Bearing fault diagnosis method based on multi-scale frequency sensing dynamic enhancement

Abstract

The invention discloses a bearing fault diagnosis method based on multi-scale frequency sensing dynamic enhancement, and the method comprises the steps: collecting a vibration signal in a bearing operation state, carrying out the preprocessing of the vibration signal, obtaining a time-frequency matrix, and dividing the time-frequency matrix into a training set and a test set; building a multi-scale frequency network sensing model, inputting a time-frequency matrix in a training set into the model to realize extraction of multi-scale features, then performing pooling, time sequence compression, flattening and dimension reduction on the extracted multi-scale features, and then outputting fault category probability distribution through a classifier; and finally, testing the trained model by using a test set, and calculating evaluation indexes such as accuracy, a confusion matrix, an ROC curve and the like. The method has high accuracy while keeping light weight, breaks through double limitations of fixed frequency band and sensitive rotating speed of a traditional method, and can provide a high-precision and low-cost light-weight solution for engineering application of variable-rotating-speed mechanical fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a bearing fault diagnosis method based on multi-scale frequency-sensing dynamic enhancement. Background Art

[0002] Rolling bearings are core functional parts of rotating mechanical systems. Their working status directly affects the working efficiency and production safety of the mechanical system. Therefore, their fault diagnosis is crucial to the maintenance of industrial equipment. On the one hand, bearings under actual working conditions are often in dynamic conditions such as sudden speed changes and load disturbances, which makes the vibration signal show significant time-varying non-Gaussian characteristics and multi-scale coupling characteristics. With the rise of machine learning technology, deep learning methods can save the tedious steps of manual feature extraction and do not require expert experience. In response to the time-varying characteristics of vibration signals, the current research focus of rolling bearing fault diagnosis has gradually shifted to data-driven methods based on deep learning. This method circumvents the technical barriers of manual feature engineering through an end-to-end architecture and can complete fault diagnosis. It is a representative method for intelligent extraction of eigenvalues. Therefore, deep learning is used in the field of variable speed rolling bearing fault diagnosis and has developed rapidly.

[0003] Currently, common deep learning methods focus on convolutional neural networks (CNNs), deep belief networks (DBNs), and autoencoders. CNNs, leveraging their local receptive field properties, enable hierarchical decoupling of fault-sensitive features, while DBNs construct high-level fault representations through layer-by-layer feature abstraction. Dong et al. (DONG Y, JIANG H, YAOR, et al. Rolling bearing intelligent fault diagnosis towards variable speed and imbalanced samples using multiscale dynamic supervised contrast learning [J]. Reliability Engineering & System Safety, 2024, 243.) proposed a multi-scale adaptive feature extraction network to enhance feature discrimination ability, but they did not establish a performance comparison with the contrastive learning paradigm, and the superiority of this method still needs to be verified; Zheng Xinhui et al. (Zheng Xinhui, Ma Chao, Wang Shaohong, et al. Rolling bearing fault diagnosis method using EEMD and deep belief network [J]. Journal of Chongqing University of Technology (Natural Science), 2024, 38(06): 133-40.) innovatively integrated ensemble empirical mode decomposition (EEMD) and DBN to achieve damage degree perception diagnosis, but did not prove the contribution of each module through ablation experiments, which led to doubts about the interpretability of the model; Zhao et al. (Zhao D, Liu S, Du H, et al. Deep branch attention networkand extreme multi-scale entropy based single vibration signal-driven variable speed fault diagnosis scheme for rolling bearing[J].Advanced Engineering Informatics, 2023, 55:101844.) proposed a single vibration signal-driven scheme based on extreme multi-scale entropy (EMSEn), developed a deep branch attention network (DBANet), and flexibly integrated information through dual parallel channels and attention mechanism. The experimental results verified its effectiveness and performance superior to other multi-scale entropy methods; Wang et al. (Wang R, Jiang H, Zhu K, et al.A deep feature enhanced reinforcement learning method for rolling bearing fault diagnosis[J]. Advanced Engineering Informatics, 2022, 54:101750.) innovatively integrated the neural network improved by Elu activation function, attention model and deep Q network to realize rolling bearing composite fault diagnosis. Experiments have verified that it is superior to other intelligent diagnosis methods. Traditional deconvolution methods are limited by filter length selection and random interference. Although CSF is excellent, it still has limitations. Therefore, Miao et al. (Miao Y, LiC, Shi H, et al. Deep network-based maximum correlated kurtosis deconvolution: Anovel deep deconvolution for bearing fault diagnosis[J]. Mechanical Systems and Signal Processing, 2023, 189: 110110) proposed maximum correlation kurtosis deconvolution based on deep network. Through optimization such as Hanning window initialization and correlation kurtosis cost function, experiments show that it is superior to CSF ​​and improved MCKD. Ma et al. (Ma Y, Yang J, Li L. Collaborative and adversarial deep transfer auto-encoder for intelligent fault diagnosis[J]. Neurocomputing, 2022, 486: 1-15.) proposed a collaborative adversarial transfer model based on convolutional autoencoders (CADTA). This model promotes category discrimination and domain similarity learning through a multi-task classifier, duplex adversarial learning, and phased training. While this approach improves diagnostic automation, the inertia of heuristic feature design makes it difficult to meet the needs of cross-operating-condition diagnosis. While all of the above-mentioned papers have used different deep learning methods to complete fault diagnosis under variable operating conditions and achieved good results, they still suffer from long model training times and slow inference speeds, ignoring the deep learning framework's significant reliance on a large number of hardware devices and intensive computing processes.

