Improved motor bearing fault diagnosis method based on SE-CAE denoising and two-stage domain adaptive

Abstract

The application discloses a motor bearing fault diagnosis method with improved SE-CAE denoising and two-stage domain adaptation, and belongs to the technical field of motor state monitoring and intelligent fault diagnosis. The method first constructs an improved SE-CAE network, adopts a joint loss function composed of a time domain reconstruction loss, a sparse constraint and a frequency domain consistency to train, and realizes signal denoising enhancement; then constructs a diagnosis network with ResNet as a backbone and a domain-specific batch normalization introduced, adopts two-stage training, that is, in the first stage, a cross-entropy and a center loss are jointly optimized to construct an intra-class compact feature space; in the second stage, a source domain supervision constraint, a domain adversarial loss and a high-confidence pseudo-label loss embedded with a center loss are jointly optimized to realize gradual migration from a global level to a category level. The application can significantly improve the precision and robustness of bearing fault diagnosis under a strong mixed noise and a variable working condition coupling scene, and has good engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of motor condition monitoring and intelligent fault diagnosis technology, specifically involving an improved SE-CAE denoising and two-stage domain adaptive motor bearing fault diagnosis method. Background Technology

[0002] Bearings are critical components in rotating electrical machines, and their health directly affects the safety and reliability of the entire machine. Early bearing failures produce weak, short-duration impact signals that are easily masked by the complex noise environment of industrial sites. In actual operation, vibration signals not only contain Gaussian white noise (originating from sensor noise and environmental interference) but are often superimposed with random impulse noise (originating from non-Gaussian interference such as load changes and transient impacts), thus creating a strongly mixed noise environment.

[0003] To address the aforementioned noise interference issues, existing technologies primarily employ signal processing methods or deep learning models for denoising. Specifically, signal processing techniques such as Wavelet Transform (WT), Empirical Mode Decomposition (EMD), and Variational Mode Decomposition (VMD) rely on manual parameter adjustments, which can easily lead to modal aliasing and over-smoothing of critical fault impacts under conditions of strong mixed noise, resulting in significant loss of subtle fault features. In contrast, existing deep learning denoising models often employ standard Convolutional Autoencoders (CAEs), whose loss functions typically optimize only the minimization of temporal signal reconstruction errors (such as mean square error). However, this single-constraint approach is not specifically designed for scenarios with strong mixed noise; it only pursues minimizing temporal reconstruction loss, easily neglecting the consistency and sparsity of frequency domain fault features. This results in a significant amount of noise remaining in the denoised signal, failing to provide high-quality input for backend diagnostics.

[0004] Meanwhile, motors often operate under varying conditions in real-world conditions. Different speeds and loads can cause data distribution shifts, leading to a decline in model transfer performance. To address this cross-condition fault diagnosis problem, existing technologies often employ domain-adaptive methods. Global alignment methods, such as Correlation Alignment (CORAL), Domain-Adversarial Neural Network (DANN), and Maximum Mean Discrepancy (MMD), focus solely on reducing the overall distribution difference between the source and target domains, neglecting the compactness of intra-class features. This can easily lead to feature dispersion within the target domain and overlapping inter-class boundaries, ultimately causing negative transfer and severely reducing the accuracy and stability of cross-condition diagnosis.

[0005] Furthermore, most existing studies treat denoising and cross-condition adaptive diagnosis as two separate steps, which makes it difficult to cope with coupled scenarios where strong mixed noise and variable condition distribution drift coexist.

[0006] In summary, achieving high-precision and robust diagnosis of motor bearing faults under complex conditions of strong mixed noise interference and drift in variable operating range has become an urgent technical problem to be solved in this field.

[0007] In view of this, the present invention is hereby proposed. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide an improved SE-CAE denoising and two-stage domain adaptive method for motor bearing fault diagnosis. This method primarily addresses the technical problems of existing denoising methods easily losing fault features under strong mixed noise, existing domain adaptive methods easily inducing intra-class dispersion and negative migration under varying operating conditions, and the lack of a collaborative mechanism between denoising and diagnosis. The technical solution provided by this invention can achieve high-precision and robust identification and diagnosis of motor bearing health status and fault types in complex industrial environments where strong mixed noise interference and variable operating condition domain drift coexist.

[0009] The objective of this invention is achieved through the following technical solution:

[0010] This invention provides an improved SE-CAE denoising and two-stage domain adaptive method for motor bearing fault diagnosis, comprising the following steps:

[0011] Step 1, Noise Injection: Obtain the original vibration signal of the motor bearing, and superimpose noise onto the original vibration signal to construct a strong mixed noise sample;

[0012] Step 2, Denoising and Enhancement Processing: Construct an improved compressed-excited convolutional autoencoder (SE-CAE) network. The SE-CAE network consists of an encoder, a decoder, and a channel attention mechanism embedded in the encoder. It is trained using a joint loss function consisting of temporal reconstruction loss, sparse constraint loss, and frequency domain consistency loss. The trained SE-CAE network is then used to denoise and enhance the strongly mixed noise samples to obtain reconstructed samples.

[0013] Step 3, Data partitioning: The reconstructed samples are divided into labeled source domain data and unlabeled target domain data according to the working conditions, and the target domain data is further divided into target domain training set and target domain test set according to a preset ratio;

[0014] Step 4, Domain Adaptive Training: Construct a domain adaptive diagnostic network with ResNet as the backbone and introduce domain-specific batch normalization, and train the diagnostic network using a two-stage training strategy, wherein:

[0015] Phase 1: Supervised pre-training of the diagnostic network using the source domain data, with optimization objectives including cross-entropy loss and center loss, to obtain pre-trained weights;

[0016] The second stage involves loading the pre-trained weights as initial parameters for the diagnostic network, and then using the source domain data and the target domain training set to perform unsupervised domain adaptive training on the diagnostic network. The optimization objectives include the cross-entropy loss and center loss of the source domain, the domain adversarial loss, and the high-confidence pseudo-label loss of the embedding center loss.

