An adaptive variable scale rolling bearing fault diagnosis method based on lightweight CNN

Abstract

The application discloses a kind of self-adapting variable scale rolling bearing fault diagnosis methods based on lightweight CNN, comprising: obtaining original vibration signal, constructs multimodal input feature map;Multimodal input feature map is input to lightweight main network and is extracted to feature, and basic feature is obtained;Basic feature is input to cascaded self-adapting variable scale feature extraction module, the module includes multiple cascaded variable scale feature extraction unit, each variable scale feature extraction unit generates corresponding sample level dynamic weight coefficient according to input basic feature, and self-adapting executes multi-scale feature extraction operation or executes identity mapping operation, and obtains output feature;The output feature of last unit is used as the input feature of next unit, and after being processed gradually, finally, enhanced feature is obtained;According to fusion feature, fault diagnosis result is output.The application provides a kind of solution with high precision and high efficiency for bearing fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of mechanical bearing fault diagnosis technology, and particularly relates to an adaptive variable-scale rolling bearing fault diagnosis method based on lightweight CNN. Background Technology

[0002] As a critical load-bearing component of rotating machinery, rolling bearings are subjected to harsh conditions such as variable loads, strong impacts, and complex couplings over long periods, making them highly susceptible to pitting, spalling, and wear failures. Their operational status directly impacts the overall safety and production efficiency of high-end equipment in wind power generation, aerospace, rail transportation, and industrial robots. Therefore, conducting efficient, accurate, and real-time fault diagnosis research on rolling bearings has become an urgent need for both academia and industry.

[0003] Traditional fault diagnosis methods mainly rely on signal processing techniques such as Fast Fourier Transform, Wavelet Transform, and Empirical Mode Decomposition to extract discriminative feature indicators from the time or frequency domain. However, these methods heavily depend on expert prior knowledge for manual feature engineering, making it difficult to effectively address the high-dimensional, nonlinear, and non-stationary characteristics of massive monitoring data.

[0004] With the development of deep learning technology, data-driven methods, represented by convolutional neural networks, have achieved breakthroughs in the field of fault diagnosis due to their end-to-end feature self-learning capabilities. Researchers have converted one-dimensional vibration signals into two-dimensional images and used deep networks to automatically mine semantic features, significantly improving diagnostic accuracy under complex working conditions. However, in practical industrial applications, especially in resource-constrained edge computing scenarios, existing mainstream network architectures face severe challenges. Edge devices are typically equipped with limited computing power and storage space, making it difficult to support deep network models with millions of parameters.

[0005] To meet the needs of edge deployment, lightweight network design has become a research hotspot. Existing lightweight frameworks effectively reduce the computational complexity of models through methods such as depthwise separable convolution and knowledge distillation. However, this lightweighting is often accompanied by a decrease in feature extraction capabilities, making it difficult to capture subtle features of weak, complex, or early faults.

[0006] Specifically, rotating machinery fault signals exhibit significant non-stationary and multi-scale characteristics: early, weak faults typically manifest as high-frequency, short transient impacts, while complex faults or later-stage damage are often accompanied by low-frequency periodic modulations and long-term dependence. This dramatic change in time scale necessitates networks with flexible variable-scale feature extraction capabilities. Existing lightweight networks typically employ fixed-size convolutional kernels, resulting in models with fixed receptive fields, making it difficult to accurately capture both microscopic impacts and macroscopic controls simultaneously within a single network. Some studies have attempted to introduce multi-scale structures, but simple parallel stacking of multiple scales significantly increases the number of parameters, contradicting the initial intention of lightweight design.

[0007] Furthermore, network pruning techniques have been shown to effectively reduce model complexity by removing redundant weights or neurons. However, traditional pruning methods typically follow a cumbersome process of "training-evaluation-pruning-fine-tuning," which is time-consuming and labor-intensive, and relies on expert experience to manually set pruning thresholds or ratios, making it difficult to achieve automated model optimization.

[0008] Therefore, for fault diagnosis, how to design a neural network model that can adaptively capture multi-scale features, maintain extreme lightweight design, and require no manual intervention is a key problem that urgently needs to be solved. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention provides an adaptive variable-scale rolling bearing fault diagnosis method based on lightweight CNN, comprising the following steps: The original vibration signal is acquired, and a multimodal input feature map is constructed based on the original vibration signal. The multimodal input feature map is input into a lightweight backbone network for feature extraction to obtain basic features; The basic features are input into a cascaded adaptive variable-scale feature extraction module, which includes multiple cascaded variable-scale feature extraction units. Each variable-scale feature extraction unit generates corresponding sample-level dynamic weight coefficients based on the input basic features, and adaptively performs multi-scale feature extraction or identity mapping operations on the input basic features based on these dynamic weight coefficients to obtain output features. The output features of the previous unit serve as the input features of the next unit, and after step-by-step processing, the enhanced features are finally obtained. The fault diagnosis results are output based on the fusion features.

[0010] Optionally, a multimodal input feature map is constructed based on the original vibration signal, specifically including: The original vibration signal is preprocessed, and two-dimensional feature representations of the Gram angle difference field are constructed in the time domain and frequency domain, respectively. The two-dimensional feature representations of Gram angle difference fields in the time and frequency domains are fused to form a multimodal input feature map.

[0011] Optionally, the multimodal input feature map is input into a lightweight backbone network for feature extraction to obtain basic features, specifically including: Intermediate features are obtained by downsampling the multimodal input feature map through cascaded standard convolutional layers and pooling layers; The intermediate features are extracted by using depthwise separable convolutional layers, and a channel attention mechanism is embedded in the depthwise separable convolutional layers to adaptively weight the channel features to obtain the basic features.

