Bearing fault diagnosis method based on multi-scale convolutional neural network and knowledge graph

Abstract

The invention discloses a bearing fault diagnosis method based on a multi-scale convolutional neural network and a knowledge graph, and belongs to the technical field of mechanical fault monitoring and artificial intelligence. The method comprises the following steps: firstly, acquiring an original bearing vibration signal and mechanism knowledge, and performing entity extraction and label standardization; inputting the vibration signal into a multi-scale residual attention convolution network for end-to-end feature extraction and fault recognition, wherein the network performs cooperative processing through a residual convolution module, a multi-scale depth separable Inception module and a time sequence channel attention module in sequence; and finally, constructing a hierarchical bearing fault knowledge graph based on an identification result and mechanism knowledge, and storing the hierarchical bearing fault knowledge graph into a graph database. According to the method, fault features can be automatically and comprehensively extracted, the diagnosis precision and cross-working-condition robustness are effectively improved, and interpretability and decision support of the diagnosis process are realized through deep fusion of the knowledge graph.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mechanical fault monitoring and artificial intelligence, and particularly relates to a bearing fault diagnosis method based on a multi-scale convolutional neural network and a knowledge graph. BACKGROUND

[0002] Rolling bearings are key transmission components in industrial rotating machinery, and their running state directly affects the overall reliability of the equipment. For bearing fault diagnosis, existing technologies mainly rely on data-driven methods. On the one hand, traditional machine learning-based methods, such as random forests, support vector machines, and Bayesian networks, extract time-domain, frequency-domain, or time-frequency domain features from vibration signals through manual extraction, and combine classifiers to achieve fault recognition. Such methods have certain diagnostic capabilities under specific working conditions. On the other hand, with the development of deep learning technology, convolutional neural networks and other models have been introduced into the field of fault diagnosis, which can automatically learn features from raw vibration data or transformed time-frequency images, reducing the dependence on manual feature engineering and improving the generalization and diagnostic efficiency of feature representation. In addition, as a structured knowledge representation tool, knowledge graphs have begun to be explored in fault diagnosis to integrate mechanism knowledge, working condition information, and diagnostic results to improve the explainability and decision support capability of the diagnostic process.

[0003] However, the existing technology still has the following shortcomings: First, traditional machine learning methods highly depend on expert experience for manual feature design and selection, which is tedious and prone to subjective bias, resulting in unstable feature quality and limited generalization ability in varying working conditions. Second, existing deep learning models are mostly based on single-scale convolution kernels or standard network structures, which are insufficient in capturing multi-scale impact and modulation features in bearing vibration signals. At the same time, as the network depth increases, it is prone to gradient vanishing or model degradation problems, affecting the extraction of deep features and diagnostic accuracy. In addition, existing methods mostly use vibration data for end-to-end classification, lacking structured integration of domain mechanism knowledge and diagnostic results, resulting in poor explainability of the diagnostic process and difficulty in tracing fault causes and assisting decision-making. Finally, the application of knowledge graphs in fault diagnosis is still in its early stages, facing challenges such as complex domain knowledge extraction, reliance on a large amount of manual entity and relationship definition, and difficulty in integrating with diagnostic models, which limits its effective deployment in practical engineering. SUMMARY

[0004] To solve the above technical problems, the application provides a bearing fault diagnosis method based on a multi-scale convolutional neural network and a knowledge graph to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above-mentioned purpose, the application provides a bearing fault diagnosis method based on a multi-scale convolutional neural network and a knowledge graph, comprising: S1, obtaining original bearing vibration signals and corresponding mechanism knowledge, performing entity extraction and label standardization based on the mechanism knowledge, the entity including equipment, working condition and fault position; S2, inputting the original bearing vibration signals into a multi-scale residual attention convolutional network for end-to-end feature extraction and fault recognition to obtain a multi-level fault classification result; wherein the processing procedure of the multi-scale residual attention convolutional network includes: first performing bottom-layer time-series feature extraction through a residual convolution module, then extracting multi-scale features in parallel through a multi-scale deep separable Inception module, and finally weighting and strengthening the features through a time-series channel attention module; S3, constructing a bearing fault knowledge graph based on the multi-level fault classification result and the mechanism knowledge, and writing the diagnosis result, working condition and mechanism event into a graph database to form a traceable knowledge graph.