[0017] Step 5: Output the results: Input the target domain test set into the diagnostic network after the two-stage training is completed, and output the fault diagnosis results of the motor bearing.

[0018] Furthermore, in step 1, the strong mixed noise sample is constructed by superimposing Gaussian white noise and random impulse noise onto the original vibration signal;

[0019] The signal-to-noise ratio of the Gaussian white noise ranges from -8dB to -3dB; the sparsity of the random impulse noise ranges from 1% to 3%, and the amplitude ranges from 4 to 7 times the signal standard deviation.

[0020] Furthermore, in step 2, the encoder consists of multiple one-dimensional convolutional layers, used to extract multi-scale features of strongly mixed noise samples;

[0021] The decoder is symmetrical to the encoder and consists of multiple one-dimensional deconvolutional layers, used to reconstruct the features output by the encoder into a time-domain signal.

[0022] The channel attention mechanism is embedded in each layer of the encoder, including: a global average pooling layer, used to compress the feature maps output by each layer of the encoder into channel statistics; a first fully connected layer and a second fully connected layer connected in sequence, used to learn the non-linear dependencies between channels and generate normalized channel weights; and a gating mechanism, used to multiply the normalized channel weights with the feature maps output by the encoder channel by channel to obtain a reweighted output, thereby realizing feature recalibration.

[0023] Further, in step 2, the joint loss function is expressed as:

[0024]

[0025] The temporal reconstruction loss is expressed as:

[0026]

[0027] The sparse constraint loss is expressed as:

[0028]

[0029] The frequency domain consistency loss is expressed as:

[0030]

[0031] In the formula, For joint losses; For the first The sample at the th The original vibration signals at each time point For the first The sample at the th Reconstructed signals at each time point This represents the total number of samples within the batch. The time sequence length of a single original vibration signal. For the first One original vibration signal In the The STFT spectrum of the [number] time frame The amplitude of each frequency component, For the first One reconstructed signal In the The STFT spectrum of the [number] time frame The amplitude of each frequency component, The total number of frequency points. For time frames, For a norm, The square of the L2 norm. The sparse constraint loss balance coefficient, This is the frequency domain consistency loss balance coefficient.

[0032] Furthermore, the sparse constraint loss balance coefficient and frequency domain consistency loss balance coefficient All were determined using the grid search method;

[0033] Wherein, the sparse constraint loss balance coefficient The value range is from 0.005 to 0.05, and the frequency domain consistency loss balance coefficient is... The value range is from 0.0005 to 0.005.

[0034] Further, in step 4, the domain-specific batch normalization is expressed as:

[0035]

[0036] In the formula, The output features are those after domain-specific batch normalization. The features to be normalized are the outputs of the convolutional layer. , The values ​​are, in order, the mean and standard deviation of the source domain. , The parameters are, in order, the scaling and offset parameters of the source domain. , The values ​​are, in order, the mean and standard deviation of the target domain. , The scaling and offset parameters for the target domain are listed in order.

[0037] Furthermore, in the first stage of step 4, the joint optimization objective of the cross-entropy loss and the center loss is expressed as:

[0038]

[0039] The cross-entropy loss of the source domain is expressed as:

[0040]

[0041] The central loss of the source domain is expressed as:

[0042]

[0043] This represents the total loss for the first phase. For the first The reconstructed signal belonging to the source domain is the signal that has been denoised by the SE-CAE network and then divided into the source domain in step 3. For the first The actual fault category label corresponding to each reconstructed signal; To diagnose the network, the signal will be reconstructed. Predicted as the true category The probability of; This represents the total number of samples in the source domain training batch. For the first The feature vector of the reconstructed signal is extracted by the diagnostic network; For the corresponding category The feature center; The square of the L2 norm; The center loss weight has a value range of 0.02 to 0.06.

[0044] Furthermore, in the second stage of step 4, the total loss function is expressed as:

[0045]

[0046] In the formula, This represents the total loss for the second phase. The total loss for the first stage includes the cross-entropy loss of the source domain and the center loss of the source domain; To achieve domain adversarial loss, a gradient inversion layer is used to align the global domain distribution. The high-confidence pseudo-label loss incorporates center loss, and the cross-entropy loss is derived from the target domain. and the central loss of the target domain Composition, that is , , where represents the weighting coefficient of the center loss in the pseudo-label loss; Dynamic weights for domain adversarial loss. Dynamic weights are used to compensate for high-confidence pseudo-label loss.

[0047] Furthermore, the dynamic weights of the domain adversarial loss The weights of the high-confidence pseudo-label loss are smoothly increased from 0 to 1.0 during training; The value was linearly increased from 0.10 to 0.25 during training.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. The motor bearing fault diagnosis method provided by this invention achieves significant synergistic enhancement effects in complex scenarios where strong mixed noise and variable operating condition domain drift coexist. This is achieved through the synergistic cooperation of an improved SE-CAE network and a domain adaptive diagnosis network constructed using ResNet and DSBN. Specifically, the improved SE-CAE network is first trained using a triple joint loss function consisting of temporal reconstruction loss, sparse constraint loss, and frequency domain consistency loss. This network effectively preserves subtle fault impact details and frequency domain feature structures while suppressing Gaussian white noise and random impulse noise, providing high-quality reconstruction samples for the backend diagnosis network. Based on this, the diagnosis network is trained using a two-stage training strategy: the first stage utilizes labeled source domain data to jointly optimize cross-entropy loss and center loss, constructing a compact intra-class feature space to obtain pre-trained weights; the second stage loads these pre-trained weights as initial parameters for the diagnosis network, and simultaneously utilizes source domain data and the target domain training set to jointly optimize the cross-entropy loss, center loss, domain adversarial loss, and high-confidence pseudo-label loss of the embedding center loss, achieving a gradual transfer from global alignment to class alignment. The two steps above complement each other: if the front-end denoising is not thorough, the back-end domain adaptation will be affected by noise, leading to increased negative migration; if the back-end domain adaptation capability is insufficient, the high-quality signal after front-end denoising still cannot overcome the domain drift problem under varying operating conditions. This invention organically combines the two, significantly improving the accuracy of motor bearing fault diagnosis.