[0012] Optionally, each variable-scale feature extraction unit generates corresponding sample-level dynamic weight coefficients based on the input basic features, specifically including: Global average pooling is performed on the basic input features in the spatial dimension to obtain a channel-level global feature description vector. The channel-level global feature description vector is input into the weight prediction network to generate dynamic weight coefficients between 0 and 1.

[0013] Optionally, each variable-scale feature extraction unit adaptively performs multi-scale feature extraction or identity mapping operations on the input basic features based on dynamic weight coefficients to obtain output features, specifically including: Multi-scale feature extraction is performed on the input basic features to obtain intermediate enhanced features; Based on the dynamic weight coefficient, the input basic features and intermediate enhancement features are weighted and fused. The fusion weight of the input basic features is 1 minus the dynamic weight coefficient, and the fusion weight of the intermediate enhancement features is the dynamic weight coefficient. The output features are obtained after weighted fusion.

[0014] Optionally, the cascaded adaptive variable-scale feature extraction module includes four cascaded variable-scale feature extraction units, wherein: The first variable-scale feature extraction unit extracts multi-scale features from the input features using dilated convolutions with a first dilation rate. The second variable-scale feature extraction unit performs multi-scale feature extraction by using dilated convolution with a second dilation rate on the input features, where the second dilation rate is greater than the first dilation rate. The third variable-scale feature extraction unit uses the hollow spatial pyramid pooling module to extract multi-scale features from the input features; The fourth variable-scale feature extraction unit performs feature smoothing and calibration on the input features using dilated convolution with a first dilation rate.

[0015] Optionally, the hollow spatial pyramid pooling module includes multiple parallel dilated convolution branches with different dilation rates and a global average pooling branch, and the outputs of each branch are concatenated in the channel dimension to form a multi-scale aggregated feature.

[0016] Optionally, the fault diagnosis result is output based on the fusion features, specifically including: Global average pooling is performed on the fused features to obtain the global feature vector; The global feature vector is input into the fully connected layer, and the probability distribution of each category is output through the normalized exponential function. The category corresponding to the highest probability is taken as the fault diagnosis result.

[0017] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0018] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0019] Compared with the prior art, the present invention has the following advantages and technical effects: This invention first constructs a multimodal input feature map based on the original vibration signal, fusing time-domain and frequency-domain information as input. Then, the multimodal input feature map is fed into a lightweight backbone network for feature extraction, yielding basic features. This backbone network employs a lightweight design, effectively controlling the number of parameters while maintaining feature extraction capabilities, thus laying the foundation for subsequent processing.

[0020] To address the problem that existing lightweight networks using fixed-size convolutional kernels result in a fixed receptive field, making it difficult to simultaneously capture the high-frequency transient impact features of weak faults and the low-frequency periodic modulation features of complex faults, this invention inputs basic features into a variable-scale feature extraction module. This module extracts features from the basic features at multiple scales through multiple cascaded variable-scale feature extraction units, constructing a variable receptive field that covers everything from local textures to global dependencies. Through this technique, this invention effectively solves the problem that a single-scale convolutional kernel cannot simultaneously capture various fault features, significantly improving the diagnostic accuracy for complex fault modes.

[0021] To address the problems of existing network pruning methods, which follow a cumbersome "training-evaluation-pruning-fine-tuning" process, are time-consuming and labor-intensive, and rely on expert experience to manually set pruning thresholds, this invention generates dynamic weight coefficients at the sample level based on basic features, and then weights and fuses basic and enhanced features according to these dynamic weight coefficients. This technique endows the network with "self-awareness": for samples with sufficient information and low complexity, the dynamic weight coefficients approach zero, and the network automatically performs structural pruning, skipping redundant deep computational paths; for samples with insufficient information and high complexity, the dynamic weight coefficients approach one, and the network forcibly activates deep multi-scale modules to extract richer semantic features. Through this mechanism, this invention achieves a dynamic balance between computational resources and diagnostic accuracy, and the entire optimization process is carried out synchronously with network training, requiring no manual intervention.

[0022] Ultimately, this invention achieves high-precision fault diagnosis while maintaining an extremely low parameter count by classifying and outputting the fused features. Through the synergistic effect of the above-mentioned technical means, this invention breaks through the feature extraction bottleneck of traditional lightweight models with a lightweight network structure, achieving a dynamic balance between accuracy and efficiency, and providing an engineering solution with both high accuracy and high efficiency for resource-constrained edge computing scenarios. Attached Figure Description

[0023] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the AVS-LCNN model structure according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the confusion matrix under different domains in an embodiment of the present invention; Figure 4 This is a visualization of t-SNE in different domains according to an embodiment of the present invention. Detailed Implementation

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0026] Example 1 This embodiment provides an adaptive variable-scale rolling bearing fault diagnosis method based on lightweight CNN, including the following steps: The original vibration signal is acquired, and a multimodal input feature map is constructed based on the original vibration signal. The multimodal input feature map is input into a lightweight backbone network for feature extraction to obtain basic features; The basic features are input into a cascaded adaptive variable-scale feature extraction module, which includes multiple cascaded variable-scale feature extraction units. Each variable-scale feature extraction unit generates corresponding sample-level dynamic weight coefficients based on the input basic features, and adaptively performs multi-scale feature extraction or identity mapping operations on the input basic features based on these dynamic weight coefficients to obtain output features. The output features of the previous unit serve as the input features of the next unit, and after step-by-step processing, the enhanced features are finally obtained. The fault diagnosis results are output based on the fusion features.

[0027] As a feasible implementation method, the method provided in this embodiment achieves intelligent diagnosis of rolling bearing faults by constructing a hierarchical variable-scale feature extraction mechanism, a sample-aware adaptive dynamic weighting strategy, and an extremely lightweight multimodal feature fusion structure.