[0004] At present, there are problems such as the difficulty in extracting features due to the non-stationarity of rolling bearing fault signals under variable speed conditions, and the model complexity and slow inference speed caused by the large number of parameters in traditional deep learning. In recent years, lightweight requirements have been put forward for industrialization, and the data-driven methods using deep learning often require a large amount of computing resources and memory, resulting in problems such as difficulty in deploying models to edge devices. Therefore, the development of a diagnostic model with high generalization ability, strong robustness and lightweight has become a key breakthrough in solving the needs of fault feature extraction, cross-operating condition adaptation and edge deployment in variable speed scenarios. Summary of the Invention

[0005] In response to the difficulties in feature extraction caused by the non-stationarity of rolling bearing fault signals under variable speed conditions, as well as the model complexity and slow inference speed caused by the large number of parameters in traditional deep learning, the purpose of the present invention is to provide a bearing fault diagnosis method based on multi-scale frequency-sensitive dynamic enhancement. The vibration signal is converted into a time-frequency matrix as input through short-time Fourier transform (STFT), and a lightweight network architecture combining deep separable convolution and dilated convolution is constructed to achieve collaborative capture of local details and global patterns of multi-scale frequency features. A channel attention mechanism is introduced to adaptively and dynamically enhance key frequency bands, and time dimension compression is combined to eliminate speed fluctuation interference. The present invention has high accuracy while maintaining lightweight. This framework breaks through the dual limitations of traditional methods of fixed frequency band and speed sensitivity, and can provide a high-precision, low-cost lightweight solution for the engineering application of variable speed machinery fault diagnosis.

[0006] The technical solution adopted by the present invention is: A bearing fault diagnosis method based on multi-scale frequency sensing dynamic enhancement includes the following steps: S1. Collecting vibration signals of a bearing in operation, preprocessing the vibration signals to obtain a time-frequency matrix, and then dividing the time-frequency matrix into a training set and a test set, wherein the preprocessing includes downsampling, normalization, and short-time Fourier transform; S2. Build a multi-scale frequency network perception model, which consists of a multi-scale frequency perception module, a dynamic frequency band enhancement module, a time domain invariance compression module, and an output layer. Input the time-frequency matrix in the training set of step S1 into the multi-scale frequency network perception model, use the multi-scale frequency perception module to extract multi-scale features, recalibrate the multi-scale features through the dynamic frequency band enhancement module, and perform time series compression through the time domain invariance compression module to obtain time series compression features. After flattening and dimensionality reduction in the output layer, the time series compression features are used by the classifier to generate a probability distribution of fault categories. S3, determining whether the number of training times M of the multi-scale frequency network perception model reaches the preset number of iterations N, if so, proceeding to the next step; otherwise, repeating step S2; S4. Use the test set to test the trained model, and introduce accuracy, confusion matrix, ROC curve, and PRC curve to evaluate model performance.

[0007] Preferably, in step S1, the specific steps of the preprocessing are: down-sampling the vibration signal using an FIR low-pass filter, then normalizing the down-sampled signal to the [-1, 1] interval, and then performing a short-time Fourier transform to convert it into a time-frequency matrix.

[0008] Preferably, in step S1, the time-frequency matrix is ​​divided into a training set and a test set in a ratio of 8:2.

[0009] Preferably, in step S2, the multi-scale frequency perception module adopts a cascade structure of depthwise separable convolution and dilated convolution to extract the features of the training set, that is, the depthwise separable convolution extracts the local frequency features of the time-frequency matrix, and the dilated convolution expands the receptive field to extract the global frequency features, thereby extracting multi-scale features; The depthwise separable convolution performs independent spatial filtering operations on each channel of the input time-frequency matrix to obtain the local spatial features of each channel. The response calculation of the input channels is used to output the local spatial features. The formula is:

[0010] Where, is the local spatial feature of the output, is the bias term of the corresponding input channel in the depth convolution stage, is the convolution kernel size; and Represents the spatial position index within the convolution kernel, , , The corresponding input channel in the spatial position The weight, Indicates the The convolution kernel of the channel is at position The weight of Indicates the first Channels are located at the convolution window position Pixel value of Use dilated convolution to perform nonlinear fusion of multi-scale context information on local spatial features to obtain multi-scale features; for the input local spatial feature position , the output response is formalized as:

[0011] Where, The multi-scale features of the output are at position The value of is the convolution kernel weight, Represented as an interval The sampled input value.