[0050] 2. This invention improves the SE-CAE network at two levels: loss function and network structure, specifically for scenarios with strong mixed noise. At the loss function level, a triple joint loss function is designed, consisting of temporal reconstruction loss, sparsity constraint loss, and frequency domain consistency loss. This triple synergistic constraint achieves a balance between noise suppression and fault feature preservation: temporal reconstruction loss ensures overall signal restoration accuracy, sparsity constraint loss highlights the sparsity characteristics of fault impacts, and frequency domain consistency loss preserves the frequency structure of fault features. At the network structure level, a compression and excitation channel attention mechanism is embedded in the encoder. Channel weights are generated through global average pooling, fully connected layers, and gating mechanisms to recalibrate the feature map, highlighting key channel features relevant to the fault and suppressing noise interference from irrelevant channels. Ablation experiments show that the triple joint loss significantly improves diagnostic accuracy compared to single or dual loss combinations.

[0051] 3. This invention addresses the domain-adaptive diagnostic network problem caused by varying operating conditions by improving the network structure and training strategy. At the network structure level, DSBN is introduced to maintain independent normalized statistics for the source and target domains while sharing convolutional layer parameters. This fundamentally decouples domain difference information from fault feature information, resolving the intra-class dispersion, boundary aliasing, and negative transfer problems caused by global alignment in existing domain adaptive methods (CORAL, DANN, MMD). At the training strategy level, a two-stage training strategy is designed, with a dynamic weight strategy employed in the second stage. The domain adversarial loss weights smoothly increase from 0 to 1.0, while the high-confidence pseudo-label loss weights linearly increase from 0.10 to 0.25, achieving a gradual transfer from global alignment to class alignment. Ablation experiments show that embedding the center loss in pseudo-label learning effectively suppresses negative transfer and further improves diagnostic accuracy.

[0052] 4. The motor bearing fault diagnosis method provided by this invention only requires a clean vibration signal under a single no-load condition as a supervision benchmark during the training phase of the denoising network. It does not require clean data covering all operating conditions and loads, thus aligning with actual industrial application scenarios (such as new equipment commissioning or regular maintenance where only a single operating condition benchmark signal can be obtained). Furthermore, this invention optimizes the balance coefficients of the joint loss function using a grid search method and provides optimized engineering parameters such as a high-confidence pseudo-label threshold and strong mixed noise construction parameters, facilitating direct application. In addition, the subsequent two-stage domain adaptive network effectively compensates for differences in signal amplitude and distribution caused by different load conditions, maintaining high diagnostic accuracy under variable load and strong noise scenarios. Attached Figure Description

[0053] The accompanying drawings are incorporated in and form part of this specification, and together with the description serve to explain the principles of the invention.

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 The flowchart shows the improved SE-CAE denoising and two-stage domain adaptive motor bearing fault diagnosis method of this invention.

[0056] Figure 2 This is a schematic diagram of the improved SE-CAE network structure in this invention;

[0057] Figure 3 This is a schematic diagram illustrating the two-stage training strategy used in this invention to train the diagnostic network.

[0058] Figure 4 The confusion matrix diagram for the diagnostic results of Comparative Example 1;

[0059] Figure 5 The confusion matrix diagram for the diagnostic results of Comparative Example 2;

[0060] Figure 6 The confusion matrix diagram for the diagnostic results of Comparative Example 3;

[0061] Figure 7 This is a confusion matrix diagram of diagnostic results in an embodiment of the present invention;

[0062] Figure 8 Confusion matrix diagram of diagnostic results for Comparative Example 4;

[0063] Figure 9 A T-SNE visualization of the feature distribution of Comparative Example 1;

[0064] Figure 10 A T-SNE visualization of the feature distribution of Comparative Example 2;

[0065] Figure 11 A T-SNE visualization of the feature distribution in Comparative Example 3;

[0066] Figure 12 This is a T-SNE visualization of the feature distribution according to an embodiment of the present invention;

[0067] Figure 13 The T-SNE visualization of the feature distribution in Comparative Example 4. Detailed Implementation

[0068] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples consistent with some aspects of the invention as detailed in the appended claims.

[0069] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0070] Example

[0071] This invention provides an improved SE-CAE denoising and two-stage domain adaptive method for motor bearing fault diagnosis, applicable to motor bearing fault monitoring and diagnosis. This method can achieve high-accuracy bearing fault identification in complex scenarios where strong mixed noise and variable operating condition domain drift coexist.

[0072] Please see Figure 1 , Figure 1 This is a flowchart of the improved SE-CAE denoising and two-stage domain adaptive motor bearing fault diagnosis method of the present invention. The method specifically includes the following steps:

[0073] Step 1, Noise Injection: Obtain the original vibration signal of the motor bearing, and superimpose noise on the original vibration signal to construct a strong mixed noise sample.

[0074] Specifically, vibration signals of motor bearings under different health conditions and operating conditions are collected to obtain raw one-dimensional time-domain data. The raw vibration signals are then segmented and standardized by Z-score sample by sample. In this embodiment, the raw vibration signals are preferably sliced ​​into overlapping slices of 1024 samples to obtain standardized raw samples.