[0028] The method first constructs a multimodal input feature map. Specifically, the original vibration signal is acquired, and Gram angle difference field coding technology is used to convert the one-dimensional vibration signal in the time domain and the frequency domain signal after fast Fourier transform into two-dimensional images, respectively. The time domain image and the frequency domain image are then superimposed in the channel dimension to form a dual-channel multimodal input feature map, making full use of the complementarity between time domain and frequency domain information.

[0029] The proposed method employs depthwise separable convolutions to reconstruct a lightweight backbone network. Multimodal input feature maps are fed into the lightweight backbone network, sequentially downsampled through standard convolutional and pooling layers, and then feature extraction is performed through depthwise separable convolutional layers. A channel attention mechanism is embedded within the depthwise separable convolutions to adaptively weight the channel features, yielding the base features. This backbone network significantly reduces the number of parameters while maintaining feature extraction capabilities.

[0030] The method constructs a hierarchical variable-scale feature extraction mechanism. Basic features are input into a variable-scale feature extraction module, which consists of multiple cascaded variable-scale feature extraction units. Each unit employs dilated convolution, gradually expanding the model's receptive field by setting different dilation rates. Specifically, the first variable-scale feature extraction unit uses dilated convolution with a small dilation rate to capture local transient impact features; the second variable-scale feature extraction unit uses dilated convolution with a larger dilation rate than the first to extract periodic modulation signals and establish temporal contextual relationships; the third variable-scale feature extraction unit employs a dilated spatial pyramid pooling module, using multiple parallel dilated convolution branches with different dilation rates and a global average pooling branch to aggregate features at multiple scales, explicitly fusing contextual information at different scales; the fourth variable-scale feature extraction unit uses dilated convolution with a small dilation rate to smooth and calibrate the multi-scale fused features. Through this coarse-to-fine feature extraction process, the module constructs a variable receptive field covering everything from local texture to global dependencies.

[0031] The proposed method introduces a sample-aware adaptive dynamic weighting strategy. In the variable-scale feature extraction module, each variable-scale feature extraction unit corresponds to a learnable dynamic weight coefficient. This dynamic weight coefficient is generated by inputting the basic features into the weight prediction network after global average pooling, and is a continuous value between 0 and 1. Based on this dynamic weight coefficient, the basic features and the enhanced features output by the corresponding variable-scale feature extraction units are linearly weighted and fused. That is, the fused feature equals one minus the dynamic weight coefficient multiplied by the basic feature, plus the dynamic weight coefficient multiplied by the enhanced feature. This mechanism enables the network to adaptively adjust the feature enhancement strength according to the complexity of the input samples: for simple samples, the dynamic weight coefficient approaches 0, automatically performing structural pruning to accelerate inference; for complex samples, the dynamic weight coefficient approaches 1, forcibly activating deep multi-scale modules to enhance robustness, achieving a dynamic balance between computational accuracy and efficiency.

[0032] Finally, global average pooling is performed on the fused features to obtain a global feature vector. The probability distribution of each category is then output through a fully connected layer and a normalized exponential function. The category with the highest probability is taken as the fault diagnosis result.

[0033] The feasible basic principle of Gram angular difference field (GADF): To address the challenge of characterizing the global temporal dependencies in one-dimensional time series signals, the Gramian Angular Field (GAF) method maps the time series to polar coordinate space and constructs a two-dimensional Gram matrix, achieving a structured representation of one-dimensional signals into two-dimensional images. Among these, the Gramian Angular Difference Field (GADF), an important form of GAF, effectively characterizes the relative changes between different time points in a time series and has been widely applied in time series signal analysis and fault diagnosis. The specific workflow is as follows: Let the original time series signal be: To avoid numerical instability and satisfy the polar coordinate mapping condition, the signal is first normalized and mapped to the interval [-1, 1]: In the formula, and These represent the mean and standard deviation of the signal, respectively. For the original time series elements, These are the normalized time series elements.

[0034] Subsequently, the normalized time series is mapped to polar coordinate space, and the polar angle is defined using the inverse cosine function: During this mapping process, the time index is implicitly preserved in the order of the polar angles, thus ensuring that the time dependency is not broken.

[0035] Based on the polar angle sequence, the GADF matrix is ​​defined as: In the formula, , and For normalized time series The corresponding cosine angle, , = .

[0036] Implementable basic principles of convolution: Convolutional Neural Networks (CNNs) are a class of deep neural network structures that achieve efficient feature learning through local connections and weight sharing mechanisms, and have been widely used in fields such as image recognition and signal analysis. For image features formed after two-dimensional processing of vibration signals, CNNs can effectively mine their local spatial correlations and multi-level discriminative information.

[0037] In CNNs, convolutional layers extract local features by performing local sliding operations on the input feature map using convolutional kernels. The basic computational form can be expressed as: In the formula, This represents the feature map of the c-th channel. For convolution kernel weights, This is the output feature map. This local receptive field mechanism enables the network to focus on local structural patterns in the image, which is particularly effective in the two-dimensional feature representation of vibration signals.

[0038] The fundamental principle of feasible, depthwise separable convolution: To further reduce the computational complexity of the model and improve its feasibility for engineering applications, this embodiment uses a depthwise separable convolution structure instead of traditional convolution operations. This structure decomposes the standard convolution into two parts: depthwise convolution and pointwise convolution, thereby significantly reducing the parameter size and computational cost. The computation process can be represented as follows: In the formula, This indicates the spatial location of the c-th channel in the input feature map. The pixel value at position () is used; c is the channel index of the input feature map, where c = 1, 2, ..., C, and C represents the total number of input channels. ( ) represents the kernel weight corresponding to the c-th channel in the depthwise convolution, where ( ) represents the index of the convolution kernel in the spatial dimension; M×N represents the spatial size of the depthwise convolution kernel, which is usually 3×3 or 5×5; This indicates that the c-th channel in the feature map output by the depthwise convolution stage is at position ( The feature response value at point (). The characteristic of this stage is that each input channel performs convolution operations independently, and no information fusion occurs between channels.