[0012] Preferably, in step S2, the dynamic frequency band enhancement module generates a channel weight vector from the multi-scale feature through a channel attention mechanism, and then multiplies the channel weight vector with the multi-scale feature output by the multi-scale frequency perception module channel by channel to enhance the important channel and realize feature recalibration; the channel attention mechanism includes two modules: feature compression Squeeze and weight generation Excitation. The feature compression Squeeze module compresses the spatial dimension of the multi-scale feature through global average pooling to generate channel-level statistics. Assume that the input multi-scale feature , then the compressed channel vector Calculated as:

[0013] Where, Indicates the channel vectors, is the number of channels, is the spatial dimension, Represents the input multi-scale feature Channels in space The value of The channel weight vector is calculated using the weight generation Excitation module , the formula is:

[0014] Where, , , and is a learnable parameter, is the dimensionality reduction ratio, is the Sigmoid function, is the ReLU activation function; Finally, the channel weight vector Multiply the multi-scale features output by the multi-scale frequency perception module channel by channel to obtain the enhanced features. The formula is:

[0015] Where, represents the enhanced features of the output, is the feature scaling function, represents channel-level multiplication, For the The weight of the channel, Multi-scale feature aisle.

[0016] Preferably, in step S2, the time domain invariance compression module uses a time series compression module to perform time series compression to extract time domain invariant features, to retain the main frequency component of the vibration signal, and to eliminate speed fluctuation interference. The formula for time series compression is:

[0017] Where, Is the time point after time series compression The characteristic value of is the window size; Subsequently, the time axis receptive field is expanded by using dilated convolution and applied to the time series compression module, covering 1.8 times the original period range. The formula is:

[0018] Where, For time point The output eigenvalues ​​of Is the convolution kernel at position The weight parameter, Indicates the input signal at time point The value of is the void ratio, is the convolution kernel width; Finally, the time dimension is compressed by mean pooling to retain the statistical characteristics between channels. The formula is expressed as:

[0019] Where, For the The temporal compression characteristics of the channel compression output, Is the input feature in the channel , time point The value of is the original time step, .

[0020] Preferably, in step S2, in the output layer, the time series compression feature is flattened into a one-dimensional feature vector by the flattening layer, and then input into the fully connected layer, and 16 ReLU neurons are used to perform feature dimensionality reduction to obtain a 16-dimensional feature vector, and then the 16-dimensional feature vector is input into the classifier, and the fault category probability distribution is generated through 4 Softmax neurons.

[0021] Compared with the prior art, the present invention has the following beneficial effects: The present invention has high accuracy while maintaining lightweight. The framework of the present invention breaks through the dual limitations of traditional methods of fixed frequency band and speed sensitivity, and can provide a high-precision, low-cost lightweight solution for the engineering application of variable speed mechanical fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of a bearing fault diagnosis method based on multi-scale frequency sensing dynamic enhancement according to the present invention; Figure 2 This is the structure diagram of the channel attention mechanism of the present invention; Figure 3 This is a structural diagram of the multi-scale frequency network perception model (MSFAN-DFE) of the present invention; Figure 4 This is a flow chart of the dynamic frequency band attention mechanism of the present invention; Figure 5 This is a diagram of the bearing experimental platform of the present invention; Figure 6 This is a curve diagram of the accuracy of the training set and the test set of the present invention; Figure 7 This is a curve diagram of the loss value of the training set and the test set of the present invention; Figure 8 It is a T-SNE visualization diagram of the feature space; Figure 9 This is the confusion matrix diagram of the test set data; Figure 10 It is a multi-category ROC curve diagram; Figure 11 It is the precision-recall curve; Figure 12 This is the frequency response attention weight distribution diagram. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of this application to further clearly and completely describe the technical solutions in the embodiments of this application. It should be noted that the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making any creative work shall fall within the scope of protection of this application.

[0024] In order to make the invention purpose, technical solutions and advantages of this application clearer, the embodiments of this application are further described in detail in conjunction with the drawings of the specification: In order to more clearly understand the above-mentioned purposes, features and advantages of the present invention, the advantages of the present invention will be further illustrated by comparing the embodiments in conjunction with the drawings and specific implementation methods.