[0075] To simulate a strongly mixed noise environment in an industrial setting, Gaussian white noise and random impulse noise are superimposed on the original sample to form a strongly mixed noise sample. Gaussian white noise is used to simulate sensor noise and environmental interference, with a preferred signal-to-noise ratio range of -8dB to -3dB. Random impulse noise is used to simulate non-Gaussian interference such as load changes and transient impacts, with a preferred sparsity range of 1% to 3% and an preferred amplitude range of 4 to 7 times the signal standard deviation.

[0076] Step 2, Denoising and Enhancement Processing: Construct an improved Squeeze-and-Excitation Convolutional Autoencoder (SE-CAE) network and train it using a joint loss function. Use the trained SE-CAE network to perform denoising and enhancement processing on the above-mentioned strongly mixed noise samples to obtain reconstructed samples.

[0077] Please see Figure 2 , Figure 2 This is a schematic diagram of the improved SE-CAE network structure in this invention. The SE-CAE network consists of an encoder, a decoder, and a channel attention mechanism embedded in the encoder.

[0078] Specifically, the encoder consists of multiple one-dimensional convolutional layers used to extract multi-scale features from strongly mixed noise samples. In this embodiment, the number of channels in the three one-dimensional convolutional layers of the encoder are 16, 32, and 64 respectively, with a kernel size of 9 and a stride of 2 for each layer.

[0079] The decoder is symmetrical to the encoder and consists of multiple one-dimensional deconvolutional layers. In this embodiment, the number of channels in the three one-dimensional deconvolutional layers of the decoder are 64, 32, and 16 respectively, with a kernel size of 9 and a stride of 2 in each layer, used to reconstruct the features output by the encoder into a temporal signal. A channel attention mechanism (Squeeze-and-Excitation, SE) is embedded after each encoding layer. This SE mechanism learns channel weights through compression and excitation operations.

[0080] First, the feature map output by the encoder is processed by a global average pooling layer. Compressed into channel statistics , of which Statistics of each channel Represented as:

[0081]

[0082] In the formula, For compression functions, For the first Feature map of each channel The length of the feature map;

[0083] Then, by learning the non-linear dependencies between channels through the sequentially connected first and second fully connected layers, normalized channel weights are generated. The normalized channel weights Represented as:

[0084]

[0085] In the formula, Let be the excitation function. These are the learnable weight parameters for the fully connected layer. , The activation functions are ReLU and Sigmoid, in that order. , The weight matrices for the first and second fully connected layers are shown in order.

[0086] Finally, the normalized channel weights are multiplied channel by channel by the feature map output by the encoder through a gating mechanism to obtain the reweighted output. , of which Output feature map after channel recalibration Represented as:

[0087]

[0088] In the formula, For recalibrating functions, For the first Normalized weights for each channel, For the first The feature maps of each channel are used to achieve feature recalibration, highlighting the key channel features related to the fault.

[0089] The reconstruction process of encoding and decoding is as follows:

[0090]

[0091]

[0092] In the formula, , , , All of these are network-learnable parameters. The activation function is ReLU. In the denoising task, the optimization objective is to minimize the reconstructed output. Compared with the original vibration signal The mean square error.

[0093] It should be noted that during the training phase of the denoising network, this embodiment divides the original vibration signal into lengths of... Slice the window to obtain There are 10 samples. The original vibration signal corresponding to each sample is denoted as _____. ,in The sample number. is the time series index. Simultaneously, Gaussian white noise and random impulse noise are superimposed on the same original vibration signal to construct a corresponding strongly mixed noise sample. This sample is input into the SE-CAE network, and the resulting denoised reconstructed signal is denoted as . , indicating the first The sample at the th The reconstructed signal at each time point.

[0094] This invention addresses strongly mixed noise scenarios by designing a joint loss function consisting of temporal reconstruction loss, sparse constraint loss, and frequency domain consistency loss, expressed as:

[0095]

[0096] In the formula, For joint losses, For temporal reconstruction loss, For sparse constraint loss, For frequency domain consistency loss, The sparse constraint loss balance coefficient, This is the frequency domain consistency loss balance coefficient;

[0097] Temporal reconstruction loss (Mean Squared Error, MSE) Used to ensure the overall fidelity of the signal, it is expressed as:

[0098]

[0099] In the formula, For the first The sample at the th The original vibration signals at each time point For the first The sample at the th Reconstructed signals at each time point This represents the total number of samples within the batch. The time sequence length of a single original vibration signal. The square of the L2 norm;

[0100] Sparsity Loss Used to highlight the sparsity of fault impacts and suppress noise redundancy, it is represented as:

[0101]

[0102] In the formula, For the first The sample at the th Reconstructed signals at each time point This represents the total number of samples within the batch. The time sequence length of a single original vibration signal. It is a norm;

[0103] Frequency Domain Consistency Loss (STFT Loss) To preserve the frequency structure of fault characteristics and avoid frequency domain distortion, it is represented as follows:

[0104]

[0105] In the formula, For the first One original vibration signal In the The STFT spectrum of the [number] time frame The amplitude of each frequency component, For the first One reconstructed signal In the The STFT spectrum of the [number] time frame The amplitude of each frequency component, The total number of frequency points. For time frames, It is the square of the L2 norm.

[0106] To facilitate understanding of the joint loss function proposed in this invention, this embodiment selects mini-batch samples and provides the calculation process for the temporal reconstruction loss:

[0107] Select small batch: batch The time sequence length (number of sampling points) of a single original vibration signal. The data was normalized to the 0~1 range using MinMax, which meets the requirements for preprocessing.

[0108] Original vibration signal: , ;

[0109] SE-CAE network reconstruction output signal: , ;

[0110] Substitute into the formula for calculating temporal reconstruction loss

[0111] ,

[0112] The temporal reconstruction loss for this batch was calculated. .

