category results diagnosed by the intelligent core layer, the assessment unit quantitatively assesses the health status of the rolling bearing and the entire equipment. The intelligent strategy unit automatically generates predictive maintenance, repair, and replacement intelligent strategies based on the assessment results. The scheduling and management unit performs dynamic optimization scheduling of personnel, spare parts, and production resources based on the intelligent strategies, realizing intelligent operation and maintenance throughout the entire equipment lifecycle and improving equipment reliability and production efficiency.
[0073] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for diagnosing composite faults in bearings with zero samples, characterized in that: Includes the following steps: S1. Multimodal data acquisition and preliminary identification of abnormal data: Collect multi-source data of bearings and upload the abnormal data after initial screening. Convert the one-dimensional vibration data in the abnormal data into a two-dimensional time-frequency diagram through continuous wavelet transform. S2. Fine-grained feature extraction: Construct a fine-grained feature extractor FGFE, which extracts multi-scale, fine-grained fault features through wavelet packet convolution WPC layer and two pyramid convolution PyConv layers; S3. Distinguishing between single and compound faults: The MSVDD model combined with the KNN algorithm is used to perform secondary discrimination on misclassified samples. Based on the distance between the sample and the center of the hypersphere of each type of single fault, the fault is determined to be a single fault or a compound fault. S4. Adaptive Feature Knowledge Base Construction: Preprocess and extract fault features from known fault data, generate standard symmetric principal feature point attribute map SPFDA through symmetric point pattern SDP transformation, construct an adaptive feature knowledge base and complete spatial mapping and tensor quantization processing, and define attribute prototype vectors for each fault category. S5. Zero-shot model construction and attribute prediction: Based on known fault data, construct an attribute topology network, build a MultiLabel-GCN model, embed and concatenate attributes of abnormal fault features, and output the predicted attribute probability vector. S6. Similarity Matching and Fault Category Output: Perform cosine similarity matching between the predicted attribute probability vector and the attribute prototype vector of the unseen fault category, and combine the single fault or compound fault determination results to output the final fault category.
2. The bearing zero-sample composite fault diagnosis method according to claim 1, characterized in that: In step S1, the multi-source data of the bearing includes bearing vibration data, image data, acoustic emission data, pressure data, rotational speed data, current data, and temperature data. The initial screening calculates the total vibration value of the bearing using an NI cRIO embedded controller. When the total vibration value exceeds a set threshold, it is determined to be abnormal data and uploaded.
3. The bearing zero-sample composite fault diagnosis method according to claim 1, characterized in that: In step S2, FGFE includes a WPC layer and two PyConv layers. The output of the WPC layer is: Where Z is the input feature map, For convolution operations with different kernel sizes, Concat is a feature concatenation operation along the channel dimension; The PyConv layer feature fusion output is: in, For the first Feature maps output by each convolutional branch It is the transpose of the feature map height. This is the transpose of the feature map width. The total number of channels before fusion; the loss function for FGFE training cross-entropy is: in, For real labels, For predicting labels, N is the number of single fault types, and V is the total number of samples participating in training.
4. The bearing zero-sample composite fault diagnosis method according to claim 1, characterized in that: In step S3, the hypersphere radius of the MSVDD model is: Where R is the radius of the optimal hypersphere. For the first in the dataset One sample, For the first in the dataset One sample, For the first Sample Lagrange multipliers, For the first Sample Lagrange multipliers, Let be any support vector on the boundary of the hypersphere; The fault detection formula for the MSVDD model is: in, To determine the distance from the test sample to the center of the hypersphere of the k-th type of fault, R k Let M be the radius of the k-th type of hypersphere, and M be the total number of known single fault categories; The MSVDD model calculates the Euclidean distance between a misclassified sample and its K nearest neighbors in the feature space, retrieves the K nearest neighbor samples, counts the fault category labels corresponding to the nearest neighbor samples, and assigns the category with the highest frequency to the misclassified sample through a majority voting mechanism.
