Fault diagnosis method for rolling bearings based on three-path adaptive time-frequency fusion network with few samples

Abstract

The present application relates to the technical field of fault diagnosis, in particular to a few sample rolling bearing fault diagnosis method based on three-path adaptive time-frequency fusion network, the method comprises collecting rolling bearing vibration signal, constructing sample; in the training stage, randomly selecting noise segment in the set amplitude range with a preset probability to perform pulse erasing or masking, data enhancement is carried out; dynamic initialization driven by physical priori; the output characteristics of three paths are unified time sequence length; adaptive denoising feature is reconstructed; cross feature recalibration between paths; three-way recalibration features are weighted and fused, and the time sequence feature channel is reweighted; the output rolling bearing each health state, fault type prediction probability, and the fault diagnosis accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and specifically to a method for diagnosing rolling bearing faults based on a three-path adaptive time-frequency fusion network with a small number of samples. Background Technology

[0002] Localized faults in rolling bearings generate periodic impact pulses, which excite high-frequency resonant carriers at the bearing's natural frequencies. Simultaneously, the rotational frequency and its harmonics modulate the amplitude and create sidebands on the resonant carriers. Traditional diagnostic methods typically employ signal processing techniques such as Fourier transform, wavelet analysis, and envelope demodulation to extract features, followed by manually setting thresholds or using classifiers like support vector machines for identification. These methods rely on expert prior knowledge and manual feature selection, limiting their adaptability to non-stationary signals, complex operating conditions, and strong noise.

[0003] With the development of deep learning, one-dimensional convolutional neural networks, recurrent neural networks, and their combined models have been used for end-to-end fault diagnosis. Existing methods with limited sample sizes mainly employ data augmentation or generative models to expand the sample size, regularization and lightweight structures to suppress overfitting, or transfer learning and meta-learning to reuse knowledge. Multi-scale convolutions have also been used to obtain features from different receptive fields.

[0004] Existing technologies suffer from the following main drawbacks: high cost of acquiring fault samples and uneven class distribution; conventional deep networks are prone to memorizing noise and overfitting when there are only a few dozen samples per class, resulting in insufficient diagnostic accuracy and stability during repeated training; most existing multi-scale convolutions only change the kernel size, failing to divide the work according to the fault generation mechanism of "impact-resonance-modulation" and failing to explicitly model the coupling relationship between various physical components, leading to low learning efficiency with a small number of samples; convolution kernels are usually randomly initialized, requiring blind searching in a wide frequency domain during the initial training phase; when data is insufficient, convolution kernels tend to focus on random noise bands; recurrent neural networks rely on serial recursion, resulting in a large number of parameters and low parallel efficiency, making them more prone to overfitting in small sample scenarios, which is not conducive to deployment on edge devices or online systems; strong background noise can form pseudo-impacts and pseudo-spectral peaks, and fixed filtering or uniform gating cannot adaptively adjust with changes in the sample signal-to-noise ratio, easily losing weak fault information or retaining excessive noise. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method for diagnosing rolling bearing faults based on a three-path adaptive time-frequency fusion network with a small number of samples.

[0006] This invention is achieved through the following technical solution: a method for diagnosing rolling bearing faults based on a three-path adaptive time-frequency fusion network using a small number of samples, comprising the following steps: S1. Collect one-dimensional acceleration vibration signals of rolling bearings under operating conditions and construct samples; S2. During the training phase of the three-path adaptive time-frequency fusion network, noise segments within a set amplitude range are randomly selected with a preset probability to perform pulse erasure or masking for data augmentation. S3. Perform dynamic initialization driven by physical priors; S4. Construct three parallel depthwise separable one-dimensional convolutional paths with different functions, and unify the temporal length of the output features of the three paths; S5. Perform singular value decomposition on the feature matrix of each path, divide the signal subspace and noise subspace according to the singular value ratio, and reconstruct the adaptive denoising features. S6. Calculate the global description for the three denoised features respectively, construct a third-order physical coupling matrix, and realize the recalibration of cross features between paths through physical adjacency initialization and sparse regularization constraints. S7. Weighted fusion of three-way recalibrated features, the fused features are input into multi-layer dilated causal convolution to model long temporal dependencies in parallel, and then the temporal feature channels are reweighted through ECA channel attention; S8. Global average pooling of the attention-weighted temporal features yields a fixed-dimensional feature vector, which is then passed through a fully connected layer and the SoftMax function to output the predicted probabilities of each health state and fault type of the rolling bearing.