[0039] in, This indicates the spatial location of the c-th channel output from the depthwise convolution stage. The eigenvalue at () is a position; The weight coefficients used in pointwise convolution to map the c-th input channel to the k-th output channel are essentially 1×1 convolution kernel weights; k is the channel index of the output feature map, k=1,2,...,K, where K represents the total number of output channels; This indicates the spatial location of the k-th channel output by the pointwise convolution stage. The feature response value at () represents the number of channels in the depthwise convolution output, which is equal to the number of channels in the input feature map. This stage achieves cross-channel feature fusion and dimension mapping through 1×1 convolution.

[0040] Feasible channel attention mechanisms: In the feature learning process of convolutional neural networks, the contribution of information contained in different channels to the fault identification task varies significantly. Traditional convolutional operations treat all channels equally during feature extraction, which can easily lead to key information being overwhelmed by redundant features, thus affecting the model's discriminative ability. To enhance the network's ability to focus on effective features, a channel attention mechanism is introduced into the feature extraction module to adaptively weight different channels.

[0041] Let the input feature map be: In the formula, C, H, and W represent the number of channels, height, and width, respectively. The channel attention mechanism first compresses the spatial dimensions through Global Average Pooling (GAP) to obtain a global statistical description of each channel: Subsequently, the obtained channel description vector =[z1,z2,…, The input is fed into a weight generation network consisting of fully connected layers and nonlinear activation functions to model the correlation between channels. The calculation process can be represented as follows: In the formula, This represents a nonlinear mapping function composed of a multilayer perceptron (MLP). It is the Sigmoid activation function. This is the channel weight vector.

[0042] Finally, the weight vector is applied to the original feature map through channel multiplication to achieve feature recalibration: In the formula, This represents the feature map output after channel attention enhancement.

[0043] The basic principle of feasible dilated convolution: In convolutional neural networks, the receptive field size directly affects the model's ability to model features at different scales. Traditional convolutions typically rely on increasing the kernel size or stacking more convolutional layers to expand the receptive field, which significantly increases the number of model parameters and computational complexity. Dilated convolution provides a more efficient way to model spatial features by introducing a spacing factor between kernel elements, thus expanding the receptive field without increasing the number of parameters.

[0044] Let the input feature map be The convolution kernel weights are The calculation form of dilated convolution can be expressed as: In the formula, This represents the dilation rate. These represent the spatial coordinate indices (row and column) of the output and input feature maps, respectively, used to determine the current convolution calculation position. When = 1, the dilated convolution degenerates into a standard convolution; when When the value is greater than 1, the convolution kernel is sparsely distributed in space, and its effective receptive field size expands to: In the formula, Let be the side length of the convolution kernel. This formula shows that as the dilation rate increases... The increase, It exhibits linear growth, thus significantly expanding the model's perception range of long-term dependent features without increasing the number of model parameters.

[0045] By adjusting the expansion rate Dilated convolution can capture a wider range of contextual information while keeping the number of convolution kernel parameters constant, making it particularly suitable for extracting vibration signal patterns with periodic or long-range dependent features.

[0046] The basic principle of feasible atrous spatial pyramid pooling: Atrous Spatial Pyramid Pooling (ASPP) is a multi-scale feature extraction structure based on dilated convolutions. Its core idea is to construct a spatial pyramid on the same feature layer by setting multiple convolutional branches with different dilation rates in parallel, thereby achieving joint modeling of contextual information at different scales. In rolling bearing vibration signals, the feature scales corresponding to different fault types differ significantly. For example, local defects typically exhibit short-term impact characteristics, while structural faults may present modulation characteristics with longer periods. Single-scale convolutional operations struggle to simultaneously capture information at different scales, easily leading to insufficient feature representation. Therefore, it is necessary to introduce a multi-scale feature extraction mechanism, enabling the network to simultaneously focus on local detailed features and global structural features, thereby improving the model's adaptability to complex fault modes.

[0047] Let the input feature map be X, the ASPP module can be represented as: In the formula, This represents a 1×1 convolution branch used to preserve original local features; These represent different expansion rates. The dilated convolutional branch is used to extract spatial features at different scales; This indicates the global average pooling branch, used to incorporate global context information; This represents a feature concatenation operation along the channel dimension.

[0048] Implementable, dynamic weight α: To achieve adaptive adjustment of the enhancement level for different sample features, this embodiment introduces a feature weighting mechanism based on sample-level dynamic weights in the convolutional neural network. This mechanism linearly fuses the basic features and enhanced features using learned continuous weight coefficients α, thereby achieving adaptive control of feature representation.

[0049] Let the input feature map be: First, perform global average pooling on the input features in the spatial dimension to obtain the channel-level global feature description vector: in, .

[0050] Then The input is fed into a weight prediction network consisting of fully connected layers and non-linear activation functions to obtain the dynamic weight coefficients corresponding to the samples: In the formula, Represents a nonlinear mapping function. For the Sigmoid function, .

[0051] set up This indicates that the input features are processed by the variable-scale feature extraction module. The enhanced feature representation obtained after transformation is the final output feature. It is obtained through the following linear weighting method: The aforementioned fusion process achieves continuous weighted adjustment of the original and enhanced features at the feature level, enabling the network to adaptively employ different scales of feature extraction for different samples during end-to-end training. This "targeted" feature extraction reduces feature redundancy caused by "unnecessary" feature extraction. This dynamic weighting mechanism achieves adaptive adjustment of the sample-level feature modeling strategy without introducing additional discrete decision rules.