[0006] Preferably, the original bearing vibration signals obtained in step S1 contain environmental noise, and the signal-to-noise ratio calculation formula is: ; wherein, is the signal power, is the noise power.

[0007] Preferably, in step S1, the label standardization specifically includes: dividing the fault information into a multi-level label system from coarse to fine according to the sampling frequency, load and rotating speed, defect size, defect position and outer ring orientation.

[0008] Preferably, the residual convolution module includes a first convolution layer, a batch normalization layer, an activation function layer, a second convolution layer, a batch normalization layer and a residual connection path connected in sequence; when the channel number or the stride of the input and the output is inconsistent, a one-dimensional convolution bypass is used to realize residual alignment.

[0009] Preferably, the multi-scale deep separable Inception module includes a plurality of parallel branches, each branch including a one-dimensional convolution layer for channel dimension reduction and a deep separable convolution layer with different receptive fields in sequence; the outputs of each branch are spliced in the channel dimension.

[0010] Preferably, the receptive field size of the deep separable convolution layer is selected from 3, 5 and 7.

[0011] Preferably, the step of weighting the input features by the time-series channel attention module includes: performing time-series global pooling on the input features to obtain time-series weights, and using the time-series weights to perform the first weighting on the input features; The first weighted feature is compressed and excited in the channel dimension to obtain channel weights, and the channel weights are used for second weighting of the feature.

[0012] Preferably, in step S2, during training of the multi-scale residual attention convolutional network, a weighted F1 value is used as a comprehensive evaluation index of model performance.

[0013] Preferably, in step S3, the nodes of the bearing fault knowledge graph are hierarchically organized according to the multi-level fault classification results obtained in step S2, forming a multi-layer semantic network containing device level, working condition level, size level, part level and orientation level.

[0014] Preferably, in step S3, the relationships in the knowledge graph are extracted based on the mechanism knowledge and the multi-level fault classification results, and are stored in the graph database in the form of triples of head entity, relationship and tail entity.

[0015] Compared with the prior art, the present application has the following advantages and technical effects: The present application constructs a cascaded processing flow of "residual convolution module → multi-scale deep separable Inception module → time sequence channel attention module", first uses the residual convolution module to stably extract the basic time sequence features and ensure the training stability, then captures the impact and modulation mode of different frequency bands in the vibration signal through the multi-scale parallel branch, and overcomes the receptive field limitation of single scale convolution, and finally uses the time sequence channel attention module designed for one-dimensional signal to adaptively strengthen the key time segments and feature channels. This collaborative design enables the model to automatically and fully extract multi-level and discriminative features from the original vibration signal, effectively improving the diagnostic accuracy and generalization ability under variable working conditions and noise interference.

[0016] The present application deeply integrates the multi-level classification results output by the deep learning model with the domain mechanism knowledge, not only uses the knowledge graph to structure the diagnostic results and trace back, but also realizes the fine-grained semantic association and query from fault type to specific working condition, defect part, size and orientation through the entity hierarchical network from coarse to fine. This provides an effective way to solve the problem of lack of explainability of existing data-driven methods, and provides rich context information support for auxiliary decision-making.

[0017] The present application directly uses the multi-level fault recognition results of the multi-scale residual attention convolutional network model as the input source for constructing and updating the knowledge graph. This design enables the entities and relationships of the knowledge graph to evolve dynamically during the model diagnosis process, forming a closed-loop system of "data-driven diagnosis → result structured storage → knowledge optimization". This surpasses the simple result post-processing mode, and realizes the deep coupling of data representation ability and knowledge reasoning ability.