[0025] Example 1

[0026] This embodiment provides a bearing fault diagnosis method based on multi-scale frequency sensing dynamic enhancement, including: S1. Collecting vibration signals of the bearing under running state, preprocessing the vibration signals to obtain a time-frequency matrix, and then dividing the time-frequency matrix into a training set and a test set in a ratio of 8:2. The preprocessing includes downsampling, normalization, and short-time Fourier transform. The specific steps of the preprocessing are: downsampling the vibration signal using an FIR low-pass filter, then normalizing the downsampled signal to the [-1, 1] interval, and then performing a short-time Fourier transform to convert it into a time-frequency matrix; S2. Build a multi-scale frequency network perception model, which consists of a multi-scale frequency perception module, a dynamic frequency band enhancement module, a time domain invariance compression module and an output layer; input the time-frequency matrix in the training set of step S1 into the multi-scale frequency network perception model, use the multi-scale frequency perception module to extract multi-scale features, recalibrate the multi-scale features through the dynamic frequency band enhancement module, and perform time series compression through the time domain invariance compression module to obtain time series compression features; the time series compression features are flattened into a one-dimensional feature vector through the flattening layer, and then input into the fully connected layer, and 16 ReLU neurons are used to perform feature dimensionality reduction to obtain a 16-dimensional feature vector, and then the 16-dimensional feature vector is input into the classifier, and the fault category probability distribution is generated through 4 Softmax neurons; The multi-scale frequency perception module uses a cascade structure of depthwise separable convolution and dilated convolution to extract the features of the training set, that is, the depthwise separable convolution extracts the local frequency features of the time-frequency matrix, and the dilated convolution expands the receptive field to extract the global frequency features, thereby extracting multi-scale features; Depthwise Separable Convolution (DSC) achieves a balanced optimization of computational efficiency and model capacity by decoupling spatial domain feature extraction and channel domain feature fusion. Its core architecture consists of two independent stages: depthwise convolution and pointwise convolution. The depthwise convolution stage performs independent spatial filtering operations on each channel of the input time-frequency matrix, and the weight matrix dimension is ,in is the number of input channels, is the convolution kernel size, for the The response calculation of the input channels is used to output the local spatial features. The formula is:

[0027] Where, is the local spatial feature of the output, is the bias term of the corresponding input channel in the depth convolution stage, is the convolution kernel size; and Represents the spatial position index within the convolution kernel, , , The corresponding input channel in the spatial position The weight, Indicates the The convolution kernel of the channel is at position The weight of Indicates the first Channels are located at the convolution window position Pixel value of This operation is done by The filter captures the local spatial features of each channel, such as the vibration pattern recognition of surface microcracks; compared with traditional convolution, the computational complexity of depthwise separable convolution is reduced from down to , when the number of input channels Much larger than the convolution kernel area When the computational complexity can be reduced to the traditional method level; in terms of parameter size, the number of weights of depthwise separable convolution is reduced to that of traditional convolution , significantly reducing model storage and computational load; This architecture uses a spatial-channel decoupling design to enhance the model's ability to represent multi-scale signals while retaining local feature extraction capabilities. In dynamic working scenarios, the lightweight characteristics of deep separable convolution make it suitable for the real-time diagnosis needs of edge computing devices. At the same time, the subsequent model dynamically adjusts the feature response through the channel attention mechanism to suppress the impact of noise interference on classification performance. Its mathematical form and engineering practice show that this method can effectively maintain the ability to distinguish fault-sensitive features while reducing computational costs.

[0028] Dilated Convolution reconstructs the spatial sampling pattern of the convolution kernel by introducing a dilation rate, expanding the receptive field while maintaining the number of parameters, and achieving nonlinear fusion of multi-scale contextual information. Its core mechanism is to insert fixed-interval dilation units between the convolution kernel weights, so that the effective receptive field increases exponentially. Specifically, when the convolution kernel size is And the void rate is When , the effective receptive field can be extended to , compared to the standard convolution , significantly improving the ability to model long-range dependencies. This feature enables dilated convolution to capture multi-scale features at the cost of linear parameter growth, providing theoretical support for the analysis of cross-cycle modulation patterns in complex vibration signals.

[0029] Mathematically, the weight matrix dimensions of the dilated convolution are consistent with those of the traditional convolution, both of which are , but its sampling position is separated by a hole Recode.

[0030] Use dilated convolution to perform nonlinear fusion of multi-scale context information on local spatial features to obtain multi-scale features; for the input local spatial feature position , the output response is formalized as:

[0031] Where, The multi-scale features of the output are at position The value of is the convolution kernel weight, Represented as an interval The sampled input value.

[0032] when When , it degenerates into standard convolution. The space-time hole structure forms a sparse connection. When the effective convolution kernel covers the area from Expand to , while the number of parameters remains unchanged. This design balances computational efficiency and receptive field expansion requirements through sparse sampling, and its computational complexity is still , but the effective receptive field is increased to .

