Motor bearing fault detection system and method based on robust deep learning

Abstract

The application discloses a motor bearing fault detection system and method based on robust deep learning, belongs to the technical field of mechanical fault detection and intelligent sensing, and extracts features of original vibration signals with labels based on a supervised learning branch network, takes the original vibration signals as reference samples, takes the reference samples with the same fault category and different fault categories as positive samples and negative samples, introduces a triplet loss to optimize the inter-class separation and intra-class aggregation relationship in an embedding feature space, and generates embedding representation; based on an unsupervised learning branch network, the original vibration signals after artificial feature extraction based on the time domain and the frequency domain are encoded, a triplet loss is introduced for unsupervised embedding learning, and high-level feature embedding representation is generated; the embedding representation output by the two branch networks is fused, a double loss of the triplet loss and the center loss is introduced for training, and the bearing fault detection model after training is used for bearing fault detection.

Description

Technical Field

[0001] This invention belongs to the field of mechanical fault detection and intelligent sensing technology, specifically relating to a motor bearing fault detection system and method based on robust deep learning. Background Technology

[0002] In key sectors such as industrial manufacturing, rail transportation, and wind power operation and maintenance, motor bearings, as core components of rotating machinery, directly impact the safety and stability of the equipment. Therefore, conducting research on high-precision and robust motor bearing fault detection is crucial for achieving health management and predictive maintenance of equipment. In recent years, with the rapid development of deep learning technology, researchers have gradually shifted from traditional manual feature extraction methods to end-to-end neural network modeling, utilizing techniques such as convolutional neural networks (CNNs), autoencoders, and transfer learning to automatically extract high-level semantic features from raw vibration signals, enabling automatic fault type identification. However, existing deep learning methods still suffer from insufficient robustness and weak inter-class discrimination when facing practical application scenarios with scarce and unevenly distributed labeled samples and variable operating conditions. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies and achieve adaptability and robustness in bearing fault detection under various operating conditions, this invention adopts the following technical solution:

[0004] A robust deep learning-based method for detecting motor bearing faults includes the following steps:

[0005] Step 101: Based on the semi-supervised learning branch network, feature extraction is performed on the original vibration signal with labels. The original vibration signal is used as the reference sample. The samples with the same fault category as the reference sample are positive samples, and the samples with different fault categories are negative samples. Triple loss is introduced to optimize the inter-class separation and intra-class aggregation relationship in the embedding feature space, generate the embedding representation, and use cross-entropy loss and triple loss for joint training to perform classification and structure optimization.

[0006] Step 102: Based on the unsupervised learning branch network, the original vibration signal after manual feature extraction in the time and frequency domains is encoded using a multilayer perceptron, and a triplet loss is introduced for unsupervised embedding learning to generate a high-level feature embedding representation. The unsupervised learning branch can extract structurally effective high-level feature representations in the unlabeled sample environment, making up for the modeling defects caused by insufficient supervision information.

[0007] Step 103: Construct a fusion optimization module to fuse the embedded representations output by the two branch networks to obtain a unified feature representation. Then, introduce a dual loss, including triplet loss and center loss, for training. Traditional triplet loss can only optimize the inter-class distance, but it has no clear constraint on the distribution of intra-class samples, which can easily lead to the scattering of intra-class samples. In this invention, the center loss is used to gather intra-class samples and alleviate sample drift. The combination of the two can form a clear and compact distribution cluster structure in the feature space, which not only enhances the separability of fault types in the feature space, but also accelerates the stability of convergence, realizes the fusion and alignment of embedded features of different modalities, and strengthens the structural discriminativeness and fault tolerance robustness of the model by means of the dual loss optimization strategy, thereby greatly improving the detection accuracy and stability.

[0008] Step 104: Generate the trained bearing fault detection model for bearing fault detection.

[0009] Further, in step 101, a semi-supervised learning branch network is constructed to acquire one-dimensional time-series data of the original vibration signal and extract temporal features. Multiple convolutional residual blocks are used to capture local temporal structures and long-term dependencies. A channel attention mechanism is employed to weight and adjust the convolutional feature maps to highlight key fault modes. A global average pooling layer compresses the high-dimensional feature maps into fixed-length global feature vectors. Simultaneously, fully connected layers and a Softmax output layer generate embedded vector representations and category prediction distributions. Each residual block contains two consecutive 1D convolutional layers and residual connection structures. This invention can train a model with strong identification capabilities for motor bearing fault categories using only partial labeled data.