[0007] Preferably, in S1, a vibration sensor is arranged on the bearing housing or an adjacent structure to collect the one-dimensional acceleration signal of the rolling bearing during operation; The continuous acceleration signal is segmented into batch samples according to a sliding window. Where B is the batch size and L is the sample length.

[0008] Preferably, in S2, during the training phase of the three-path adaptive time-frequency fusion network, the data augmentation is represented as follows: ，

[0009] ; In the formula, To enhance the amplitude of the i-th sample at time t; To enhance the preceding amplitude; t is a binary mask that takes values ​​of 0 or 1; t is the sampling point number within the window; The erase probability has a value range of [0,1]. This is a small-amplitude threshold, and its unit is consistent with the input signal. This indicates the probability of an event occurring.

[0010] Preferably, S3 includes the following sub-steps: S3-1. Quickly calculate the power spectral density for each batch of input samples and extract at least three main peak frequencies; S3-2. Calculate the center frequency of each path and generate the corresponding parameterized basis filter; S3-3. Generate adaptive convolution kernels based on dynamic hybrid basis filters of global sample descriptors.

[0011] Preferably, in S3-1, the power spectral density is calculated using the following formula for a sample of length L: ; ; In the formula, The discrete Fourier complex coefficients of the i-th sample at the k-th frequency grid point; This corresponds to the power spectral density; Let be the t-th time-domain sampled value; t is the time-domain sample point index; k is the frequency grid point index; L is the sample length; j is the imaginary unit, satisfying... .

[0012] Preferably, the formula for calculating the center frequency of the m-th path in S3-2 is as follows: ; In the formula, Let m be the center frequency of the m-th path; The frequency of the main spectrum assigned to the m-th path; The initial value is 0, which is the learnable frequency offset. , These are the lower and upper limits of the prior frequency interval, respectively; m is the path number, which can be 1, 2, or 3; Clip() means to truncate the value within the specified interval; The parameterized basis filter is expressed as: ; In the formula, Let be the impulse response of the parameterized basis filter for the m-th path at time t; This is the set of parameters for the filter; The amplitude coefficient; The scale parameters for controlling pulse width and equivalent bandwidth are expressed in terms of available sampling points or seconds. t represents the center frequency; t represents the discrete time index or the converted time.

[0013] Preferably, the center frequency and scale parameters of the fundamental filter in S3-3 participate in end-to-end backpropagation, and the gated network generates dynamic coefficients based on the input global descriptor, which are represented as follows: ； ; .

[0014] In the formula, α is a vector composed of the mixed weights of each basis filter; W and b are the weight matrix and bias of the gated network, respectively; z is the global description vector of the input features; The non-negative normalized weights of the r-th basis filter; The kernel is a dynamically generated convolution kernel based on the samples; M is the number of basis filters. Let r be the r-th parameterized basis filter; r is the basis filter index; SoftMax is used to make all... The sum is 1.

[0015] Preferably, the three paths in S4 include the impact sensing path P1, the frequency domain modulation path P2, and the time-frequency joint path P3; Each path includes channel-wise depthwise convolution and pointwise convolution, and their operations are represented as follows: ; ; In the formula, This is the output of the c-th channel after depthwise convolution; The characteristics of the o-th output channel; Features of the c-th input channel; The q-th coefficient of the convolution kernel of the c-th channel; t represents the pointwise convolution weights mapped from input channel c to output channel o; t is the time index; q is the kernel position index; K is the kernel length; d is the dilation rate; c and o are the input and output channel indices, respectively; C is the number of input channels; The formulas for calculating the output length of each path are as follows: ; In the formula, is the length of the output sequence after one-dimensional convolution; L is the length of the input sequence; P is the padding amount used on both sides of the input; d is the dilation rate; K is the kernel length; s is the convolution stride; This indicates rounding down to the nearest integer.

[0016] Preferably, in S5, singular value decomposition is performed on the feature map of each path, and the boundary is determined according to the position of the singular spectrum change; After dividing the features into signal subspace and noise subspace, soft gating is used to obtain denoised features.

[0017] Preferably, the path intersection feature recalibration in S6 is represented as follows: ; In the formula, This is the output after cross-recalibration of the m-th path; , These are the input features for the m-th and n-th paths, respectively. , This is the channel modulation vector for the corresponding path; The weights are the cross-path residuals from path n to path m; m and n are path numbers and n ≠ m; This indicates element-wise multiplication.