[0052] A feasible fault diagnosis framework for Adaptive Variable-Scale Lightweight Convolutional Neural Network (AVS-LCNN): This embodiment constructs an end-to-end intelligent diagnostic system, aiming to solve the problems of difficulty in extracting bearing fault features, strong background noise, and limited computing resources at the edge in industrial environments.

[0053] The core design philosophy is "dynamic perception and on-demand computation." Unlike traditional CNNs that treat all samples "equally," this model introduces a dynamic weighting mechanism. For samples of varying complexity, features are extracted at different depths to avoid computational efficiency losses due to over-extraction. The model mainly consists of three stages: multimodal perception input, lightweight backbone extraction, and dynamic variable-scale feature extraction and fusion. Figure 1 As shown.

[0054] The original vibration signal is a one-dimensional time series. Although it contains rich frequency information, it is easy to lose temporal correlation when directly input into a convolutional neural network. To enable AVS-LCNN to better "understand" the signal, a multi-modal two-dimensional mapping strategy was designed. Modality 1 (temporal domain encoding): The original vibration signal is encoded into a two-dimensional image using GADF (Gram Angular Difference Field). GADF can preserve the temporal dependence of the signal and map the numerical transformations in the time series to the texture and pattern of the image. Modality 2 (frequency domain encoding): After performing an FFT transformation on the signal, GADF encoding is also used. The frequency domain image can intuitively reflect the energy distribution of fault characteristic frequencies (such as the outer and inner fault frequencies). Modality 3 (dual-channel fusion): The time domain and frequency domain images are superimposed to form a dual-channel input, solving the problem of incomplete information from a single modality. Furthermore, using the GADF image as input can fully utilize the texture extraction capabilities of AVS-LCNN.

[0055] The lightweight backbone extraction task is to quickly compress high-dimensional images into abstract feature maps, laying the foundation for subsequent refined analysis. The first two layers of this part use standard convolution plus max pooling to rapidly reduce the feature map size (64×64→32×32→16×16), reducing the computational cost. The third layer uses depthwise separable convolution. It separates "spatial convolution" and "channel convolution," with only one-eighth to one-ninth of the parameters of ordinary convolution. A channel attention mechanism module is embedded at the end of the backbone extraction layer. It can automatically identify which channels contain fault impact information and which channels are just background noise, and assign higher weights to the former. This design enables the backbone network to complete feature extraction and noise suppression with extremely low computational cost, providing "clean" and highly abstract feature maps for subsequent dynamic scale feature extraction and fusion layers.

[0056] A dynamic variable-scale feature extraction and fusion layer constructs a dynamically perceptive feature inference engine. Unlike traditional convolutional networks that use a fixed computation method, this architecture introduces a dynamic layer based on "soft gating and hard triggering," aiming to adaptively allocate computational resources for vibration signals with different signal-to-noise ratios. This module consists of four cascaded dynamic units, and its core design logic follows a hierarchical processing paradigm of "variable-scale feature extraction—multi-scale aggregation—feature smoothing." Each dynamic unit embeds a weight decision sub-network to quantify the dependence of the current input features on the transformation of this layer. This mechanism allows the network to build a "short-circuit" path (α→0) for simple samples, preserving the original features and reducing redundant computation; while for complex samples with high entropy values, the network automatically increases α, activating deep nonlinear transformations to provide abstract features.

[0057] The design principle of the four-layer variable-scale feature extraction is as follows: The basic features extracted by the lightweight backbone network are input into the first variable-scale feature extraction unit, which uses dilated convolution with a first dilation rate to extract features, resulting in a first enhanced feature. The first enhanced feature is then input into the second variable-scale feature extraction unit, which uses dilated convolution with a second dilation rate to extract features, resulting in a second enhanced feature, wherein the second dilation rate is greater than the first dilation rate. The second enhanced feature is then input into the third variable-scale feature extraction unit, which uses a dilated spatial pyramid pooling module to extract features at multiple scales, resulting in a third enhanced feature. Finally, the third enhanced feature is input into the fourth variable-scale feature extraction unit, which uses dilated convolution with a first dilation rate to smooth and calibrate features, resulting in an enhanced feature.

[0058] Specifically, all four layers employ dilated convolution. Without increasing computational cost, the receptive field of the model is gradually expanded by changing the dilation rate of each layer, enabling the model to extract more complex fault features. The first layer is a dilated convolution with a dilation rate of 2. While maintaining the feature map resolution, the dilation rate initially expands the receptive field, aiming to capture local transient impact features such as bearing ball scratches. The second layer is a standard dilated convolution with a dilation rate of 4, further expanding the receptive field to cover a longer time window. This layer focuses on extracting periodic modulation signals caused by low-frequency faults or speed fluctuations, establishing the temporal context of the signal. The third layer is Atrous Spatial Pyramid Pooling, which integrates parallel dilated convolutions into a (6, 12, 18) convolutional stream. This layer aims to address the multi-scale effect of fault features, as a single-scale convolutional kernel cannot simultaneously handle high-frequency pulses (small scale) and low-frequency envelopes (large scale). ASPP achieves explicit fusion of multi-scale features through multi-path parallel sampling. The model's ability to represent complex fault features is significantly enhanced. The fourth layer uses dilated convolution with an expansion rate of 2 as the final stage of feature extraction. This layer is used to smooth out the grid effect that may be introduced by large dilated convolution, and to finely reorganize and calibrate the features after multi-scale fusion to ensure that the features input to the classifier have high robustness.

[0059] This model achieves deep collaboration between low-level physical signals and high-level semantic representations by constructing a coarse-to-fine feature extraction process. Its core dynamic weighting mechanism endows the network with the ability of "variable-scale feature extraction," and α-soft gating enables on-demand allocation of computational resources, effectively overcoming the challenge of balancing model capacity and sample difficulty. Ultimately, this architecture achieves over 99% diagnostic accuracy with extremely low parameters, providing a highly efficient solution for fault diagnosis in resource-constrained scenarios that combines theoretical depth with practical engineering value.