[0018] According to the characteristics of the bearing fault signal, the time sequence channel attention module is used to realize the serial and targeted attention to the time sequence and channel dimension. Compared with the general attention mechanism, the customized design can more accurately capture the key local impact related to the fault in the one-dimensional vibration signal and the channel response, thereby showing better feature selection and anti-interference ability in a strong noise environment. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings, which form a part of the present application, are intended to provide further understanding of the present application, and the illustrative embodiments of the present application and their description serve the purpose of explaining the present application. The accompanying drawings should not be construed as an inappropriate limitation on the present application. In the drawings: Figure 1 A schematic diagram of a residual backbone module A of an embodiment of the present application; Figure 2 A schematic diagram of an Inception module B of an embodiment of the present application; Figure 3 A schematic diagram of an Inception module C of an embodiment of the present application; Figure 4 A schematic diagram of an attention mechanism module D of an embodiment of the present application; Figure 5 A schematic diagram of a fault diagnosis overall framework and flowchart of an embodiment of the present application; Figure 6 A schematic diagram of an accuracy rate curve of a fifth level classification experiment of different algorithms of an embodiment of the present application; Figure 7 A CWRU knowledge graph of an embodiment of the present application; Figure 8 A model input layer visualization result graph of an embodiment of the present application; Figure 9 A model attention module layer visualization result graph of an embodiment of the present application; Figure 10 A fault diagnosis graph of an embodiment of the present application under different algorithm variable load conditions; Figure 11 A comparison graph of accuracy rate curves of different algorithms of an embodiment of the present application. DETAILED DESCRIPTION

[0020] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0021] It is noted that the steps shown in the flowcharts of the drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowcharts, in some cases the steps shown or described can be executed in an order different from that shown here.

[0022] The following first describes the technical terms involved in the following embodiments.

[0023] A convolutional neural network (CNN) is a kind of multi-layer feedforward model that learns parameters by means of error backpropagation and is composed of convolution, pooling and full connection modules. The convolution layer is responsible for automatically extracting multi-scale features from the original data; the pooling layer suppresses overfitting while compressing features and reducing computational complexity; and the full connection layer completes the mapping of features to categories. The network performs convolution operation on the input tensor in the convolution unit with a local receptive field, realizing the step-by-step representation of semantics from low to high levels.

[0024] A knowledge graph organizes domain knowledge in a structured way, covers entities, attributes and their mutual relationships, and is a semantic network composed of nodes and edges, which can be represented in the form of triples (E1, R, E2): where E1 is the head entity, R is the relationship, and E2 is the tail entity. In the fault diagnosis scenario, key entities (such as components, fault types, working conditions, etc.) can be extracted from mechanism information, and the relationships between entities can be inferred according to a classification model or rules, so as to construct a bearing fault diagnosis knowledge graph for knowledge organization, result explanation and decision support.

[0025] Embodiment one The embodiment provides a bearing fault diagnosis method based on a multi-scale convolutional neural network and a knowledge graph, comprising: S1, acquiring original bearing vibration signals and corresponding mechanism knowledge, performing entity extraction and label normalization based on the mechanism knowledge, the entities including equipment, working conditions and fault positions; Further, in step S1, the label normalization is specifically: according to the sampling frequency, load and rotating speed, defect size, defect position and outer ring orientation, the fault information is divided into a multi-level label system from coarse to fine.

[0026] S2, input the original bearing vibration signal into a multi-scale residual attention convolutional neural network (MRTCNN) for end-to-end feature extraction and fault identification, and obtain a multi-level fault classification result; wherein, the processing procedure of the MRTCNN comprises: first, performing bottom-layer time sequence feature extraction through a residual convolution module, then performing parallel multi-scale feature extraction through a multi-scale deep separable Inception module, and finally performing weighting and strengthening on the features through a temporal-channel attention module (TCAM) for 1-D. Further, the residual convolution module comprises a first convolution layer, a batch normalization layer, an activation function layer, a second convolution layer, a batch normalization layer and a residual connection path connected in sequence; when the channel number or the step size of the input and the output is inconsistent, a one-dimensional convolution bypass is used to realize residual alignment.