[0018] Preferably, the prediction probability calculation formula in S8 is as follows: ; ; In the formula, The scalar value of the c-th channel after global average pooling; L represents the attention feature; L′ represents the time length. Let be the predicted probability that the sample belongs to the j-th class, with a value of [0,1]. and These are the weight vector and bias of the j-th classifier, respectively; g is the weight vector and bias of all... The feature vector is composed of C, which is the total number of fault categories; j and r are the category indices. , denoted as the weight vector and bias of the r-th classifier, respectively.

[0019] Preferably, when training the three-path adaptive time-frequency fusion network, a composite loss function is used for training. The composite loss includes cross-entropy classification loss, supervised contrastive prototype loss, and physical coupling regularization loss. Backpropagation synchronously updates all network parameters.

[0020] Compared with the prior art, the beneficial effects of the present invention are: This invention aims to overcome the aforementioned shortcomings of existing methods for diagnosing rolling bearing faults with a limited number of samples, and provides a method for diagnosing rolling bearing faults with a limited number of samples based on a three-path adaptive time-frequency fusion network. It leverages the physical mechanism of bearing faults to achieve decoupled extraction of vibration signals along different paths, increases information interaction between paths with physical constraints, adaptively adapts to noise environments with different signal-to-noise ratios, and completes long-term feature modeling with a lightweight parallel structure. Under conditions of scarce fault-labeled samples and strong background noise, it simultaneously improves fault diagnosis accuracy, training stability, and adaptability to edge device deployment.

[0021] The physical prior-driven dynamic initialization of this application constrains the center frequency of the convolution kernel by combining the input spectral peaks and the inherent frequency range of the device, reducing the broadband blind search in the random initialization stage and improving the efficiency and stability of small sample training.

[0022] The cross-path feature recalibration in this application achieves bidirectional supplementation between paths through a learnable coupling matrix that is subject to physical prior constraints, and explicitly utilizes the relationship of "impact-induced resonance and resonance positioning feedback to impact-based spurious removal".

[0023] The noise perception subspace gating adjusts the noise retention ratio according to the subspace structure of the sample or batch, and can adaptively compromise between different signal-to-noise ratios to enhance robustness under noisy conditions.

[0024] Multi-path fusion and lightweight temporal-channel attention aggregation replace recurrent structures with expanded causal convolutions, supporting parallel computation; combined with depthwise separable convolutions and ECA attention, it significantly reduces parameters while maintaining long-range dependency modeling capabilities, making it suitable for online monitoring and edge deployment.

[0025] The composite loss directly constrains the intra-class compactness and inter-class separation of the feature space, while maintaining the physical rationality of path coupling, further suppressing overfitting of a small number of samples. Attached Figure Description

[0026] Figure 1 This is the overall network architecture diagram of the three-path adaptive time-frequency fusion network based on this application. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0028] Example 1 Combination Figure 1 This invention starts from the generation chain of "periodic impact pulse - high-frequency resonant carrier - frequency conversion modulation sideband" of local faults in rolling bearings, constructs three parallel feature paths with clear physical division of labor, and uses power spectrum peaks to drive the dynamic initialization of each path with physical priors; performs soft gating of the noise subspace within the path, and performs cross-recalibration under physical coupling constraints between the paths; after fusion, parallel temporal modeling is completed through dilated causal convolution and channel attention, and finally outputs the fault category probability.

[0029] A method for diagnosing rolling bearing faults based on a three-path adaptive time-frequency fusion network with a small number of samples includes the following steps: S1. Collect one-dimensional acceleration vibration signals of rolling bearings under operating conditions and construct samples.

[0030] One-dimensional acceleration signals of rolling bearings during operation are acquired using vibration sensors mounted on the bearing housing or adjacent structures. The continuous acceleration signals are then segmented into batch samples using a sliding window. Where B is the batch size and L is the sample length.

[0031] Preferred sample length sliding step size .

[0032] Furthermore, the sample requirements can be divided into training set, validation set, and test set according to equipment or operating conditions.

[0033] The sliding window segmentation process is represented as follows: ; ; In the formula, Let be the sample vector of the i-th window; s(·) be the continuous acceleration vibration signal; i be the window index; H be the sliding step size; L be the window length; N be the total number of windows generated from the continuous signal; and T be the total number of sampling points of the continuous vibration signal. This expression clarifies the mapping relationship between the continuous signal and the fixed-length network input.