[0060] Furthermore, this embodiment designs a lightweight rolling bearing fault diagnosis method based on the AVS-LCNN network framework with variable-scale feature extraction. The overall flowchart is as follows: Figure 2 As shown.

[0061] Feasible experimental verification and discussion: Experimental hardware and software configuration: To verify the fault diagnosis performance of the proposed AVS-LCNN network, this embodiment uses the High-Speed ​​Train Bogie Fault Dataset (HTBF) and the Paderborn University (PU) Bearing Dataset for experimental verification. All algorithms were implemented in PyTorch 1.13.0 and Python 3.8.20, with a computer configuration of i5-12500 CPU and GTX 1660 SUPER. The Adam optimizer was used with an initial learning rate of 0.001, a size of 64, and a total training period of 30 epochs. To quantify the behavior of the dynamic weights, during the testing phase, the proportion of samples in the four dynamic layers whose α exceeded the threshold (L1: 0.5, L2: 0.6, L3: 0.7, L4: 0.7) was statistically analyzed and defined as the "layer trigger rate".

[0062] Dataset introduction: Dataset 1: HTBF Dataset. Vibration signals were collected from the HTBF failure simulation platform. The experimental platform was designed as a 1:2 scale high-speed train bogie model. It was driven by a 2.2 kW three-phase asynchronous motor, with each axle box containing one HRB 352213 rolling bearing. The platform could simulate different operating conditions by adjusting the motor speed and horizontal and vertical loads. The bogie failure simulation experimental platform was equipped with four MS101050 single-axis accelerometers and one TES001T three-axis accelerometer, collecting signals from three measurement points: the axle box, gearbox, and motor, for a total of five channels. The sampling frequency of each channel was 25.6 kHz. A detailed description of the HTBF dataset is shown in Table 1.

[0063] Table 1 Dataset 2: PU Dataset. This dataset was acquired using a modular test bench and contains vibration signals of the 6203 bearing under four different speeds (900-1500 rpm) and load conditions, with a sampling rate of 64 kHz. A significant feature of the PU dataset is that it includes 14 sets of real fatigue damage samples generated through accelerated life testing (mainly pitting and wear on the inner and outer rings, 12 sets of artificially processed damage samples, and 6 sets of healthy samples). A detailed description of the PU dataset is shown in Table 2.

[0064] Table 2 Further experimental verification: HTBF dataset: For the HTBF dataset, this experiment uses bearing data in 9 different health states. The failure types cover the inner ring, outer ring and rolling elements of rolling bearings under different damage levels (level 1, level 2 and level 3), as well as cage noise and normal condition.

[0065] First, the original vibration signal was preprocessed, with a sliding window length of 1024 data points and a sampling rate of 12kHz. To ensure class balance, 1000 samples were randomly selected for each fault class, totaling 9000 samples. Then, the signal was encoded using multimodal GADF encoding, resulting in 64×64 time-domain, frequency-domain, and time-frequency-domain GADF images, which served as input to the neural network. The network was randomly divided into training and test sets in a 7:3 ratio. Table 3 summarizes the performance comparison of the HTBF dataset models. Table 3 Experimental results show that time-frequency domain features achieved the best classification accuracy (99.41%), followed closely by frequency domain features (99.22%), while time domain features performed relatively poorly (72.48%). This significant difference not only verifies the importance of data preprocessing but also reveals the dynamic adaptation mechanism of AVS-LCNN at a deeper level.

[0066] As shown in Table 3, the model achieves near-perfect classification results with extremely low computational cost. Furthermore, thanks to depthwise separable convolutions and a dynamic skipping mechanism, ADW-LCNN has only about 0.17M parameters. In terms of training efficiency, the model completes 30 training rounds in just over 80 seconds, averaging about 2.6 seconds per round, demonstrating high computational efficiency.

[0067] To explore the core "variable scale" and "dynamic weight" design of AVS-LCNN, this embodiment analyzes in depth the weight values ​​and trigger rate distribution of the four dynamic layers α under different modalities. The results reveal that the model has highly intelligent computational resource scheduling capabilities.

[0068] In the weaker temporal modalities, the trigger rates of Layer 3 and Layer 4 reached as high as 98.34% and 97.48%, respectively. This indicates that the model correctly identified the high uncertainty of temporal samples. The decision network forcibly activated the deep ASPP layer, attempting to compensate for insufficient input information through the large receptive field of ASPP. This behavior demonstrates the model's intelligent characteristic of proactive compensation when facing low-quality inputs.

[0069] In the frequency domain, the trigger rate exhibits a clear stepwise upward trend (L1: 14.27% → L2: 90.14% → L3: 96.52% → L4: 98.00%). This indicates that for the spectral image, the model tends to construct features layer by layer, gradually aggregating from local frequency peaks (L1 / L2) into global fault modes (L3 / L4), which aligns with the physical intuition of frequency domain analysis.

[0070] In the information-rich time-frequency domain modality, the trigger rates of Layer 1 and Layer 2 were 0.02% and 8.39%, respectively, while Layer 3 and Layer 4 remained at 97.96% and 96.30%. This indicates that the model determined that the dual-channel input already provided sufficiently clear low-level features, thus automatically "shutting down" the first two feature expansion modules and directly invoking high-level multi-scale semantic analysis. This dynamic inference strategy of "skipping the shallow layers and directly addressing semantics" not only saves computational resources but also avoids the noise amplification that may be introduced by shallow convolutions.