[0027] Specifically, to improve the network depth and feature expression ability without significantly increasing the parameters, the MRTCNN uses a residual convolution block with small convolution kernel stacking as the front-end backbone. Each block is composed of two layers of (3x1) one-dimensional convolution and batch normalization (BN) layers, and is equipped with ReLU activation; when the channel number or the step size changes, a (1x1) bypass is used to realize residual alignment. This design not only inherits the advantage of "small convolution kernel multi-layer stacking is superior to large convolution kernel", but also stabilizes deep training with the help of residual path, alleviates gradient vanishing and accelerates convergence. Unlike directly applying VGG, the residual connection and small kernel convolution are tightly coupled in this embodiment, so that the first two layers complete the transition from the original vibration to the middle-level semantics while keeping the parameters friendly. The residual backbone module A is as shown in Figure 1 .

[0028] Further, the multi-scale deep separable Inception module comprises a plurality of parallel branches, each branch comprising a one-dimensional convolution layer for channel dimension reduction and a deep separable convolution layer with different receptive fields in sequence; the outputs of each branch are spliced in the channel dimension.

[0029] Further, the receptive field size of the deep separable convolution layer is selected from 3, 5 and 7.

[0030] Specifically, the traditional single-scale convolution has limited adaptability to bearing vibration under varying conditions and frequency bands. To this end, MRTCNN introduces a multi-scale deep separable Inception structure in the middle section: parallel branches use deep separable convolutions with different receptive fields such as (3, 5, 7) to cover both narrow-band impact and wide-band modulation patterns at low cost; each branch first performs channel dimension reduction with (1x1) convolution, then performs deep convolution of the corresponding scale, and finally concatenates in the channel dimension. Inception module B and Inception module C, as shown in Figures 2-3 Compared with the traditional Inception, this module significantly reduces complexity with deep separable convolution while enhancing feature diversity; the number of channels is evenly distributed among branches to avoid dominance by a certain path, thereby improving robustness to different fault frequency bands and speed disturbances.

[0031] Further, the step of weighting the input features by the time-series channel attention module TCAM includes: temporally globally pooling the input features to obtain a time-series weight, and weighting the input features for the first time using the time-series weight; compressing and exciting the features after the first weighting in the channel dimension to obtain a channel weight, and weighting the features for the second time using the channel weight.

[0032] Specifically, to highlight key pulses and their channel responses, MRTCNN concatenates TCAM (Temporal-Channel Attention for 1-D) after high-level features. Attention mechanism module D, as shown in Figure 4 TCAM first obtains a time weight through temporal global pooling to suppress smooth redundancy; then it performs compression-excitation in the channel dimension to obtain a channel weight, emphasizing filter responses that are strongly related to fault patterns. The two weights act on the feature tensor in turn, achieving "temporal first, then channel" step-by-step emphasis. Unlike the two-dimensional spatial attention of CBAM, TCAM is directed at one-dimensional vibration scenarios, directly modeling the time dimension and channel dimension, and is more suitable for expressing impact-type faults. To stabilize training, the module is equipped with BN / Dropout and residual projection, allowing attention enhancement to naturally integrate with backbone representation.

[0033] Further, in step S2, when training the MRTCNN, the weighted F1 value is used as a comprehensive evaluation indicator of model performance.

[0034] Specifically, the overall framework and process of fault diagnosis are as follows: Figure 5The method takes the original bearing vibration sequence as input without time-frequency transformation or manual features. First, the entity and attribute extraction (components, working conditions, fault types, etc.) is completed through rules and mechanism entries, and the vibration segment is sent to MRTCNN in parallel: module A captures low-level time edges and short cycles, module B and module C gather multi-scale impact and bandwidth information, and module D uses TCAM to weight and strengthen key time slices and channels. Then, the global average pooling and classification head are used to get the fault category, and the regularization methods such as BN and Dropout are used to alleviate overfitting in the training stage, and the optimal model is selected according to the accuracy and weighted F1 indicators. When the classification result is stable, the triplets are constructed according to the mechanism relationship and model output to form a fault diagnosis knowledge graph, realizing information integration, result explanation and auxiliary decision-making.

[0035] Considering the class imbalance of the data, accuracy alone is not enough to fully evaluate performance, and because precision and recall have advantages and disadvantages, the weighted F1 is used as a comprehensive indicator.