[0034] S2. During the training phase of the three-path adaptive time-frequency fusion network, noise segments within a set amplitude range are randomly selected with a preset probability to perform pulse erasure or masking for data augmentation.

[0035] This application only performs pulse erasure or masking on small-amplitude noise segments randomly selected with a preset probability during the training phase, reducing the network's dependence on local noise patterns and focusing on the overall distribution of periodic impacts. This enhancement operation is not performed during the testing phase. The enhancement phase is represented as follows: ，

[0036] ; In the formula, To enhance the amplitude of the i-th sample at time t; To enhance the preceding amplitude; t is a binary mask that takes values ​​of 0 or 1; t is the sampling point number within the window; The erase probability has a value range of [0,1]. This is a small-amplitude threshold, with the unit consistent with the input signal; This indicates the probability of an event occurring.

[0037] The erase probability control enhances the intensity, the small amplitude threshold is determined by the amplitude quantile, and the indicator function is used to limit the masked sampling points. This prioritizes masking small amplitude noise segments to avoid disrupting the main impact peak.

[0038] S3. Perform dynamic initialization driven by physical priors.

[0039] S3 includes the following sub-steps: S3-1. Quickly calculate the power spectral density for each batch of input samples and extract at least three main peak frequencies.

[0040] For a discrete sample of length L, the power spectral density is calculated using the following formula: ; ; In the formula, The discrete Fourier complex coefficients of the i-th sample at the k-th frequency grid point; This corresponds to the power spectral density; Let be the t-th time-domain sampled value; t is the time-domain sample point index; k is the frequency grid point index; L is the sample length; j is the imaginary unit, satisfying... .

[0041] S3-2. Calculate the center frequency of each path and generate the corresponding parameterized basis filter; The center frequency of the m-th path in S3-2 is preheated using the following formula: ; In the formula, Let m be the center frequency of the m-th path; The main spectral peak frequency (Hz) assigned to the m-th path; For learnable frequency offset; , These are the lower and upper limits of the prior frequency range, respectively; m is the path number, which can be 1, 2, or 3; Clip() means truncating the value within the specified range.

[0042] Among them, learnable frequency offset The initial value is 0, and the upper and lower limits of the frequency are... , The a priori range of natural frequencies determined by bearing and equipment parameters is given.

[0043] The parameterized basis filter is expressed as: ; In the formula, Let be the impulse response of the parameterized basis filter for the m-th path at time t; This is the set of parameters for the filter; The amplitude coefficient; The scale parameters for controlling pulse width and equivalent bandwidth are expressed in terms of available sampling points or seconds. t represents the center frequency; t represents the discrete time index or the converted time.

[0044] S3-3. Generate adaptive convolution kernels based on dynamic hybrid basis filters of global sample descriptors.

[0045] The center frequency and scaling parameters of the fundamental filter in S3-3 participate in end-to-end backpropagation. The gated network generates dynamic coefficients based on the input global descriptor, as shown below: ； ; .

[0046] In the formula, α is a vector composed of the mixed weights of each basis filter; W and b are the weight matrix and bias of the gated network, respectively; z is the global description vector of the input features; The non-negative normalized weights of the r-th basis filter; The kernel is a dynamically generated convolution kernel based on the samples; M is the number of basis filters. Let r be the r-th parameterized basis filter; r is the basis filter index; SoftMax is used to make all... The sum is 1.

[0047] This allows the convolution kernel to focus on the main energy band in the early stages of training and to adaptively combine with the spectral structure of different samples.

[0048] S4. Construct three parallel, depth-separable one-dimensional convolutional paths with different functions, and unify the temporal length of the output features of the three paths.

[0049] The three paths in S4 include the following: Impact sensing path P1: A depthwise separable one-dimensional convolution with a kernel size of 5, a stride of 1, and a dilation rate of 1 is used to maintain a high temporal resolution for extracting transient impact pulse sequences.

[0050] Frequency domain modulation path P2: A depthwise separable one-dimensional convolution with kernel size 31, stride 2, and dilation rate 3 is used to extract the high-frequency resonant carrier with a larger receptive field and smoothing effect.

[0051] The time-frequency joint path P3 employs a depthwise separable one-dimensional convolution with a kernel size of 15, a stride of 1, and an expansion rate of 2, simultaneously sensing the low-frequency transmodulation sideband and the high-frequency attenuation envelope.