[0033] Compared with traditional dilated convolution, dilated convolution achieves multi-scale feature aggregation by adjusting the dilation rate while maintaining the same number of parameters. When the dilation rate is 0.000, a single dilated convolution layer is equivalent to stacking multiple convolution layers with different dilation rates, but the number of parameters is reduced by about 67%. This feature gives it a unique advantage in multi-scale signal processing: by configuring a multi-branch dilated convolution structure, local impact features can be extracted in parallel. and global periodic modulation characteristics , and achieve multi-scale time-frequency coupling representation through feature splicing. In addition, the sparse connection characteristics of the void convolution can suppress the interference of adjacent point noise. Its mathematical form shows that when there is local noise in the input signal, the void rate The increase of will dilute the noise weight ratio, thereby improving the feature robustness; In step S2, the dynamic frequency band enhancement module generates a channel weight vector from the multi-scale features through the channel attention mechanism, and then multiplies the channel weight vector with the multi-scale features output by the multi-scale frequency perception module channel by channel to enhance the important channels and realize feature recalibration; the channel attention mechanism includes two modules: feature compression Squeeze and weight generation Excitation, such as Figure 2 As shown in the figure, the feature compression Squeeze module compresses the spatial dimension of multi-scale features through global average pooling to generate channel-level statistics. Suppose the multi-scale features of the input , then the compressed channel vector Calculated as:

[0034] Where, Indicates the channel vectors, is the number of channels, is the spatial dimension, Represents the input multi-scale feature Channels in space The value of The channel weight vector is calculated using the weight generation Excitation module , the formula is:

[0035] Where, , , and is a learnable parameter, is the dimensionality reduction ratio, is the Sigmoid function, is the ReLU activation function; Finally, the channel weight vector Multiply the multi-scale features output by the multi-scale frequency perception module channel by channel to obtain the enhanced feature map. The formula is:

[0036] Where, represents the enhanced features of the output, is the feature scaling function, represents channel-level multiplication, For the The weight of the channel, Multi-scale feature aisle; In order to solve the time-varying characteristics brought by variable speed as much as possible, the present invention designs a time series compression module. The core is to achieve feature dimensionality reduction by eliminating data redundancy in the time dimension, extract time domain invariant features through a sliding window, retain the main frequency component of the vibration signal, and eliminate speed fluctuation interference; the time domain invariance compression module uses the time series compression module to perform time series compression to extract time domain invariant features, retain the main frequency component of the vibration signal, and eliminate speed fluctuation interference. The formula for time series compression is:

[0037] Where, Is the time point after time series compression The characteristic value of is the window size; Subsequently, the time axis receptive field is expanded by using dilated convolution and applied to the time series compression module, covering 1.8 times the original period range. The formula is:

[0038] Where, For time point The output eigenvalues ​​of Is the convolution kernel at position The weight parameter, Indicates the input signal at time point The value of is the void ratio, is the convolution kernel width; Finally, the time dimension is compressed by mean pooling to retain the statistical characteristics between channels. The formula is expressed as:

[0039] Where, For the The timing compression characteristics of the channel compression output, Is the input feature in the channel , time point The value of is the original time step, .

[0040] S3, determining whether the number of training times M of the multi-scale frequency network perception model reaches the preset number of iterations N, if so, proceeding to the next step; otherwise, repeating step S2; It is important to note that the following factors should be considered when setting the number of model training times in this step: First, problem complexity: Complex problems may require more training and iterations. If the problem has highly nonlinear or abstract features, more iterations may be needed to capture these complexities. Second, dataset size: Larger datasets generally require more training times. More data helps the model generalize better and reduces the risk of overfitting. Third, early stopping: This strategy stops training when performance on the validation set no longer improves. This helps prevent overfitting and helps find an appropriate number of training times without wasting computing resources.

[0041] S4. Use the test set to test the trained model, and introduce accuracy, confusion matrix, ROC curve, and PRC curve to evaluate model performance.

[0042] The core of the multi-scale frequency network perception model (MSFAN-DFE) of the present invention is composed of a multi-scale frequency perception module, a dynamic frequency band enhancement module and a time domain invariance compression module. Its structure is as follows: Figure 3 As shown in the figure, a cascade structure of depth-wise separable convolution and dilated convolution is used to construct a frequency-axis multi-scale feature extractor to achieve cross-scale frequency feature extraction. A depth-wise separable convolution kernel of (1×7) is used to perform lightweight feature extraction along the frequency axis, which can significantly reduce the number of parameters. Dilated convolution is used instead of two-layer point-by-point convolution, the dilation rate is set to 3, the number of channels is 32, and the receptive field is expanded to 13 sampling points. Multi-scale feature extraction is achieved through this cascade structure, implicitly capturing frequency features of different scales from local to global. The channel attention mechanism extracts global statistical information of each channel through global average pooling and compressing spatial dimensions. After two dimensional adjustments (Permute), the convolution operation is adapted to use lightweight Convolution generates a channel weight vector, which is normalized to Its adaptive characteristics are due to the strong correlation between the weight and the frequency domain energy distribution of the input data: if a channel contains a critical frequency band for bearing faults, such as the impact resonance frequency band or modulation sideband, the weight approaches 1, and the channel characteristics are globally enhanced through the broadcast mechanism; conversely, the weight of the noise-dominated channel approaches 0, achieving adaptive suppression without the need for manual threshold setting. The process is as follows Figure 4 shown.