[0036] S3, based on the multi-level fault classification result and the mechanism knowledge, a bearing fault knowledge graph is constructed, and the diagnosis result, working condition and mechanism event are written into the graph database to form a traceable knowledge graph.

[0037] Further, in step S3, the nodes of the bearing fault knowledge graph are hierarchically organized according to the multi-level fault classification result obtained in step S2 to form a multi-level semantic network including device level, working condition level, size level, part level and orientation level.

[0038] Further, in step S3, the relationships in the knowledge graph are extracted based on the mechanism knowledge and the multi-level fault classification result, and are stored in the graph database in the form of triplets of head entity, relationship and tail entity.

[0039] The embodiment carries out experiments on the above technical solutions, which are divided into two parts: one is the construction of the fault knowledge graph, which is stored in the graph database using Neo4j, and the programmed writing of entities and relationships is completed through py3neo; the second is the bearing fault classification, which takes the MRTCNN (Multi-scale Residual TCAM 1-DCNN, MRTCNN) proposed in the embodiment as the core recognizer, and trains and evaluates the CWRU dataset, and discusses the model performance combined with the separability measure and visualization.

[0040] The overall flow of the experiment is: based on the mechanism information, extract entities and attributes, complete the rule-based annotation; the original vibration fragment is directly input into MRTCNN, without time-frequency transformation or manual features; after obtaining stable classification results, the "sample-working condition-defect" triplets are mapped to the graph database to form a queryable and interpretable knowledge base. The experiment specifically includes: 1. Entity extraction; The experiment uses the bearing data publicly released by Case Western Reserve University (CWRU). This dataset also records key mechanism factors, including motor load and speed, defect location and size, sampling location and frequency, etc., which can be used as the basis for entity and relationship extraction. Combined with the data description, this embodiment converts the rule-based knowledge into a unified set of entities and attributes: (1) Sampling location and frequency: both the drive end and the fan end are sampled, with the drive end including 12 kHz and 48 kHz versions, and the fan end commonly having 12 kHz; normal samples and various fault samples cover the above settings.

[0041] (2) Load and speed: the test is conducted under 0 / 1 / 2 / 3 HP four-grade loads, corresponding to speeds of about 1797, 1772, 1750, and 1730 r / min; this information is used to distinguish the same type of faults under different working conditions.

[0042] (3) Defect size: common defect sizes are 0.007, 0.014, and 0.021 inches, with a small amount of 0.028 inches, used to refine the severity of defects at the same location.

[0043] (4) Defect location: covering rolling elements, inner rings, and outer rings; the outer ring is further marked as 3 / 6 / 12 o'clock directions to express the influence of installation orientation on fault response.

[0044] On this basis, the entity names are regularized and symbolized, and the hierarchical graph nodes are organized in a "from coarse to fine" manner to form a six-layer fault semantic network: the test bench / platform is the zero-level core node to unify the data source; above it are Type (working condition / location type, 4 types), Rev (speed / load combination, 16 types), Size (defect size, 44 types), Location (location level, 119 types), and Clock (outer ring orientation expansion, 160 types). Each node records the label, type, and brief description, and the relationship edge is used to express the semantics of "belongs to / occurs in / corresponds to working condition / corresponds to size". After entering according to this specification, the "specific sample-working condition-defect location-size-orientation" link can be retraced in Neo4j in the form of path query to realize the interpretable presentation of the classification results. Part of the fault node examples and coding specifications of the knowledge graph are shown in Table 1.

[0045] Table 1 2. relation extraction; The embodiment adopts the constructed MRTCNN to perform hierarchical recognition on the fault samples, and extracts the "entity-relation" pairs from the classification results according to the hierarchical recognition. According to the mechanism hierarchy of the data set, the samples are respectively labeled with four categories, 16 categories, 44 categories, 119 categories and 160 categories. The data preprocessing takes 1024 data points as a time sequence segment, and after expanding the samples through a sliding window, the training / test set is randomly divided according to 7:3. The different hierarchical classifications of the MRTCNN model are shown in Table 2.