[0052] The three paths in this application are not simply scaled side-by-side, but correspond to the impulse pulse, resonant carrier, and modulation sideband respectively. This allows a small number of samples to preferentially learn interpretable fault physical characteristics, reducing dependence on random noise patterns. The small convolutional kernel of P1 avoids excessive smoothing of the impulse peak; the large kernel and expansion rate of P2 expand the coverage area and reduce sensitivity to local noise; P3 uses a medium kernel to balance time localization and frequency band resolution. The three paths form a complementary combination of short, medium, and long receptive fields.

[0053] Each path includes channel-wise depthwise convolution and pointwise convolution, and their operations are represented as follows: ; ; In the formula, This is the output of the c-th channel after depthwise convolution; The characteristics of the o-th output channel; Features of the c-th input channel; The q-th coefficient of the convolution kernel of the c-th channel; t represents the pointwise convolution weights mapped from input channel c to output channel o; t is the time index; q is the kernel position index; K is the kernel length; d is the dilation rate; c and o are the input and output channel indices, respectively; C is the number of input channels.

[0054] The parameter relationship between depthwise separable convolution and standard convolution is as follows: ; In the formula, The number of learnable parameters for depthwise separable convolution; The number of learnable parameters for a standard convolution; K is the kernel length; , These represent the number of input channels and the number of output channels, respectively; all parameters are dimensionless counts.

[0055] The formulas for calculating the output length of each path are as follows: ; In the formula, is the length of the output sequence after one-dimensional convolution (sampling points); L is the length of the input sequence; P is the padding amount (sampling points) used on both sides of the input; d is the dilation rate; K is the kernel length; s is the convolution stride; This indicates rounding down to the nearest integer.

[0056] Adaptive pooling or equivalent size transformation is performed on the three outputs to ensure that their time dimensions are consistent. The kernel size, step size, and expansion rate mentioned above are preferred values, but neighboring values ​​can also be selected while maintaining the differences in receptive field and physical division of labor among the three paths.

[0057] S5. Perform singular value decomposition on the feature matrix of each path, divide the signal subspace and noise subspace according to the singular value ratio, reconstruct the adaptive denoising features, and complete the noise perception subspace gating (NSG).

[0058] In S5, singular value decomposition is performed on the feature map of each path, and the boundary is determined based on the location of the singular spectrum change, for example: ; ; In the formula, Let m be the feature matrix of the m-th path; , These are the left and right singular vector matrices, respectively; It is a singular value diagonal matrix; Let be the i-th singular value in descending order; i is the singular value index; k is the boundary order between the signal subspace and the noise subspace; m is the path number; the superscript T indicates matrix transpose; argmax indicates the index that maximizes the adjacency ratio of the singular values.

[0059] The formulas for calculating the signal subspace and noise subspace are as follows: ; ; In the formula, The signal subspace reconstruction matrix for the m-th path; Reconstruct the matrix for the noise subspace; Let i be the i-th singular value; , These are the corresponding left and right singular vectors; k is the subspace boundary order; r is... The rank or the total number of singular values ​​retained; i is the singular component index; m is the path number.

[0060] After dividing the features into signal subspace and noise subspace, soft gating is used to obtain the denoising features, which are represented as follows:

[0061] In the formula, The denoised features of the m-th path after gating; and These are the signal subspace and the noise subspace, respectively. is the soft-preservation coefficient for the noise subspace of the m-th path, with a value range of [0,1].

[0062] The soft gating coefficients are obtained from learnable parameters via a sigmoid mapping or generated from signal-to-noise statistics. These coefficients are decreased under low signal-to-noise ratio (SNR) conditions to suppress the noise subspace, and increased under high SNR conditions to preserve complete information. Only a small number of gating parameters are required for the three paths.

[0063] S6. Calculate the global description for each of the three denoising features and construct a third-order physical coupling matrix.

[0064] The third-order physical coupling matrix described in this application is a 3-row × 3-column learnable coupling gain matrix corresponding to three feature paths. The off-diagonal elements of the matrix represent the information transfer gain between different paths. The matrix parameters are initialized with physical adjacency and sparse regularization constraints are applied. The calculation is completed in combination with the global descriptors of each path, thereby realizing the recalibration of cross-features between paths.