[0043] Then, by performing a global mean pooling operation along the time axis, the vibration signal features of the nine time segments in the multidimensional feature map are compressed into a single frequency domain representation, achieving feature fusion and dimensionality reduction across time dimensions. Time series compression combines the dual characteristics of time-domain noise suppression and frequency-domain feature focus. Firstly, it smoothes random impact interference through mean operations, improving the model's robustness in strong noise environments; secondly, it removes redundant time-domain information, enabling the subsequent attention mechanism to precisely target fault-sensitive frequency bands.

[0044] Experimental design and validation 1. Selected experimental dataset In order to verify the effectiveness of the multi-scale frequency network perception model MSFAN-DFE model, this paper selects the Ottawa bearing dataset of Canada as the experimental dataset. The experimental platform of the variable working condition bearing dataset of the University of Ottawa, Canada consists of a motor, an AC drive, an encoder, a coupling and an ER16K ball bearing, such as Figure 5As shown. The vibration signal is collected by an accelerometer with a sampling frequency of 200 kHz and a duration of 10 seconds. The data is stored in a MAT file, which contains vibration signals (Channel_1) and speed signals (Channel_2). The dataset covers four types of faults: healthy bearings, inner ring faults, outer ring faults, and ball faults. Each fault type simulates four variable speed conditions: speed increase, deceleration, speed increase and then deceleration, and deceleration and then speed increase. The characteristic of this dataset is the simulation of variable speed conditions. It is suitable for studying the impact of speed changes on fault characteristics and transfer learning tasks. This experiment sets up four experimental categories and selects datasets under 10 working conditions for experimental verification; as shown in Table 1.

[0045] Table 1 Bearing dataset

[0046] 2. Experimental parameter settings The bearing data was processed in blocks, with each data set divided into 1,000 samples, for a total of 10,000 samples. Before downsampling, the experimental data was filtered, and an FIR low-pass filter was designed to eliminate high-frequency components and avoid aliasing. After downsampling, each sample had 1,000 data points. The differential equation is:

[0047] Where, is the filter order (the number of coefficients is ), are the filter coefficients (also called tap weights), The input signal at time The value of The data is then normalized and linearly mapped to The interval is used to accelerate the convergence of the model. The calculation formula is:

[0048] Then the training and test data sets are randomly divided into 8:2 ratios. The normalized data are subjected to short-time Fourier transform.

[0049] Where, is the number of FFT points, is the short-time Fourier transform output result, is the time index, is the frequency index, is a discrete time point, is the value of the corresponding input signal at a discrete time point, is the Hann window function, is frame shift, is the shifted window function.

[0050] 3. Experimental Results and Analysis 3.1MSFAN-DFE Effectiveness Verification The preprocessed STFT spectrum graph is input into the MSFAN-DFE model, and the ADAM optimizer is selected to update the network parameters. The learning rate is 0.001, the number of iterations is 100, and the number of samples per batch is 16. The accuracy, confusion matrix, ROC curve, and PRC curve are introduced to evaluate the diagnostic performance of the model. The accuracy and confusion matrix of MSFAN-DFE on the test set are shown in the figure. Figure 6-12 As shown in the figure, the experimental data contains a total of 10 million points, the ratio of training set to test set is 8:2, and the accuracy of the test set is as high as 99.60%.

[0051] 3.2 Ablation Experiment To verify the diagnostic performance of the three modules—Cov2D dilated convolution, dynamic frequency channel attention (DFBA), and time series compression (TC)—a systematic ablation study was conducted. Under the same training conditions, fault diagnosis was performed by removing the dilated convolution, attention mechanism, and time series compression modules, respectively. Cov2D+DFBA removes the time series compression (TC) module; Cov2D+TC removes the dynamic frequency channel attention module; DFBA+TC removes the multi-scale dilated convolution module; and MSFAN-DFE represents the complete model, including all three modules. The resulting accuracy and test loss values ​​are shown in Table 2: Table 2 Ablation experiment

[0052] Ablation experiments demonstrate that the full MSFAN-DFE model achieves an optimal balance between accuracy and efficiency through multi-module collaborative optimization. The dynamic frequency channel attention (DFBA) module plays a decisive role in model performance: the accuracy of the Cov2D+TC model with DFBA removed drops significantly to 93.85%, while the loss surges to 0.1866, demonstrating its critical importance for feature selection. The temporal compression (TC) module and the multi-scale dilated convolution module jointly optimize computational efficiency and model lightweighting: the Cov2D+DFBA model with TC removed has a parameter count of 410,000 and an inference time of 1.21 seconds. However, the DFBA+TC model with multi-scale dilated convolution removed, despite doubling the number of parameters, still boasts significantly higher inference time than the full model, demonstrating the contribution of the multi-scale architecture to hierarchical feature extraction and computational acceleration. Leveraging the complementary nature of its modules, the full model achieves the lowest per-sample latency and minimal parameter count while maintaining the highest accuracy, providing theoretical support for high-precision real-time deployment. These results fully demonstrate the design necessity of each innovative module and their synergistic mechanism.