[0046] Table 2 As can be seen from Table 2, the classification results of the third and fifth levels are basically the same. Even in the case of classification up to 160 categories, the model still achieves an accuracy of 98.90%. The fourth level has relatively weak effect, which is mainly affected by the extremely unbalanced data categories and the explosive increase in the number of categories, but its weighted F1 still reaches 97.48%, which still has reference value for engineering applications. The different hierarchical classification results of the MRTCNN model in Table 2 are for testing the generalization ability, and the embodiment compares with Inception and GRU at the fifth level. The GRU (Gate Recurrent Unit, GRU) network is a model that maintains the effect of the LSTM (Long Short Term Memory, LSTM) network, has a simpler structure, fewer parameters, and better convergence performance, and is good at processing time series data sets; the Inception network is the benchmark network of the embodiment. The test results of the three models are shown in Table 3 and Figure 6 Figure 6 It can be seen that the MRTCNN significantly exceeds the two controls after about 20 epochs, and tends to be stable around 80-100 epochs, and the gap between training and testing is always small, which reflects better generalization. The rising stage of Inception is faster than that of GRU, but both of them are significantly lower than the MRTCNN proposed in the embodiment.

[0047] Table 3 ​The advantages of MRTCNN mainly stem from its structural design: Module A quickly extracts local temporal textures and stabilizes gradients, resulting in faster early convergence; the parallel operation of modules B and C with multiple receptive fields enhances inter-class separability and mitigates degradation with depth; Module D assigns weights to key channels and key time slices, further strengthening discriminative features; and Dropout and L2 regularization are combined to suppress overfitting. Overall, MRTCNN achieves higher final accuracy, faster convergence, and more stable test curves in multi-class bearing diagnosis, providing reliable support for subsequent entity and relation extraction and knowledge graph construction based on classification results.

[0048] 3. Map construction; The extracted entities and relations are standardized into triples (head entity, relation, tail entity). Some fault nodes and relations are shown in Table 4. These are then batch-written into Neo4j using Python to construct a fault knowledge graph. CRUD operations (Create, Read, Update, Delete) can be performed using py2neo and Cypher statements. The graph visually displays the attributes and upstream / downstream relationships of each fault node, facilitating the browsing of its semantic context.

[0049] Table 4 like Figure 7 Neo4j supports custom display by level and number of nodes. In this example, the central red node is CWRU_Equipment; the three yellow nodes correspond to DE12, DE48, and FE12 (first-level nodes), and the remaining colors represent child nodes at different levels. The Cypher statement can be used to view all nodes directly connected to DE12_Rev1797_size7: MATCH p=(n:level_3{name:'DE12_Rev1797_size7'})-[:belongs to]-()RETURN p; This graph can also be used in conjunction with question-and-answer retrieval to return relevant fault information, assisting decision-making and thus improving diagnostic efficiency and interpretability.

[0050] To test the recognition capability under small sample and noisy environments, this embodiment uses three operating conditions (0, 1, and 2 HP) as the basis for data augmentation and partitioning, constructs three sets of subset datasets (A, B, and C), and the CWRU small sample dataset partitioning is shown in Table 5. Fault classification experiments are then conducted on these datasets to evaluate the model's performance under small-scale data.

[0051] Table 5 Considering the prevalence of environmental noise in actual working conditions, an additional variable noise scenario was set up to verify noise immunity. The signal-to-noise ratio is defined as: in , respectively the signal power and the noise power.

[0052] (1) Comparison experiment of attention module; In order to verify the effectiveness of the TCAM attention module added in this embodiment, SE, ECA and TCAM of this embodiment are respectively connected on the basis of the multi-scale residual convolutional backbone, and tested under the condition of B data set and SNR=-10dB. The test results of different attention models (see Table 6) show that: the TCAM of this embodiment achieves the highest accuracy, which is about 33.4% higher than the baseline; SE-ResNet and ECA-ResNet can significantly outperform the baseline, but they still lag behind TCAM; the performance of CBAM is between ECA / SE and TCAM, which shows that separate channel or spatial modeling is not enough to cover the time sequence dependence of the vibration signal, and TCAM allocates weights in time dimension and channel dimension at the same time, which is more suitable for the characteristics of bearing vibration and more robust in strong noise.