[0113] In this embodiment, the weight of the temporal reconstruction loss is set to 1 as a basic constraint, and the sparse constraint loss balance coefficient is used. and frequency domain consistency loss balance coefficient All were determined using the grid search method. Among them, the sparse constraint loss balance coefficient... The value ranges from 0.005 to 0.05, which is the frequency domain consistency loss balance coefficient. The value range is from 0.0005 to 0.005, preferably... , .

[0114] It is important to note that in this embodiment, during the SE-CAE network training phase, only the original vibration signal (clean vibration signal) under no-load (0HP) conditions is used as the supervision benchmark. This setting aligns with industrial realities: in actual motor operation and maintenance, it is typically only possible to obtain a weakly interfering original benchmark signal under a single no-load condition in specific scenarios such as new equipment commissioning and regular maintenance, making it difficult to cover clean data under all operating conditions and loads. This invention, through temporal reconstruction, sparse constraints, and frequency domain consistency joint loss optimization of the SE-CAE network, can learn general denoising characteristics under strong mixed noise under the supervision of a single benchmark signal, effectively compensating for differences in signal amplitude and distribution caused by different load conditions.

[0115] After the improved SE-CAE network is trained, the strongly mixed noise samples constructed in step 1 are denoised and enhanced to output reconstructed samples, which are used as inputs for subsequent domain adaptive diagnostic networks.

[0116] Step 3: Data partitioning: The reconstructed samples are divided into labeled source domain data and unlabeled target domain data according to the working conditions, and the target domain data is further divided into target domain training set and target domain test set according to a preset ratio.

[0117] Specifically, the denoised and enhanced samples are divided into a source domain and a target domain according to the operating conditions (different loads and speeds). Source domain data. ,in For the first The reconstructed signal after denoising of the source domain samples using an SE-CAE network For the first The true fault category labels of each source domain sample, including fault location (generally inner ring, outer ring, and ball bearings for rolling bearings) and fault severity; target domain data. It only contains the denoised vibration signal and needs to determine the health and fault status. The target domain data is divided into a target domain training set and a target domain test set according to a preset ratio (e.g., 8:2 in this embodiment), and sample-level normalization is performed to unify the amplitude distribution.

[0118] Step 4, Domain Adaptive Training: Construct a domain adaptive diagnostic network with ResNet as the backbone network and introduce Domain-Specific Batch Normalization (DSBN), and train the diagnostic network using a two-stage training strategy.

[0119] Please see Figure 3 , Figure 3 This is a schematic diagram illustrating the two-stage training strategy used in this invention to train the diagnostic network.

[0120] First, a fault diagnosis network with ResNet as the backbone is constructed, and DSBN is introduced. DSBN maintains independent normalized statistics for the source and target domains respectively, while sharing convolutional layer parameters, thereby decoupling domain difference information from fault feature information from the root cause.

[0121]

[0122] In the formula, The output features are those after domain-specific batch normalization. The features to be normalized are the outputs of the convolutional layer. , The values ​​are, in order, the mean and standard deviation of the source domain. , The parameters are, in order, the scaling and offset parameters of the source domain. , The values ​​are, in order, the mean and standard deviation of the target domain. , The scaling and offset parameters for the target domain are listed in sequence. It should be noted that the domain-specific batch normalized convolutional layer parameters are shared between the source and target domains, while the normalization statistics are independent for each domain, thus decoupling domain differences from fault characteristics.

[0123] Then, the diagnostic network is trained using the following two-stage training strategy:

[0124] Phase 1, Source Domain Supervised Pre-training: The diagnostic network is pre-trained using labeled data from the source domain, jointly optimizing cross-entropy loss and center loss to make similar features compact and dissimilar features clearly defined.

[0125]

[0126] In the formula, This represents the total loss for the first phase. The cross-entropy loss is the source domain. The central loss of the source domain, The center loss weight has a value range of 0.02 to 0.06, and is preferably 0.04 in this embodiment;

[0127] Cross-entropy loss This ensures basic classification accuracy, which is expressed as:

[0128]

[0129] In the formula, For the first The reconstructed signal belonging to the source domain is the signal that has been denoised by the SE-CAE network and then divided into the source domain in step 3. For the first The actual fault category label corresponding to each reconstructed signal; To diagnose the network, the signal will be reconstructed. Predicted as the true category The probability of; This represents the total number of samples in the source domain training batch.

[0130] Central loss This is used to force similar samples to cluster together in the feature space, enhancing feature separability. It is represented as:

[0131]

[0132] In the formula, For the first The feature vector of the reconstructed signal is extracted by the diagnostic network; For the corresponding category The feature centers are normalized features output by domain-specific batch normalization. After processing by subsequent network layers, the deep feature vector of the sample is finally obtained through global average pooling; It is the square of the L2 norm.

[0133] After the first phase of training is completed, the pre-trained weights are obtained.

[0134] To facilitate understanding of the present invention, this embodiment uses the following small batch as an example to illustrate the calculation process of the cross-entropy loss in the source domain:

[0135] If single batch There are six categories of faults.

[0136] Sample unique hot tag: , , ;

[0137] Model predicted probabilities: , , ;

[0138] Substitute the cross-entropy loss formula into the source domain The cross-entropy loss of the source domain in this batch is calculated. .

[0139] The second stage, unsupervised domain adaptive training: The pre-trained weights obtained in the first stage are loaded as the initial parameters of the diagnostic network. Then, the diagnostic network is trained unsupervised domain adaptively using source domain data and the target domain training set. Relying solely on domain adversarial loss can easily lead to fragmented intra-class features and blurred inter-class boundaries; simultaneously, directly applying pseudo-labels to unlabeled samples in the target domain can easily introduce low-confidence mislabeling noise, causing severe negative transfer. Therefore, this invention introduces center loss and high-confidence pseudo-labels on top of domain adversarial training.