[0071] Time-domain models are prone to serious confusion between "rolling element failure" and "outer race failure." For example... Figure 3 As shown in the temporal confusion matrix, obvious misclassification blocks appear at the off-diagonal positions corresponding to these two types of faults, indicating that the simple temporal waveform (GADF texture) has low feature discrimination when describing different damage levels of the same type of fault and is susceptible to random noise interference. In contrast, the dual-channel input (time-frequency domain) significantly enhances the identification of periodic impact features by introducing frequency domain information. Figure 3 As shown in the mid-time-frequency domain confusion matrix, the brightness of the diagonal elements is concentrated, and the number of misclassified terms on the off-diagonal side is significantly reduced, demonstrating the effectiveness of multimodal input in improving classification consistency.

[0072] exist Figure 4 In the t-SNE visualization results, the nine fault classes in the time-frequency domain exhibit a highly compact and highly separable cluster structure with no obvious outliers. This indicates that after multi-scale aggregation of Layer 3 (ASPP) and feature calibration of Layer 4, the model successfully maps the original signal to a linearly separable high-dimensional semantic space. AVS-LCNN not only achieves a high-precision diagnosis of 99.41% under multimodal input, but more importantly, the difference in the distribution of α-trigger rates (full activation in the frequency domain vs. shallow pruning in the time-frequency domain) provides direct evidence for the model's dynamic reasoning ability. Figure 4As shown, while the frequency domain model achieves acceptable clustering results, some inter-class boundaries still exhibit fuzzy areas. In contrast, the time-frequency domain model shows highly clustered feature points across all categories and clear inter-class boundaries, verifying the model's ability to dynamically reconstruct the computation path based on the input information entropy. This "on-demand computation" characteristic enables it to achieve an optimal balance between performance and efficiency when dealing with industrial signals of varying quality.

[0073] PU dataset: To further evaluate the generalization robustness of the AVS-LCNN framework under different working conditions and real-world damage scenarios, this embodiment introduces the Paderborn University (PU) bearing dataset for supplementary validation.

[0074] The experiment selected vibration signals under the N15-M07-F04 operating condition, including five typical healthy states: healthy state (K001, Class A), outer ring fault 1 (KA01, Class B), outer ring fault 2 (KA04, Class C), inner ring fault 1 (KI08, Class D), and inner ring fault 2 (KI21, Class E). Compared with the artificial fault data in the previous experiments, the PU dataset contains real damage such as fatigue spalling, with more complex signal components and stronger background noise.

[0075] The data preprocessing strategy remained consistent: the sample length was 1024, the image size generated by GADF encoding was consistent, and three input modalities were constructed: time domain, frequency domain, and time-frequency domain. Table 4 summarizes the performance comparison of the PU dataset models: Table 4 As shown in Table 4, the model maintains extremely high diagnostic performance on the PU dataset. In the time-frequency domain, the model achieves an extremely high accuracy of 99.73% and converges rapidly. This not only proves that AVS-LCNN can effectively overcome the strong noise interference in the PU dataset, but also shows that it maintains state-of-the-art diagnostic capabilities under different damage mechanisms (artificial vs. real). Compared to the fluctuations in time-domain accuracy in the main experiment, the time-domain accuracy in this experiment remained stable at 96.40%. This indicates that GADF encoding, combined with a variable-scale structure, can extract discriminative features from the complex impact textures generated by real damage.

[0076] By analyzing the triggering distribution of the dynamic weight α in this experiment, the model exhibits a highly stable "minimalist inference strategy." This strategy highlights the two core advantages of variable scaling and dynamic weight design: 1. "Universal Dominance" of Variable-Scale Features: Regardless of whether it's in the time domain or the time-frequency domain, the trigger rate of Layer 3 (ASPP) consistently remains above 93%. This strongly demonstrates the necessity of multi-scale feature extraction. Real fatigue damage signals simultaneously contain high-frequency transient impacts (microscopic erosion) and low-frequency periodic modulations (macroscopic spalling). Fixed-scale convolutional kernels struggle to handle both simultaneously, while the ASPP module successfully decouples these complex hybrid modes through parallel multipath dilated convolutions. The high activation rate of Layer 3 indicates that it is the "core engine" for the model to cope with complex operating conditions.

[0077] 2. Intelligent Structural Pruning of Dynamic Weights: In the time-frequency domain modality, the trigger rate of Layer 2 is 0.0%, while that of Layer 1 is only 3.2%. This phenomenon is highly consistent with the performance of Experiment 1 (HTBF) in the optimal modality. It reveals the deep advantage of the dynamic weight mechanism—automatic identification of redundancy. When the dual-channel input provides sufficiently clear features, the model intelligently determines that the feature expansion of the lower layers (Layer 1 / 2) is redundant, thus automatically "closing" these layers. Efficiency Verification: This dynamic path of "skipping the shallow layers and directly addressing deep semantics" is essentially a sample-level network pruning. It allows the model to maintain extremely low computational overhead (training only takes 51.34s) when processing more complex real-world data, achieving a win-win situation of inference efficiency and accuracy, which are crucial in industrial edge deployments.

[0078] Supplementary experiments on the PU dataset further validated the superiority of AVS-LCNN. Faced with real-world damage data, the model demonstrated the irreplaceable role of the variable-scale ASPP module in extracting complex features through sustained high activation in Layer 3; simultaneously, through sustained suppression in Layers 1 and 2, it showcased the intelligent characteristics of the dynamic weight mechanism in eliminating redundant computations and optimizing inference paths. This adaptive capability of "complex when necessary (multi-scale aggregation) and simple when necessary (automatic pruning)" is precisely the core competitive advantage that distinguishes this method from traditional static CNNs.

[0079] Comparative experiment: To comprehensively evaluate the performance of the proposed AVS-LCNN, this embodiment includes comparative experiments. The comparisons primarily focus on three aspects: diagnostic accuracy, number of parameters, and training time.

[0080] To verify the advantages of AVS-LCNN in "lightweight design" and "high-precision diagnosis", this embodiment selects six representative deep learning models as benchmarks. All models are trained and tested under the same experimental conditions, using three feature input types: time domain, frequency domain, and time-frequency domain.