[0053] 3.3 Comparative Experiment In order to verify the model's excellence in accuracy and lightweightness, it is compared with other deep learning models. Figure 8 In this paper, SE-ResNet (WU X, PENG H, CUIX, et al. Multichannel Vibration Signal FusionBased on Rolling Bearings and MRST-Transformer Fault Diagnosis Model [J]. Ieee Sensors Journal, 2024, 24(10): 16336-46), MobileNetV2 (HU B, TANG J, WU J, et al. An Attention EfficientNet-Based Strategy for Bearing Fault Diagnosisunder Strong Noise [J]. Sensors, 2022, 22(17)), and EfficientNetB0 represent the use of the corresponding models to replace the multi-scale frequency-aware network in this paper, and the subsequent attention mechanism and temporal compression module are not modified. MSFAN-DFE represents the method proposed in this study. The first layer of SE-ResNet uses a Conv1D with a kernel size of 5 followed by a MaxPooling1D with 16 output channels, followed by two simplified residual blocks with a kernel size of 3 and 32 output channels, and adds Dropout (0.3). Finally, it uses global average pooling to a 32-dimensional fully connected layer to output the classification result. The first layer of MobileNetV2 uses a Conv1D with a stride of 2 and a kernel size of 3, with 16 output channels, followed by 6 inverted residual blocks, and finally uses global average pooling to a 16-dimensional fully connected layer to output the classification result. The first layer of EfficientNetB0 uses a Conv1D with a stride of 2, a kernel size of 3, and 32 output channels, followed by 7 groups of MBConv blocks with a kernel size of 3 / 5 and a scaling factor of 6, and finally uses global average pooling to output a fully connected classification layer.

[0054] Table 3 Model comparison experiment

[0055] This study compares the performance of four models: SE-ReaNet, EfficientNetB0, MobileNetV2, and MSFAN-DFE. As shown in Table 3, the comprehensive advantages of the MSFAN-DFE method in fault diagnosis tasks are verified.

[0056] MSFAN-DFE significantly outperforms other models with an accuracy of 99.60%, while its loss is only 18%-7% of the comparison models, demonstrating its stronger feature representation capabilities and training stability. This demonstrates that the proposed model possesses excellent classification accuracy and generalization capabilities. Furthermore, in terms of lightweightness, MSFAN-DFE reduces the number of parameters by 76.4% compared to MobileNetV2 and 57.0% compared to SE-ReaNet. Its single-sample inference latency is comparable to that of MobileNetV2, while its total inference time is 31.3% lower than that of EfficientNetB0, validating its innovative design in terms of model compression and computational efficiency. Overall, MSFAN-DFE achieves a 44.1% improvement in inference speed compared to SE-ReaNet while maintaining the highest accuracy, demonstrating the synergistic optimization effect of temporal compression (TC) and dynamic frequency-channel attention (DFBA). Compared to EfficientNetB0, which has a similar number of parameters, MSFAN-DFE achieves a 1.35 percentage point improvement in accuracy, demonstrating the effectiveness of the multi-scale dilated convolution module in capturing local features.

[0057] The MSFAN-DFE's 0.29ms single-sample latency meets real-time monitoring requirements. Its model memory occupies only 0.18MB, making it suitable for deployment in embedded devices. It provides a high-precision, low-cost solution for engineering applications of mechanical fault diagnosis.

[0058] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0059] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A bearing fault diagnosis method based on multi-scale frequency sensing dynamic enhancement, characterized in that: The following steps are involved: S1. Collecting vibration signals of a bearing in operation, preprocessing the vibration signals to obtain a time-frequency matrix, and then dividing the time-frequency matrix into a training set and a test set, wherein the preprocessing includes downsampling, normalization, and short-time Fourier transform; S2. Build a multi-scale frequency network perception model, which consists of a multi-scale frequency perception module, a dynamic frequency band enhancement module, a time domain invariance compression module, and an output layer. Input the time-frequency matrix in the training set of step S1 into the multi-scale frequency network perception model, use the multi-scale frequency perception module to extract multi-scale features, recalibrate the multi-scale features through the dynamic frequency band enhancement module, and perform time series compression through the time domain invariance compression module to obtain time series compression features. After flattening and dimensionality reduction in the output layer, the time series compression features are used by the classifier to generate a probability distribution of fault categories. S3, determining whether the training number M of the multi-scale frequency network perception model reaches the preset number of iterations N, and if so, proceeding to the next step; Otherwise, repeat step S2; S4. Use the test set to test the trained model, and introduce accuracy, confusion matrix, ROC curve, and PRC curve to evaluate model performance.