[0010] Furthermore, in step 101, the semi-supervised learning uses a joint loss function for training. This joint loss function includes cross-entropy loss and triplet loss. The cross-entropy loss is used to optimize classification accuracy during the semi-supervised learning stage. Most existing supervised learning methods for bearing fault detection only use cross-entropy loss for classification learning, failing to explicitly optimize the clustering structure of sample distribution in the feature space. This invention, however, relies solely on cross-entropy loss, jointly training category discrimination and embedding space structure. By jointly introducing triplet loss, the embedding structure is optimized simultaneously outside of classification, thereby improving the model's sensitivity and separability to the boundaries between similar fault types. This effectively enhances the model's ability to distinguish fault types in the embedding space, improving the model's adaptability and robustness under various operating conditions.

[0011] Furthermore, in step 102, an unsupervised learning branch network is constructed. Statistical feature engineering is used to extract artificial features from the time and frequency domains respectively. Unlabeled artificial features are then embedded and optimized, rather than through deep automatic feature extraction. This reduces the risk of misleading the model on unlabeled data. A multilayer perceptron is used to generate high-level feature embedding vectors of fixed dimensions to describe the distribution structure of the input samples. Existing unsupervised learning methods, if they directly perform deep encoding on the original signal, are prone to pseudo-structure shifts due to a lack of label guidance, leading to model embedding space failure. This invention, through a combination of shallow feature extraction and deep embedding learning, enables the system to still possess structural representation capabilities under unsupervised conditions, achieving structured modeling of the original vibration signal in the embedding space. The multilayer perceptron includes multiple fully connected layers and nonlinear activation functions. Each fully connected layer is followed by batch normalization to improve training stability.

[0012] Furthermore, in step 102, the original vibration signal is first segmented by a sliding window to obtain a fixed-length signal segment. Then, various manual feature extraction methods are used to obtain characterization information from the time domain and frequency domain, respectively. The time domain features include, but are not limited to, mean, standard deviation, kurtosis, waveform factor, and impulse factor; the frequency domain features include, but are not limited to, spectral centroid, dominant frequency, bandwidth, and frequency domain energy distribution.

[0013] Furthermore, in step 102, the original vibration signal feature vector is used as a reference sample. A triplet is constructed by introducing the reference sample and positive samples of the same fault category and negative samples of different fault categories. The triplet loss function is used as the training target in the unsupervised learning stage to control the embedding distance relationship between bearing vibration signals. Through training, positive samples are made closer to the reference sample and negative samples are made farther away, thereby enhancing the separability between fault modes, realizing the structural optimization of the feature embedding space, and further optimizing the class separability of bearing fault features in the embedding space.

[0014] Embedding optimization guided by triplet loss enables samples to form a good clustering structure in the feature space, providing more robust semantic support for subsequent fusion and classification.

[0015] Furthermore, in step 102, a group of signals with similar spectral characteristics or temporal statistical characteristics are regarded as the same cluster, and the embedding vectors with similar Euclidean distances in the high-dimensional feature space are regarded as the neighborhood structure, so that the generation of triple samples can still be performed in the unlabeled scenario. Based on the triple samples of the signal group, rather than a single time point or sampling points within a time series, each time series is mapped to a high-dimensional feature vector through a deep embedding function, and the Euclidean distance between the reference sample and the positive sample and the Euclidean distance between the reference sample and the negative sample are calculated. The separability of fault categories in the feature space is optimized by maximizing the difference between the two Euclidean distances.

[0016] Furthermore, the fusion optimization module in step 103 includes splicing operations, multi-source feature weighting mechanisms, and consistency regularization mechanisms. This invention does not simply stack two branches, but uses consistency regularization mechanisms to make the two branch networks complementary rather than conflicting in their embedding representations. It combines intra-class aggregation and inter-class separation, taking into account both space "compression" and "stretching", to achieve dual optimization of "discrimination ability + stable convergence". By constraining the cosine similarity of the embedding representations of the two branch networks, the feature representations of the same fault sample are made consistent.

[0017] Furthermore, in step 104, the unified feature representation is connected to the backend classifier to perform the final fault discrimination. At the same time, the trained bearing fault detection model is serialized, saved, and deployed. The bearing fault detection model of the present invention has strong deployment capability, lightweight structure, supports edge deployment + real-time detection in industrial field, reduces model complexity, and the optimization of embedded spatial structure brings higher "robustness" and "fault tolerance", so that the model can maintain high detection accuracy even under complex working conditions. Through good clustering of fault sample spatial distribution, new fault samples are easily classified into the correct cluster even if they are unlabeled.