[0081] Table 5 summarizes the classification accuracy (mean ± standard deviation) of seven independent experimental runs. The results show that AVS-LCNN exhibits a significant competitive advantage, especially in the frequency and time-frequency domains, both of which are crucial for fault diagnosis.

[0082] Table 5 As shown in Table 5, MSCNN performs poorly, especially in the time-frequency domain where its accuracy is only 77.55%. In contrast, AVS-LCNN achieves an accuracy of 99.41%, representing an improvement of over 21%. This indicates that the proposed dynamic variable-scale feature extraction method significantly outperforms simple static convolution stacking. Compared to attention models, AVS-LCNN consistently outperforms SKNet in all three domains (e.g., frequency domain: 99.22% vs 98.85%). This demonstrates that the proposed gating mechanism is more effective than SKNet's segmentation attention mechanism in filtering out noise. Most notably, AVS-LCNN's accuracy has reached a level comparable to heavyweight networks. In the time-frequency domain, its accuracy (99.41%) is slightly different from that of Pyramid Network (99.53%) and CBAM-ResNet (99.34%).

[0083] Training times for large models (ResNet-18, PyramidNet, CBOM-ResNet) range from 540 to 835 seconds. In contrast, AVS-LCNN takes only about 100 seconds, representing a 5 to 8-fold improvement in training efficiency. Although MSCNN is the fastest (about 44 seconds), its lower accuracy makes it difficult to apply in practice.

[0084] According to Table 5, the number of parameters in AVS-LCNN is only 0.17M (170,929), which is about 1 / 65 of that of PyramidNet (11.02M) and ResNet-18 (11.17M). Even compared with Inception (0.91M) and SKNet (0.37M), the model size of AVS-LCNN is reduced by 81% and 53%, respectively.

[0085] The network structure parameters of AVS-LCNN are shown in Table 6: Table 6 On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0086] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0087] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An adaptive variable-scale rolling bearing fault diagnosis method based on lightweight CNN, characterized in that, Includes the following steps: The original vibration signal is acquired, and a multimodal input feature map is constructed based on the original vibration signal. The multimodal input feature map is input into a lightweight backbone network for feature extraction to obtain basic features; The basic features are input into a cascaded adaptive variable-scale feature extraction module, which includes multiple cascaded variable-scale feature extraction units. Each variable-scale feature extraction unit generates corresponding sample-level dynamic weight coefficients based on the input basic features, and adaptively performs multi-scale feature extraction or identity mapping operations on the input basic features based on these dynamic weight coefficients to obtain output features. The output features of the previous unit serve as the input features of the next unit, and after step-by-step processing, the enhanced features are finally obtained. The fault diagnosis results are output based on the fusion features.

2. The method according to claim 1, characterized in that, Constructing a multimodal input feature map based on the original vibration signal specifically includes: The original vibration signal is preprocessed, and two-dimensional feature representations of the Gram angle difference field are constructed in the time domain and frequency domain, respectively. The two-dimensional feature representations of Gram angle difference fields in the time and frequency domains are fused to form a multimodal input feature map.

3. The method according to claim 1, characterized in that, The multimodal input feature map is input into a lightweight backbone network for feature extraction to obtain basic features, specifically including: Intermediate features are obtained by downsampling the multimodal input feature map through cascaded standard convolutional layers and pooling layers; The intermediate features are extracted by using depthwise separable convolutional layers, and a channel attention mechanism is embedded in the depthwise separable convolutional layers to adaptively weight the channel features to obtain the basic features.

4. The method according to claim 1, characterized in that, Each variable-scale feature extraction unit generates corresponding sample-level dynamic weight coefficients based on the input basic features, specifically including: Global average pooling is performed on the basic input features in the spatial dimension to obtain a channel-level global feature description vector. The channel-level global feature description vector is input into the weight prediction network to generate dynamic weight coefficients between 0 and 1.

5. The method according to claim 1, characterized in that, Each variable-scale feature extraction unit adaptively performs multi-scale feature extraction or identity mapping operations on the input basic features based on dynamic weight coefficients to obtain output features, specifically including: Multi-scale feature extraction is performed on the input basic features to obtain intermediate enhanced features; Based on the dynamic weight coefficient, the input basic features and intermediate enhancement features are weighted and fused. The fusion weight of the input basic features is 1 minus the dynamic weight coefficient, and the fusion weight of the intermediate enhancement features is the dynamic weight coefficient. The output features are obtained after weighted fusion.

6. The method according to claim 5, characterized in that, The cascaded adaptive variable-scale feature extraction module comprises four cascaded variable-scale feature extraction units, wherein: The first variable-scale feature extraction unit extracts multi-scale features from the input features using dilated convolutions with a first dilation rate. The second variable-scale feature extraction unit performs multi-scale feature extraction by using dilated convolution with a second dilation rate on the input features, where the second dilation rate is greater than the first dilation rate. The third variable-scale feature extraction unit uses the hollow spatial pyramid pooling module to extract multi-scale features from the input features; The fourth variable-scale feature extraction unit performs feature smoothing and calibration on the input features using dilated convolution with a first dilation rate.

7. The method according to claim 6, characterized in that, The hollow spatial pyramid pooling module includes multiple parallel dilated convolution branches with different dilation rates and a global average pooling branch. The outputs of each branch are concatenated along the channel dimension to form a multi-scale aggregated feature.

8. The method according to claim 1, characterized in that, The fault diagnosis results are output based on the fusion features, specifically including: Global average pooling is performed on the fused features to obtain the global feature vector; The global feature vector is input into the fully connected layer, and the probability distribution of each category is output through the normalized exponential function. The category corresponding to the highest probability is taken as the fault diagnosis result.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images