[0053] Table 6 (2) Model visualization analysis; In order to intuitively evaluate the representation ability of MRTCNN, this embodiment respectively maps the features of input layer, multi-scale stacking layer, attention module and fully connected layer using t-SNE. Under the setting of SNR=-10dB, the model is trained using data set B, and tested on data set C.

[0054] From the t-SNE visualization of Figure 8 and Figure 9 It can be seen that: in the input layer ( Figure 8 ), the samples of each class overlap in a large range in the two-dimensional space, and the boundary between classes is not clear; after multi-scale convolution and TCAM attention ( Figure 9 ), the same class samples converge to a compact cluster in the subspace, and the interval between different classes increases significantly, and the feature separability is obviously improved. This phenomenon also holds under the cross-condition, noisy setting of B→C, which shows that the proposed MRTCNN can effectively suppress noise and condition difference and extract discriminative features, bringing stable recognition performance and better generalization ability for subsequent classification.

[0055] (3) Fault diagnosis experiment under variable working condition and noise condition; In order to test the robustness and cross-condition generalization ability of the model in complex industrial scenes, cross-validation experiments are carried out under the noise environment of SNR=-10dB, selecting A, B and C three load conditions. Record A→B as training in A condition and testing in B condition; the rest of the combinations are similar. Each combination is independently run 10 times, and the average value is taken as the final result. The fault diagnosis of different algorithms under different load conditions is shown in Figure 10It can be seen from the figure that although Inception10 still maintains good performance under variable load conditions, the method (MRTCNN) of the embodiment achieves the highest accuracy in the six cross-condition combinations, and the overall accuracy is stable near 90% or above, which is significantly better than VGG16 and AlexNet. The results show that, thanks to the multi-scale feature extraction and time domain attention design of the model, MRTCNN still maintains stronger robustness and generalization ability in the case of strong noise and load change, meeting the requirements of cross-condition adaptability for actual industrial fault diagnosis.

[0056] To further evaluate the effectiveness of the proposed method, the embodiment carries out verification on the Mechanical Fault Prevention Technology Institute (MFPT) dataset. The data is collected by Eric Bechhoefer of NRG Systems under multi-condition conditions, covering normal condition, inner ring fault and outer ring fault vibration signals. According to the data mechanism, the following constraints and collection settings are obtained: (1) The sampling frequency is 97656 sps and 48828 sps; (2) At 97656 sps, baseline and outer ring fault data are provided, load is 270 pounds, input shaft speed is 25 Hz, at 48828 sps, it contains 7 types of outer ring fault and 7 types of inner ring fault, and the input shaft speed is also 25 Hz; (3) The baseline and outer ring fault of 270 pounds each contain 3 groups of data; the outer ring fault is obtained under the load of 25, 50, 100, 150, 200, 250 and 300 pounds; the inner ring fault is obtained under the load of 0, 50, 100, 150, 200, 250 and 300 pounds.

[0057] Accordingly, entity extraction and naming normalization are carried out, and a 4-layer fault knowledge graph is constructed: the zero-level node of the test bench as the center is taken as the hub of the graph; the first-level nodes are divided into two categories; the second-level nodes are refined into four categories; and the third-level nodes are expanded to 20 categories. Part of the fault node description of MFPT is shown in Table 7.

[0058] Table 7 Through the MRTCNN model constructed in the embodiment, the MFPT data is labeled and divided according to the classification hierarchy: 2-classification, 4-classification and 20-classification tasks are constructed respectively. Each sample is intercepted with a time domain segment of length 1024; the data enhancement strategy is consistent with the CWRU dataset experiment. The dataset is divided into training set and test set according to 7:3, and the training and test samples are not overlapped. The test results of the model on each task are shown in Table 8.