[0140] The total loss function for the second stage is expressed as:

[0141]

[0142] In the formula, This represents the total loss for the second phase. The first stage loss includes the cross-entropy loss of the source domain and the center loss of the source domain, which are used to maintain the classification ability of the source domain. To achieve domain adversarial loss, a gradient inversion layer is used to align the global domain distribution. The high-confidence pseudo-label loss is derived from the embedding center loss, and the cross-entropy loss is derived from the target domain. and the central loss of the target domain Composition, that is , , where represents the weighting coefficient of the center loss in the pseudo-label loss; Dynamic weights for domain adversarial loss. Dynamic weights are used to compensate for high-confidence pseudo-label loss.

[0143] The high-confidence pseudo-label strategy of this invention is as follows: a sample and its pseudo-label are only used for supervised training when the maximum class probability predicted by the diagnostic network for the target domain sample is higher than a preset threshold. In this embodiment, the pseudo-label confidence threshold ranges from 0.85 to 0.95. This is because when the threshold is below 0.85, the pseudo-label noise is too large, affecting the diagnostic accuracy; when the threshold is above 0.95, there are too few usable pseudo-label samples, making it difficult for the network model to converge sufficiently. Therefore, 0.9 is preferred. Simultaneously, a dynamic weighting strategy is adopted: dynamic weighting of domain adversarial loss. During training, a standard Sigmoid-type increasing curve is used to smoothly increase the weights from 0 to 1.0, achieving gradual transfer from early global alignment to later class alignment; a high-confidence pseudo-label loss is used for dynamic weights. During training, a standard linear increment curve is used, increasing linearly from 0.10 to 0.25, which only applies to high-confidence target domain samples. The weighting coefficients of the center loss in the pseudo-label loss are... Setting it to 0.08 is used to enhance the compactness of intra-class features in the target domain and avoid negative transfer.

[0144] Step 5: Output the results: Input the target domain test set into the diagnostic network after the two-stage training is completed, and output the fault diagnosis results of the motor bearing.

[0145] Experimental verification

[0146] To verify the efficacy of the present invention, the following specific examples are used to demonstrate its performance, and ablation experiments are presented for comparison. The specific process is as follows:

[0147] 1) Data Acquisition: The test bearings are pre-configured with fault types and severity. Under certain conditions (including but not limited to speed, load, motor model, sampling frequency, and bearing specifications), vibration acceleration signals from the bearings at the motor drive end and fan end are collected using an accelerometer. In this example, six types of data from the Paderborn (PU) dataset are used to test the network's reliability. The data is shown in Table 1 below:

[0148] Table 1 Bearing Vibration Data

[0149]

[0150] 2) Noise injection: Gaussian white noise with a signal-to-noise ratio of approximately -5dB and random impulse noise with a sparsity of approximately 1.5% and an amplitude of 5.5 times the standard deviation of the signal are injected. After superposition, the overall input signal-to-noise ratio is approximately -7.13dB.

[0151] 3) Data Partitioning: Four operating conditions were selected. The source domain consisted of operating conditions of 900 rpm, 0.7 Nm torque, and 1000 N radial load, and 1500 rpm, 0.1 Nm torque, and 1000 N radial load. The target domain consisted of operating conditions of 1500 rpm, 0.7 Nm torque, and 400 N radial load, and 1500 rpm, 0.7 Nm torque, and 1000 N radial load. The target domain data was divided into training and testing sets in an 8:2 ratio.

[0152] 4) Training configuration: The AdamW optimizer was used, with an initial learning rate of 0.001 and a weight decay coefficient of 0.0005. During the denoising stage, the batch size was 32, and training lasted for 120 epochs. In the domain adaptation network stage, the first and second stages were each trained for 200 epochs, with the pseudo-label confidence threshold set to 0.9 and the dynamic weights gradually increasing.

[0153] 5) Performance testing and ablation experiments.

[0154] (I) SE-CAE Joint Loss Function Ablation Experiment

[0155] To verify the effectiveness and superiority of the improved joint loss function of the SE-CAE network proposed in step 2 of this invention, ablation comparison experiments were conducted under identical experimental conditions. All experiments used the following training parameters: batch size of 32, learning rate fixed at 1e-3, and maintained complete consistency in the SE-CAE network and its subsequent two-stage adaptive network structure, dataset, task division, and classifier structure, only changing the combination of the loss function. Specific settings are as follows:

[0156] Comparative Example 1: Compared with the example, the only difference is that the joint loss function during SE-CAE network training only includes the temporal reconstruction loss, that is, it does not include the sparsity constraint loss and the frequency domain consistency loss. All other experimental conditions are exactly the same.

[0157] Comparative Example 2: Compared with the example, the only difference is that the joint loss function during SE-CAE network training includes temporal reconstruction loss and sparse constraint loss, that is, it does not include frequency domain consistency loss; the other experimental conditions are exactly the same.

[0158] Comparative Example 3: Compared with the example, the only difference is that the joint loss function during SE-CAE network training includes temporal reconstruction loss and frequency domain consistency loss, i.e., it does not include sparsity constraint loss; the other experimental conditions are exactly the same.

[0159] Example: In this example, the joint loss function for training the SE-CAE network includes temporal reconstruction loss, sparse constraint loss, and frequency domain consistency loss, i.e., a triple joint loss is used.

[0160] The diagnostic accuracy results are as follows:

[0161] Comparative Example 1 (using only temporal reconstruction loss): diagnostic accuracy was 95.89%;

[0162] Comparative Example 2 (using temporal reconstruction loss and sparse constraint loss): diagnostic accuracy was 96.45%;

[0163] Comparative Example 3 (using temporal reconstruction loss and frequency domain consistency loss): diagnostic accuracy was 96.67%;

[0164] In this embodiment (using temporal reconstruction loss, sparse constraint loss, and frequency domain consistency loss), the diagnostic accuracy is 97.45%.