[0018] The robust deep learning-based motor bearing fault detection device includes a semi-supervised learning branch module, an unsupervised learning branch module, a fusion optimization module, and an output module. It sequentially executes the construction and training of the semi-supervised learning branch network, the unsupervised learning branch network, and the fusion optimization module in the robust deep learning-based motor bearing fault detection method, and generates a trained bearing fault detection model for bearing fault detection.

[0019] The advantages and beneficial effects of this invention are as follows:

[0020] This invention addresses the problems of scarce labeled samples, complex feature distribution, and significant differences in operating conditions in motor bearing vibration signals. It constructs a deep feature extraction architecture that integrates semi-supervised and unsupervised learning strategies. A semi-supervised learning branch receives labeled raw vibration signals, extracts convolutional feature maps, and generates embedded representations. This is then jointly trained using cross-entropy loss and triplet loss, ensuring classification accuracy while enhancing the structural discriminative power of the feature space. Simultaneously, an unsupervised learning branch receives unlabeled sample features extracted from the time and frequency domains. A high-dimensional embedded representation is achieved through a multilayer perceptron, and triplet loss is used for structural optimization, effectively improving the utilization efficiency and distribution perception capability of unlabeled samples. Furthermore, the embedding outputs of the two branches are concatenated or weighted in a fusion optimization module, and a dual loss function consisting of triplet loss and center loss is introduced as the overall optimization objective, thereby simultaneously improving intra-class compactness and inter-class separability. This invention can extract stable and robust bearing fault features in multi-condition environments, improve the model's discrimination and generalization capabilities, and enable the final output bearing fault detection model to have higher accuracy and stability in complex engineering environments. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the system structure in an embodiment of the present invention.

[0022] Figure 2 This is a flowchart of the method in an embodiment of the present invention. Detailed Implementation

[0023] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0024] like Figure 1 As shown, the motor bearing fault detection system based on robust deep learning includes a semi-supervised learning branch module, an unsupervised learning branch module, a fusion optimization module, and an output module.

[0025] The semi-supervised learning branch module receives raw vibration signal data with labels, extracts convolutional feature maps and generates embedding representations, and uses cross-entropy loss and triplet loss for joint training to perform classification and structure optimization.

[0026] The unsupervised learning branch module receives unlabeled samples extracted from manual features in the time and frequency domains, encodes the input using a multilayer perceptron, and performs unsupervised embedding learning using triplet loss.

[0027] The fusion optimization module concatenates or weights the embedded features output by the semi-supervised learning branch and the unsupervised learning branch for subsequent classification or similarity measurement modules. It also introduces a dual loss function as the optimization objective of the entire system, which is composed of a combination of triplet loss and center loss.

[0028] The output module outputs an optimized bearing fault detection model for bearing fault detection.

[0029] like Figure 2 As shown, the motor bearing fault detection method based on robust deep learning includes the following steps:

[0030] Step 101: Semi-supervised learning branch. Receive raw vibration signal data with labels, extract convolutional feature maps and generate embedding representations, and use cross-entropy loss and triplet loss for joint training to perform classification and structure optimization.

[0031] In one embodiment, the semi-supervised learning branch network structure includes an input layer, multiple convolutional residual blocks, an attention module, and an output embedding layer, which carry out the entire process from the original vibration signal to the embedded feature representation.

[0032] The original vibration signal was collected by a sensor during the operation of the motor bearing, and is a one-dimensional time series data. After normalization preprocessing, the input signal is fed into a semi-supervised learning branch network. This branch uses a one-dimensional convolutional neural network (1D-CNN) as its core structure, mainly including the following modules: Initial convolutional layer: used to extract low-level temporal features, with a kernel size of 16, a stride of 2, and 64 output channels; Three sets of convolutional residual blocks: each set of residual blocks contains two consecutive 1D convolutional layers and residual connection structures, used to capture local temporal structures and long-term dependencies; Attention mechanism module: uses a channel attention mechanism to weight and adjust the convolutional feature map to highlight key fault modes; Global average pooling layer: compresses the high-dimensional feature map into a fixed-length global feature vector; Fully connected layer and Softmax output layer: used to generate embedding vector representations and class prediction distributions. In one embodiment, the semi-supervised learning branch is trained using a joint loss function, specifically including: cross-entropy loss: used to optimize classification accuracy in the semi-supervised learning stage; triplet loss: by constructing anchor samples, positive samples, and negative sample triples, it optimizes the inter-class separation and intra-class aggregation relationships in the embedded feature space. The joint loss function is expressed as follows:

[0033]

[0034] in, For cross-entropy loss, λ is the triplet loss, and λ is the weighting coefficient used to adjust the contribution ratio of the two to the training objective.