[0065] The positive influence of the impact path on the resonance path and the joint path is preferentially preserved, while the reverse gain is constrained but can be learned.

[0066] The recalibration process is represented as follows: ; ; ; In the formula, Features of the m-th path The description vector obtained by global average pooling; The description vector after incorporating information from other paths; The coupling gain from the nth path to the mth path; The vector describing the nth path; Let be the channel modulation vector of the m-th path, with elements taking values ​​of [0,1]; GAP represents global average pooling. Let m be a multilayer perceptron for the m-th path.

[0067] The final inter-path cross-feature recalibration is represented as follows: ; In the formula, This is the output after cross-recalibration of the m-th path; , These are the input features for the m-th and n-th paths, respectively. , This is the channel modulation vector for the corresponding path; The weights are the cross-path residuals from path n to path m; m and n are path numbers and n ≠ m; This indicates element-wise multiplication.

[0068] The physical constraint terms of the coupling matrix can be expressed as: ; In the formula, For physical coupling regularization loss; This is the learnable path coupling matrix (i.e., the third-order physical coupling matrix). The physical adjacency prior matrix is ​​set based on the impact-resonance-modulation relationship; , The positive regularization coefficient is non-negative. This represents the L1 norm of the sum of the absolute values ​​of the matrix elements. The physical adjacency matrix is ​​a pre-defined prior. The cross-path residual weights are preferably initialized to 0. PCR enhances the impact characteristics and resonance band selection, while simultaneously feeding back resonance frequency domain positioning information to the impact path to suppress noise spurious peaks.

[0069] S7. Weighted fusion of three-way recalibrated features, the fused features are input into multi-layer dilated causal convolution to model long temporal dependencies in parallel, and then the channel weights of temporal features are adaptively redistributed through ECA channel attention; thus, multi-path fusion and lightweight temporal-channel attention aggregation are achieved.

[0070] The three recalibration features are element-wise summed or weighted and fused, as shown below: ; In the formula, This is a feature of multi-path fusion; This is the output of the m-th path after recalibration; is the non-negative normalized fusion weight of the m-th path; m is the path number, which can be 1, 2, or 3.

[0071] When performing additive fusion, the weights of each path can be the same or normalization can be omitted.

[0072] The fused features are sequentially processed through three layers of dilated causal convolution, the first of which... The layer output is: ; In the formula, For the first The layer's output at time t; Features of the previous layer; For the first The q-th coefficient of the convolutional kernel; t is the time index; q is the kernel position index; K is the kernel length; d_ For the first Layer expansion rate; The network layer number; BN represents batch normalization; This represents the parameterized linear rectified activation function.

[0073] The three-layer architecture increases the number of channels from 32 to 64, provides an effective receptive field covering 29 time steps, and supports parallel computation across the entire sequence.

[0074] Then, ECA channel attention is used, which is represented as follows: ; In the formula, This is the average descriptive value of the c-th channel along the time dimension; L′ represents the value of the time-series aggregated feature at channel c and time t; c is the channel index; t is the time index; and L′ is the time length of the feature sequence.

[0075] The features of the channel after reweighting are represented as follows: ; ; In the formula, ω is a vector composed of attention weights for each channel, with elements taking values ​​of [0,1]; z is a vector composed of description values ​​for all channels; and k is the length of the one-dimensional ECA convolution kernel. These are the characteristics after channel reweighting; The temporal characteristics before reweighting; This represents a one-dimensional convolution with a kernel length of k; This indicates element-wise multiplication by channel.

[0076] S8. Global average pooling of the attention-weighted temporal features yields a fixed-dimensional feature vector. This vector is then passed through a fully connected layer and the SoftMax function to output the predicted probabilities of each health state and fault type of the rolling bearing, thus completing feature aggregation and classification.

[0077] In S8, the time-series features reweighted along the time dimension are subjected to global average pooling to obtain a fixed-length vector. This vector is then input into a fully connected layer and passed through SoftMax to obtain the probability of each health state or fault category, as shown below: ; ; In the formula, The scalar value of the c-th channel after global average pooling; L represents the attention feature; L′ represents the time length. Let be the predicted probability that the sample belongs to the j-th class, with a value of [0,1]. and These are the weight vector and bias of the j-th classifier, respectively; g is the weight vector and bias of all... The feature vector is composed of C, which is the total number of fault categories; j and r are the category indices. , denoted as the weight vector and bias of the r-th classifier, respectively.