[0059] Table 8 As can be seen from Table 8, the proposed model achieves 100% recognition rate on the first and second level tasks; in the third level fault diagnosis experiment, the accuracy and weighted F1 are both 99.43%, showing strong discrimination ability and stability. To further verify the reliability of the conclusion, the third level classification is taken as the benchmark, and the commonly used algorithms are compared and evaluated. The comparison results of different algorithms in the third experiment are shown in Table 9.

[0060] Table 9 From Figure 11 It can be seen that the convergence speed of the proposed model is significantly faster than GRU and Inception. There is a certain shock in the early training, mainly affected by class imbalance and noise interference; but it quickly enters a stable platform as the training progresses. This shows that the model has stronger discriminant feature extraction and anti-noise robustness, and the final diagnosis performance is better than the two comparison networks.

[0061] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A bearing fault diagnosis method based on multi-scale convolutional neural networks and knowledge graphs, characterized in that, Includes the following steps: S1. Obtain the original bearing vibration signal and the corresponding mechanism knowledge, and perform entity extraction and label standardization based on the mechanism knowledge. The entities include equipment, working conditions and fault locations. S2. The original bearing vibration signal is input into a multi-scale residual attention convolutional network for end-to-end feature extraction and fault identification to obtain multi-level fault classification results. The processing flow of the multi-scale residual attention convolutional network includes: firstly, low-level temporal features are extracted through the residual convolution module; then, multi-scale features are extracted in parallel through the multi-scale depth-separable Inception module; and finally, the features are weighted and enhanced through the temporal channel attention module. S3. Based on the multi-level fault classification results and the mechanism knowledge, construct a bearing fault knowledge graph, and write the diagnosis results, operating conditions and mechanism events into the graph database to form a traceable knowledge graph.

2. The bearing fault diagnosis method according to claim 1, characterized in that, The original bearing vibration signal obtained in step S1 includes environmental noise, and its signal-to-noise ratio is calculated using the following formula: ； in, For signal power, This represents noise power.

3. The bearing fault diagnosis method according to claim 1, characterized in that, In step S1, the label standardization specifically involves dividing the fault information into a multi-level label system from coarse to fine according to the sampling frequency, load and rotation speed, defect size, defect location and outer ring orientation.

4. The bearing fault diagnosis method according to claim 1, characterized in that, The residual convolution module includes a first convolutional layer, a batch normalization layer, an activation function layer, a second convolutional layer, a batch normalization layer, and a residual connection path connected in sequence; when the number of channels or stride of the input and output are inconsistent, residual alignment is achieved through a one-dimensional convolution bypass.

5. The bearing fault diagnosis method according to claim 1, characterized in that, The multi-scale depth-separable Inception module includes multiple parallel branches, each branch sequentially including a one-dimensional convolutional layer for channel dimensionality reduction and a depth-separable convolutional layer with different receptive fields; the outputs of each branch are concatenated in the channel dimension.

6. The bearing fault diagnosis method according to claim 5, characterized in that, The receptive field size of the depth-separable convolutional layer is selected from 3, 5, and 7.

7. The bearing fault diagnosis method according to claim 1, characterized in that, The steps of weighting the input features in the temporal channel attention module include: Temporal global pooling is performed on the input features to obtain temporal weights, and the input features are weighted for the first time using the temporal weights; The features after the first weighting are compressed and activated along the channel dimension to obtain channel weights, and the features are then weighted a second time using the channel weights.

8. The bearing fault diagnosis method according to claim 1, characterized in that, In step S2, when training the multi-scale residual attention convolutional network, the weighted F1 score is used as a comprehensive evaluation index of the model performance.

9. The bearing fault diagnosis method according to claim 1, characterized in that, In step S3, the nodes of the bearing fault knowledge graph are organized hierarchically according to the multi-level fault classification results obtained in step S2, forming a multi-level semantic network including equipment level, operating condition level, size level, location level and orientation level.

10. The bearing fault diagnosis method according to claim 1, characterized in that, In step S3, the relationships in the knowledge graph are extracted based on the mechanistic knowledge and the multi-level fault classification results, and stored in the graph database in the form of triples of head entity, relationship, and tail entity.

Images