[0035] The innovation lies in introducing a structure-optimized triplet loss into semi-supervised learning. Instead of relying solely on cross-entropy, it jointly trains category discrimination and embedding space structure to achieve structure-aware feature learning. In contrast, most existing supervised learning methods for bearing fault detection only use cross-entropy loss for classification, failing to explicitly optimize the clustering structure of samples in the feature space. This invention, by jointly introducing triplet loss, simultaneously optimizes the embedding structure beyond classification, thereby improving the model's sensitivity and separability to boundaries between similar fault types.

[0036] Through the above embodiments, a model with strong identification ability for motor bearing fault categories can be trained using only partial labeled data. The combined use of cross-entropy loss and triplet loss effectively enhances the model's ability to distinguish fault types in the embedding space, improving its adaptability and robustness under various operating conditions.

[0037] Step 102: Construct an unsupervised learning branch, receive unlabeled samples extracted from manual features in the time and frequency domains, encode the input using a multilayer perceptron, and perform unsupervised embedding learning using triplet loss.

[0038] In one embodiment, the unsupervised learning branch makes full use of a large number of unlabeled vibration samples and achieves structured modeling of the samples in the embedding space by combining shallow feature extraction with deep embedding learning.

[0039] The input data consists of manually extracted features obtained through statistical feature engineering. Specifically, the original vibration signal is first segmented by a sliding window to obtain fixed-length signal segments. Then, various manual feature extraction methods are used to obtain representational information from the time and frequency domains, respectively. Time-domain features include, but are not limited to: mean, standard deviation, kurtosis, waveform factor, impulse factor, etc.; frequency-domain features include, but are not limited to: spectral centroid, dominant frequency, bandwidth, frequency energy distribution, etc. After standardization, the above features are concatenated into a vector, which serves as the input to the unsupervised learning branch. The innovation of constructing an unsupervised learning branch using time-frequency domain manual features + multilayer perceptron + triplet loss for embedding optimization of unlabeled data lies in using manual time-frequency features instead of deep automatic feature extraction, thus reducing the "misguided risk" of the model on unlabeled data. The combination strategy of shallow features + deep embedding enables the system to still have structural representation capabilities under unsupervised conditions; the explicit combination of the definitions of "same cluster" and "nearby structure" ensures that triplet sample generation can still be performed in unlabeled scenarios. Existing unsupervised learning methods, if they directly perform deep encoding on the original signal, are prone to falling into pseudo-structure shifts due to the lack of label guidance, leading to the failure of the model's embedding space. This invention avoids the above problems by constructing structured triples through prior manual feature extraction, ensuring the usability and generalization ability of the unsupervised path.

[0040] The unsupervised branch structure includes key components such as manually input features, stacked fully connected layers, output embedding vectors, and loss optimization paths. The unsupervised branch employs a multilayer perceptron structure, primarily composed of multiple fully connected layers and the ReLU nonlinear activation function. Each fully connected layer is followed by batch normalization to improve training stability. The network's final output is a fixed-dimensional high-level feature embedding vector, used to describe the distribution structure of the input samples.

[0041] To optimize the class separability of bearing fault features in the embedding space, the system introduces a triplet loss function as the core training objective in the unsupervised learning stage. This loss function controls the embedding distance relationship between bearing vibration signals by constructing triplet samples (i.e., reference sample, positive sample, and negative sample). In this system, the reference sample is the vibration signal feature vector of a certain fault type (e.g., outer race fault), the positive sample is other vibration signals belonging to the same class (same fault category or operating condition), and the negative sample comes from the signal features of another type of fault (e.g., inner race fault or normal state). Furthermore, 'same cluster' refers to a group of signal samples with similar spectral features or temporal statistical features (e.g., RMS, kurtosis); 'neighborhood structure' refers to the set of embedding vectors that are close in Euclidean distance in the high-dimensional feature space. Through training, positive samples are encouraged to move closer to the reference sample and negative samples to move further away, thereby enhancing the separability between fault modes; the reference sample and negative sample are "dissimilar" (e.g., sampled from different category regions); the network optimizes the structure of the feature embedding space by minimizing the distance between the reference sample and the positive sample while maximizing the distance between the reference sample and the negative sample.