2. The bearing fault diagnosis method based on multi-scale frequency sense dynamic enhancement according to claim 1 is characterized in that: In step S1, the specific steps of the preprocessing are: down-sampling the vibration signal using an FIR low-pass filter, then normalizing the down-sampled signal to the [-1, 1] interval, and then performing a short-time Fourier transform to convert it into a time-frequency matrix.

3. The bearing fault diagnosis method based on multi-scale frequency sense dynamic enhancement according to claim 1 is characterized in that: In step S1, the time-frequency matrix is ​​divided into a training set and a test set in a ratio of 8:

2.

4. The bearing fault diagnosis method based on multi-scale frequency sense dynamic enhancement according to claim 1 is characterized in that: In step S2, the multi-scale frequency perception module uses a cascade structure of depthwise separable convolution and dilated convolution to extract the features of the training set, that is, the depthwise separable convolution extracts the local frequency features of the time-frequency matrix, and the dilated convolution expands the receptive field to extract the global frequency features, thereby extracting multi-scale features; The depthwise separable convolution performs independent spatial filtering operations on each channel of the input time-frequency matrix to obtain the local spatial features of each channel. The response calculation of the input channels is used to output the local spatial features. The formula is: ; Where, is the local spatial feature of the output, is the bias term of the corresponding input channel in the depth convolution stage, is the convolution kernel size; and Represents the spatial position index within the convolution kernel, , , The corresponding input channel in the spatial position The weight, Indicates the The convolution kernel of the channel is at position The weight of Indicates the first Channels are located at the convolution window position Pixel value of Use dilated convolution to perform nonlinear fusion of multi-scale context information on local spatial features to obtain multi-scale features; for the input local spatial feature position , the output response is formalized as: ； Where, The multi-scale features of the output are at position The value of is the convolution kernel weight, Represented as an interval The sampled input value.

5. The bearing fault diagnosis method based on multi-scale frequency sense dynamic enhancement according to claim 1 is characterized in that In step S2, the dynamic frequency band enhancement module generates a channel weight vector from the multi-scale feature through the channel attention mechanism, and then multiplies the channel weight vector with the multi-scale feature output by the multi-scale frequency perception module channel by channel to enhance the important channel and realize feature recalibration; the channel attention mechanism includes two modules: feature compression Squeeze and weight generation Excitation. The feature compression Squeeze module compresses the spatial dimension of the multi-scale feature through global average pooling to generate channel-level statistics. Suppose the input multi-scale feature , then the compressed channel vector Calculated as: ； Where, Indicates the channel vectors, is the number of channels, is the spatial dimension, Represents the input multi-scale feature Channels in space The value of The channel weight vector is calculated using the weight generation Excitation module , the formula is: ; Where, , , and is a learnable parameter, is the dimensionality reduction ratio, is the Sigmoid function, is the ReLU activation function; Finally, the channel weight vector Multiply the multi-scale features output by the multi-scale frequency perception module channel by channel to obtain the enhanced features. The formula is: ; Where, represents the enhanced features of the output, is the feature scaling function, represents channel-level multiplication, For the The weight of the channel, Multi-scale feature aisle.

6. The bearing fault diagnosis method based on multi-scale frequency sense dynamic enhancement according to claim 1 is characterized in that: In step S2, the time domain invariance compression module uses the time series compression module to perform time series compression to extract the time domain invariant features, retain the main frequency components of the vibration signal, and eliminate the speed fluctuation interference. The formula for time series compression is: ； Where, Is the time point after time series compression The characteristic value of is the window size; Subsequently, the time axis receptive field is expanded by using dilated convolution and applied to the time series compression module, covering 1.8 times the original period range. The formula is: ; Where, For time point The output eigenvalues ​​of Is the convolution kernel at position The weight parameter, Indicates the input signal at time point The value of is the void ratio, is the convolution kernel width; Finally, the time dimension is compressed by mean pooling to retain the statistical characteristics between channels. The formula is expressed as: ; Where, For the The timing compression characteristics of the channel compression output, Is the input feature in the channel , time point The value of is the original time step, .

7. The bearing fault diagnosis method based on multi-scale frequency sense dynamic enhancement according to claim 1 is characterized in that: In step S2, in the output layer, the time series compression features are flattened into a one-dimensional feature vector by the flattening layer, and then input into the fully connected layer. 16 ReLU neurons are used to reduce the feature dimension to obtain a 16-dimensional feature vector. The 16-dimensional feature vector is then input into the classifier to generate the fault category probability distribution through 4 Softmax neurons.

Images