[0165] Experimental results show that while time-domain reconstruction loss alone has a certain denoising capability, it is insufficient in preserving weak fault features. Adding sparse constraint loss to the time-domain reconstruction loss enhances the sparsity of fault impacts, improving diagnostic accuracy by 0.56%. Adding frequency-domain consistency loss protects the frequency structure of fault features, improving diagnostic accuracy by 0.78%. In this embodiment, when the three loss functions are combined into a joint loss function, the constraints form a synergistic enhancement mechanism, resulting in a 1.56% improvement in diagnostic accuracy compared to Comparative Example 1, and 1.00% and 0.78% improvement compared to Comparative Examples 2 and 3, respectively. This fully demonstrates the superiority of the proposed triple joint loss function.

[0166] Please see Figures 4 to 7 , Figures 9 to 12 ,in: Figure 4 This is a confusion matrix diagram of the diagnostic results for Comparative Example 1. Figure 5 The confusion matrix diagram for the diagnostic results of Comparative Example 2 is shown. Figure 6 The confusion matrix diagram for the diagnostic results of Comparative Example 3 is shown. Figure 7 This is a confusion matrix diagram of the diagnostic results in this embodiment. Figures 9 to 11 The following are T-SNE visualizations of the feature distributions from comparative examples 1 to 3. Figure 12 This is a T-SNE visualization of the feature distribution in this embodiment. (From...) Figure 7 and Figure 12 As can be seen, the clustering of various fault features in this embodiment is more compact and the boundaries between classes are clearer, further verifying the effectiveness of the method of the present invention.

[0167] (II) Ablation Experiment of Pseudo-Label Loss Function

[0168] To further verify the effectiveness of the diagnostic network proposed in step 4 of this invention in embedding center loss during pseudo-label learning, a control experiment was conducted while keeping all other experimental conditions completely consistent. The front-end inputs were all the results after denoising using the triple joint loss function of this embodiment. The comparison settings are as follows:

[0169] Comparative Example 4: Compared with the Example, the only difference is that the pseudo-label loss function does not include the center loss; all other experimental conditions are exactly the same.

[0170] In this example, the pseudo-label loss function includes the center loss.

[0171] Please see Figure 8 , Figure 13 as well as Figure 7 and Figure 12 ,in: Figure 8 The confusion matrix diagram for the diagnostic results of Comparative Example 4 is shown. Figure 13 A T-SNE visualization of the feature distribution in Comparative Example 4; Figure 7 This is a confusion matrix diagram of the diagnostic results in this embodiment. Figure 12 This is a T-SNE visualization of the feature distribution in this embodiment.

[0172] The diagnostic accuracy results are as follows:

[0173] Comparative Example 4 (the pseudo-label loss function does not include center loss): the diagnostic accuracy was 96.67%;

[0174] In this embodiment (where the pseudo-label loss function includes center loss), the diagnostic accuracy is 97.45%.

[0175] Experimental results show that after embedding center loss in pseudo-label learning, the accuracy increased from 96.67% to 97.45%, effectively suppressing negative transfer phenomenon, and further verifying the effectiveness of introducing center loss in pseudo-label learning in this invention.

[0176] It should be noted that the specific parameters mentioned above are used to verify the diagnostic effect of this method under extreme working conditions and are not necessary limitations on the present invention. In practical applications, the above parameters can be adaptively adjusted according to the noise characteristics, working conditions and hardware resources of the specific industrial site. For example, the signal-to-noise ratio and sparsity of the noise injection can be set according to the measured noise level on site; the number of training rounds and batch size of the denoising network can be adjusted according to the data scale and computing resources.

[0177] In summary, this invention proposes an intelligent method for diagnosing motor bearing faults based on an improved SE-CAE network and a domain adaptive diagnostic network fused with DSBN. Compared to existing signal processing-based denoising methods, this invention effectively suppresses background interference and retains weak fault impact features under strong mixed noise conditions through joint loss optimization of the SE-CAE network, providing high-quality input for backend diagnosis. Simultaneously, a domain adaptive diagnostic network composed of ResNet and DSBN is employed, using a two-stage training strategy (first stage source domain pre-training, second stage domain adaptive fine-tuning): first, compact intra-class feature learning is completed in the source domain; then, class-level alignment is achieved through collaborative optimization of domain adversarial alignment, center loss, and high-confidence pseudo-labels, fundamentally decoupling domain differences and avoiding intra-class dispersion and negative transfer. The diagnostic model constructed in this invention achieves high fault diagnosis accuracy even in complex scenarios with strong mixed noise and variable operating conditions, demonstrating significant engineering application value.

[0178] The above description is merely a specific embodiment of the present invention, intended to enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments without creative effort will be readily apparent to those skilled in the art; the basic principles defined herein can be applied to other embodiments without departing from the spirit and scope of the present invention.

[0179] It should be understood that this invention is not limited to the specific content described above, and various equivalent modifications or substitutions can be made without departing from its scope of protection. The scope of protection of this invention is defined only by the appended claims.

Claims

1. An improved SE-CAE denoising and two-stage domain adaptive method for motor bearing fault diagnosis, comprising the following steps: Step 1, Noise Injection: Obtain the original vibration signal of the motor bearing, and superimpose noise onto the original vibration signal to construct a strong mixed noise sample; Step 2, Denoising and Enhancement Processing: Construct an improved SE-CAE network, which consists of an encoder, a decoder, and a channel attention mechanism embedded in the encoder. The network is trained using a joint loss function consisting of temporal reconstruction loss, sparse constraint loss, and frequency domain consistency loss. The trained SE-CAE network is then used to denoise and enhance the strongly mixed noise samples to obtain reconstructed samples. Step 3, Data partitioning: The reconstructed samples are divided into labeled source domain data and unlabeled target domain data according to the working conditions, and the target domain data is further divided into target domain training set and target domain test set according to a preset ratio; Step 4, Domain Adaptive Training: Construct a domain adaptive diagnostic network with ResNet as the backbone and introduce domain-specific batch normalization, and train the diagnostic network using a two-stage training strategy, wherein: Phase 1: Supervised pre-training of the diagnostic network using the source domain data, with optimization objectives including cross-entropy loss and center loss, to obtain pre-trained weights; The second stage involves loading the pre-trained weights as initial parameters for the diagnostic network, and then using the source domain data and the target domain training set to perform unsupervised domain adaptive training on the diagnostic network. The optimization objectives include the cross-entropy loss and center loss of the source domain, the domain adversarial loss, and the high-confidence pseudo-label loss of the embedding center loss. Step 5: Output the results: Input the target domain test set into the diagnostic network after the two-stage training is completed, and output the fault diagnosis results of the motor bearing.