[0042]

[0043] Among them, anchor point samples Typically, this represents a complete vibration time series signal under a certain fault condition (such as "outer ring peeling"), positive sample. For another set of signals with the same fault type or operating conditions, negative samples These are sample signals of different categories (such as "inner ring fault" or "normal operation"). The system uses a deep embedding function. Each time series is mapped to a high-dimensional feature vector, and the Euclidean distance between the anchor point and positive samples, and the Euclidean distance between the anchor point and negative samples are calculated based on this vector. The separability of fault categories in the feature space is optimized by maximizing the difference between these two distances. The first stage of the construction is represented by the... The triplet sample index, rather than the sampling point number within a single time point or time series.

[0044] Through the above embodiments, the constructed unsupervised learning branch can extract structurally effective high-level feature representations in unlabeled sample environments, making up for the modeling defects caused by insufficient supervision information; at the same time, the embedding optimization guided by triple loss enables the samples to form a good clustering structure in the feature space, providing more robust semantic support for subsequent fusion and classification.

[0045] Step 103: Construct a fusion optimization module to concatenate or weightedly fuse the embedded features output from the two branches for subsequent classification or similarity measurement modules, and introduce a dual loss function as the optimization objective of the entire system. The dual loss function is composed of a combination of triplet loss and center loss.

[0046] In one embodiment, the embedding vectors output by the semi-supervised learning branch and the unsupervised branch are concatenated into a unified feature representation and fed into a fusion module. The fusion module includes a concatenation operation, a multi-source feature weighting mechanism, and a consistency regularization mechanism. The innovation lies in not simply stacking the two branches, but introducing consistency regularization to ensure that the two branches are complementary rather than conflicting in representation; combining intra-class aggregation and inter-class separation, it balances space "compression" and "stretching," achieving a dual optimization of "discriminative power + stable convergence." Traditional triplet loss can only optimize inter-class distance, but it has no clear constraint on the distribution of intra-class samples, easily leading to scattered intra-class samples. In this system, the introduction of center loss effectively gathers intra-class samples and alleviates the sample drift problem; the combination of the two can form a clear and compact distribution cluster structure in the feature space, thereby greatly improving diagnostic accuracy and stability. Consistency regularization improves the consistency of feature representation for the same fault sample by constraining the cosine similarity of the embedding results of the two branches. Furthermore, a joint loss function is designed for the overall model, composed of triplet loss and center loss.

[0047]

[0048] The triplet loss expands the boundaries between different classes by maximizing the distance between the anchor sample and the negative sample while minimizing the distance between the anchor sample and the positive sample; the center loss enhances intra-class clustering by bringing the feature vectors of samples of the same class closer to their class center. This design not only enhances the separability of fault types in the feature space but also accelerates the stability of system convergence. Through the above embodiments, the fusion and alignment of embedded features from different modalities are achieved, and the structural discriminativeness and fault tolerance robustness of the model are strengthened by the dual-loss optimization strategy.

[0049] Step 104: Construct an output module to output the trained bearing fault detection model.

[0050] In one embodiment, the unified feature representation output by the fusion module can be connected to the backend classifier for final fault identification. Simultaneously, the system can serialize and save the trained model, deploying it to edge devices or industrial field platforms for real-time detection of bearing operating status and outputting fault type judgment results. The model boasts strong deployment capabilities: lightweight structure, supporting both edge deployment and real-time industrial field detection. Its innovation lies in the model design's consideration of edge deployment feasibility. For example, it uses manually generated features to simplify the unsupervised branch structure, reducing model complexity; optimized embedding space structure brings higher robustness and fault tolerance, enabling the model to maintain high detection accuracy even under complex operating conditions; and through good clustering of fault sample spatial distribution, new fault samples, even without labels, are easily classified into the correct cluster.