5. The bearing zero-sample composite fault diagnosis method according to claim 1, characterized in that: In step S4, preprocessing and fault feature extraction include: X1. Divide each group of known fault data of length n into a samples, and introduce a sample of length n. The sliding window performs overlapping sampling of known fault data, expanding the effective data length of each group of known fault data. This multiplies the original input features; X2. Perform global average pooling on the input features to obtain channel features, use convolution to obtain spatial features, and calculate the attention matrix: Where S is the attention matrix, X is the Sigmoid activation function. D X is the channel feature vector. K For spatial feature vectors; X3. The attention matrix is weighted and fused with the original input features, and residual connections are introduced to generate the final output: , Among them, S X X is the weighted feature map, and X is the original input feature map. The weight parameters are the input feature map. represents the weight parameters of the attention matrix, and Y represents the final output.
6. The bearing zero-sample composite fault diagnosis method according to claim 5, characterized in that: In step S5, the adjacency matrix is embedded in the attribute layer of the MultiLabel-GCN model, and the total loss function introduces asymmetric focus loss, negative probability pruning loss, and dynamic weight adjustment. The expression for the MultiLabel-GCN model is as follows: in, Let I be the initial attribute embedding matrix, I be the independent initial vector for each attribute, A be the number of attributes, and q be the initial attribute embedding matrix. emb For the embedding dimension, For the first Convolution output before layer activation For the first The attribute embedding matrix of the layer, M norm For normalized attribute coupling topology, For the first Layer attribute embedding matrix, For learnable weight matrix, Let L be the dropout rate, and L be the final attribute node embedding matrix. For dimensions.
7. The bearing zero-sample composite fault diagnosis method according to claim 6, characterized in that: The adjacency matrix is constructed by training independent classifiers for each attribute in the feature knowledge base, calculating residuals, and then using a Bayesian network structure to learn and extract attribute dependencies.
8. The bearing zero-sample composite fault diagnosis method according to claim 6, characterized in that: The expression for the total loss function is: Where e is the attribute index, B e r is the model's predicted probability for the attribute with index e. e For a true binary label, 1 indicates that the attribute exists. It is an asymmetric focal parameter, which is [value] for positive samples. When negative samples are It is a minimal constant that prevents logarithmic calculations from becoming invalid. The negative probability pruning loss is... This is a dynamic weight adjustment item. This represents the number of negative samples in the current batch. This represents the number of positive samples in the current batch. For the total loss function, It is balance The contribution weight hyperparameter, It is the basic weighting coefficient.
9. The bearing zero-sample composite fault diagnosis method according to claim 1, characterized in that: In step S6, the expression for cosine similarity is: in, The attribute probability vector representing the prediction. C represents the prototype vector of the defined attributes of the t-th unseen class, and C represents the total number of attributes. Represents the Euclidean length of the vector; If the fault is a single fault, the MultiLabel-GCN model calculates the cosine similarity between the predicted attribute probability vector and the attribute prototype vector, and assigns the sample to the unseen fault category with the highest similarity, thus obtaining the single fault category prediction. If the fault is a composite fault, the MultiLabel-GCN model will calculate the cosine similarity between the predicted attribute probability vector and the attribute prototype vector, and combine the unseen fault categories with higher similarity to form a composite fault category prediction.
10. A bearing zero-sample composite fault diagnosis system, characterized in that: The diagnostic system adopts a bearing zero-sample composite fault diagnosis method as described in any one of claims 1-9. The diagnostic system includes an intelligent perception layer, an intelligent core layer, and an intelligent operation and maintenance layer. The intelligent sensing layer performs multimodal real-time acquisition and preliminary fault diagnosis of bearing data, and uploads abnormal data to the intelligent core layer. The intelligent core layer performs fault identification and diagnosis on abnormal data of the bearing. The intelligent operation and maintenance layer assesses the status of the equipment, generates operation and maintenance strategies, and manages resource scheduling based on the fault categories diagnosed by the intelligent core layer.