2. The improved SE-CAE denoising and two-stage domain adaptive motor bearing fault diagnosis method according to claim 1, characterized in that, In step 1, the strong mixed noise sample is constructed by superimposing Gaussian white noise and random impulse noise onto the original vibration signal; The signal-to-noise ratio of the Gaussian white noise ranges from -8dB to -3dB; the sparsity of the random impulse noise ranges from 1% to 3%, and the amplitude ranges from 4 to 7 times the signal standard deviation.

3. The improved SE-CAE denoising and two-stage domain adaptive motor bearing fault diagnosis method according to claim 1, characterized in that, In step 2, the encoder consists of multiple one-dimensional convolutional layers, used to extract multi-scale features of strongly mixed noise samples; The decoder is symmetrical to the encoder and consists of multiple one-dimensional deconvolutional layers, used to reconstruct the features output by the encoder into a time-domain signal. The channel attention mechanism is embedded in each layer of the encoder, including: a global average pooling layer, which is used to compress the feature maps output by each layer of the encoder into channel statistics; The first fully connected layer and the second fully connected layer are connected in sequence to learn the nonlinear dependencies between channels and generate normalized channel weights; and a gating mechanism is used to multiply the normalized channel weights with the feature map output by the encoder channel by channel to obtain a reweighted output, thereby realizing feature recalibration.

4. The improved SE-CAE denoising and two-stage domain adaptive motor bearing fault diagnosis method according to claim 1, characterized in that, In step 2, the joint loss function is expressed as: The temporal reconstruction loss is expressed as: The sparse constraint loss is expressed as: The frequency domain consistency loss is expressed as: In the formula, For joint losses; For the first The sample at the th The original vibration signals at each time point For the first The sample at the th Reconstructed signals at each time point The total number of samples within the batch. The time sequence length of a single original vibration signal. For the first One original vibration signal In the The STFT spectrum of the [number] time frame The amplitude of each frequency component, For the first One reconstructed signal In the The STFT spectrum of the [number] time frame The amplitude of each frequency component, The total number of frequency points. For time frames, For a norm, The square of the L2 norm. The sparse constraint loss balance coefficient, This is the frequency domain consistency loss balance coefficient.

5. The improved SE-CAE denoising and two-stage domain adaptive motor bearing fault diagnosis method according to claim 4, characterized in that, The sparse constraint loss balance coefficient and frequency domain consistency loss balance coefficient All were determined using the grid search method; Wherein, the sparse constraint loss balance coefficient The value range is from 0.005 to 0.05, and the frequency domain consistency loss balance coefficient is... The value range is from 0.0005 to 0.

005.

6. The improved SE-CAE denoising and two-stage domain adaptive motor bearing fault diagnosis method according to claim 1, characterized in that, In step 4, the domain-specific batch normalization is expressed as: In the formula, The output features are those after domain-specific batch normalization. The features to be normalized are the outputs of the convolutional layer. , The values ​​are, in order, the mean and standard deviation of the source domain. , The parameters are, in order, the scaling and offset parameters of the source domain. , The values ​​are, in order, the mean and standard deviation of the target domain. , The scaling and offset parameters for the target domain are listed in order.

7. The improved SE-CAE denoising and two-stage domain adaptive motor bearing fault diagnosis method according to claim 1, characterized in that, In the first stage of step 4, the joint optimization objective of the cross-entropy loss and the center loss is expressed as: The cross-entropy loss of the source domain is expressed as: The central loss of the source domain is expressed as: In the formula, This represents the total loss for the first phase. For the first The reconstructed signal belonging to the source domain is the signal that has been denoised by the SE-CAE network and then divided into the source domain in step 3. For the first The actual fault category label corresponding to each reconstructed signal; To diagnose the network, the signal will be reconstructed. Predicted as the true category The probability of; This represents the total number of samples in the source domain training batch. For the first The feature vector of the reconstructed signal is extracted by the diagnostic network; For the corresponding category The feature center; The square of the L2 norm; The center loss weight has a value range of 0.02 to 0.

06.

8. The improved SE-CAE denoising and two-stage domain adaptive motor bearing fault diagnosis method according to claim 7, characterized in that, In the second stage of step 4, the total loss function is expressed as: In the formula, This represents the total loss for the second phase. The total loss for the first stage includes the cross-entropy loss of the source domain and the center loss of the source domain; To achieve domain adversarial loss, a gradient inversion layer is used to align the global domain distribution. The high-confidence pseudo-label loss incorporates center loss, and the cross-entropy loss is derived from the target domain. and the central loss of the target domain Composition, that is , , where represents the weighting coefficient of the center loss in the pseudo-label loss; Dynamic weights for domain adversarial loss; Dynamic weights are used to compensate for high-confidence pseudo-label loss.

9. The improved SE-CAE denoising and two-stage domain adaptive motor bearing fault diagnosis method according to claim 8, characterized in that, The domain adversarial loss dynamic weight The weights are smoothly increased from 0 to 1.0 during training; the high-confidence pseudo-label loss has dynamic weights. The value was linearly increased from 0.10 to 0.25 during training.

Images