[0051] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A motor bearing fault detection method based on robust deep learning, characterized in that... Includes the following steps: Step 101: Based on the semi-supervised learning branch network, feature extraction is performed on the original vibration signal with labels. The original vibration signal is used as a reference sample. The samples with the same fault category as the reference sample are positive samples, and the samples with different fault categories are negative samples. Triple loss is introduced to optimize the inter-class separation and intra-class aggregation relationship in the embedded feature space and generate the embedded representation. Step 102: Based on the unsupervised learning branch network, the original vibration signal after manual feature extraction in the time domain and frequency domain is encoded, and triplet loss is introduced to perform unsupervised embedding learning to generate high-level feature embedding representation. Step 103: Construct a fusion optimization module to fuse the embedding representations output by the two branch networks to obtain a unified feature representation, and introduce a dual loss including triplet loss and center loss for training, and use the center loss to gather in-class samples. Step 104: The trained bearing fault detection model is used for bearing fault detection.

2. The motor bearing fault detection method based on robust deep learning according to claim 1, characterized in that: In step 101, a semi-supervised learning branch network is constructed to obtain one-dimensional time series data of the original vibration signal and extract temporal features. Local temporal structure and long-term dependencies are captured through multiple convolutional residual blocks. Channel attention mechanism is used to weight and adjust the convolutional feature map to highlight key fault modes. The high-dimensional feature map is compressed into a fixed-length global feature vector through a global average pooling layer. At the same time, the embedded vector representation and category prediction distribution are generated through a fully connected layer and a Softmax output layer.

3. The motor bearing fault detection method based on robust deep learning according to claim 1, characterized in that: In step 101, semi-supervised learning is trained using a joint loss function, which includes cross-entropy loss and triplet loss. Cross-entropy loss is used to optimize classification accuracy during the semi-supervised learning stage.

4. The motor bearing fault detection method based on robust deep learning according to claim 1, characterized in that: In step 102, an unsupervised learning branch network is constructed, and artificial features are extracted from the time domain and frequency domain respectively through statistical feature engineering. The unlabeled artificial features are embedded and optimized, and a multilayer perceptron is used to generate high-level feature embedding vectors to describe the distribution structure.

5. The motor bearing fault detection method based on robust deep learning according to claim 1, characterized in that: In step 102, the original vibration signal is first segmented by a sliding window to obtain a fixed-length signal segment. Then, various manual feature extraction methods are used to obtain characterization information from the time domain and frequency domain, respectively. The time domain features include mean, standard deviation, kurtosis, waveform factor, and impulse factor. Frequency domain characteristics include spectral centroid, dominant frequency, bandwidth, and frequency domain energy distribution.

6. The motor bearing fault detection method based on robust deep learning according to claim 1, characterized in that: In step 102, the original vibration signal feature vector is used as a reference sample. A triplet is constructed by introducing the reference sample and positive samples of the same fault category and negative samples of different fault categories. The triplet loss function is used as the training target for the unsupervised learning stage to control the embedding distance relationship between bearing vibration signals. Through training, positive samples are made closer to the reference sample and negative samples are made farther away.

7. The motor bearing fault detection method based on robust deep learning according to claim 1, characterized in that: In step 102, a group of signals with similar spectral or temporal statistical characteristics are considered as the same cluster. Embedding vectors with similar Euclidean distances in the high-dimensional feature space are considered as the neighborhood structure. Based on the triplet samples of the signal group, each time series is mapped to a high-dimensional feature vector through a deep embedding function. The Euclidean distance between the reference sample and the positive sample, and the Euclidean distance between the reference sample and the negative sample are calculated accordingly. The separability of fault categories in the feature space is optimized by maximizing the difference between the two Euclidean distances.

8. The motor bearing fault detection method based on robust deep learning according to claim 1, characterized in that: The fusion optimization module in step 103 includes splicing operation, multi-source feature weighting mechanism and consistency regularization mechanism. The consistency regularization mechanism makes the two branch networks complementary in their embedding representations. It combines intra-class aggregation and inter-class separation. By constraining the cosine similarity of the embedding representations of the two branch networks, the feature expression of the same fault sample is made consistent.

9. The motor bearing fault detection method based on robust deep learning according to claim 1, characterized in that: In step 104, the unified feature representation is connected to the backend classifier to perform the final fault identification. At the same time, the trained bearing fault detection model is serialized, saved, and deployed.

10. A motor bearing fault detection device based on robust deep learning, comprising a semi-supervised learning branch module, an unsupervised learning branch module, a fusion optimization module, and an output module, characterized in that: The semi-supervised learning branch network, unsupervised learning branch network, and fusion optimization module in the robust deep learning-based motor bearing fault detection method according to any one of claims 1 to 9 are constructed and trained in sequence, and a trained bearing fault detection model is generated for bearing fault detection.

Images