[0078] Categories can include normal, inner ring failure, outer ring failure, rolling element failure, and their different damage sizes or operating conditions. Output the category corresponding to the highest probability, or output the result to be reviewed when the probability is below a threshold.

[0079] This application employs a composite loss function for training a three-path adaptive time-frequency fusion network. The composite loss includes cross-entropy classification loss, supervised contrastive prototype loss, and physical coupling regularization loss. Backpropagation synchronously updates all network parameters.

[0080] The composite loss function is expressed as follows: ; In the formula, This represents the total training loss of the network. Cross-entropy is used for classification loss; For supervised comparison of prototype loss; For physical coupling regularization loss; , These are the non-negative weighting coefficients for prototype loss and coupling loss, respectively.

[0081] ; In the formula, B is the average cross-entropy loss of the batch; B is the batch size; C is the total number of classes. The true label indicator value (0 or 1) for the i-th sample belonging to the j-th class; is the corresponding predicted probability; i is the sample index; j is the class index.

[0082] ; In the formula, B is the batch average prototype comparison loss; B is the batch size; C is the total number of categories. Let be the embedded feature vector of the i-th sample; Its true category The prototype vector; is the prototype vector of the j-th class, formed by the mean of the features of samples in that class; τ is the cosine similarity; τ is the temperature coefficient greater than 0; i is the sample index; j is the category index.

[0083] The category prototypes are formed by the mean features of similar samples, and cosine similarity is used for similarity. A temperature coefficient controls the smoothness of the contrastive learning distribution. The prototype loss brings similar samples closer together and pushes away dissimilar samples. The physical coupling regularization term constrains the directionality and sparsity of the coupling matrix. Backpropagation is used to jointly update the convolution, dynamic frequency offset, subspace gating, path coupling, attention, and classification parameters.

[0084] Example 2 This embodiment, based on Embodiment 1, uses a sliding window sampling method to construct the sample set, with a sliding step size of 400 and a sample length of 1024. The dataset includes normal, inner race fault, outer race fault, and rolling element faults, and includes different fault sizes.

[0085] The dataset was divided into four groups: A (0HP), B (1HP), C (2HP), and D (3HP) according to different load conditions. In this embodiment, group B (1HP) was used as the main experimental dataset. 60 training samples (accounting for 30% of the total samples) were randomly selected for each type of fault state, and the rest were used as test samples to establish a small number of sample diagnostic scenarios.

[0086] This embodiment is based on the PyTorch 1.8.0 framework and runs on a 12th Gen Intel(R) Core(TM) i5-12400 2.50GHz processor. The model uses the following configuration: the optimizer is AdamP, the learning rate is set to 0.001, the batch size is set to 32, the maximum number of iterations is 150, an early stopping strategy (patience=10) is adopted, the dropout rate is set to 0.2, and the weight decay coefficient is 0.0001.

[0087] To verify the diagnostic performance of the proposed network model under limited sample conditions, five scenarios were set up on dataset B with training sample ratios ranging from 0.1 (20 samples) to 0.5 (100 samples). The proposed method was compared with other methods. The comparison methods included traditional machine learning methods (PCA-SVM), classic deep learning methods (1D-CNN), a combination of convolutional and recurrent networks (CNN-BiGRU), and a limited-sample fault diagnosis method (DLWCB). All experiments were repeated five times, and the mean accuracy and standard deviation were recorded. The results are shown in Table 1.

[0088] Table 1. Model diagnostic results under limited sample conditions.

[0089] With only 20 training samples per class, the method in this application achieves an accuracy of 89.12%, which is 1.56%, 16.97%, 18.89%, and 28.76% higher than DLWCB, CNN-BiGRU, 1D-CNN, and PCA-SVM, respectively. This shows that by introducing the physical prior of the fault, the model can form an effective discrimination boundary without relying on a large number of samples.

[0090] When the number of samples per class increased to 40, 60, and 80, the accuracy of this application reached 95.67%, 98.54%, and 99.12%, respectively, which is an improvement of 1.52%, 0.98%, and 0.76% compared with the best comparative method DLWCB at the same stage. As the amount of data increases, the gap between different deep models gradually narrows, but this method still maintains its leading position, indicating that its advantage does not come from a single training scale.

[0091] When the number of samples in each class reaches 100, the accuracy of this application reaches 99.75%, which is 0.17% higher than DLWCB. Its standard deviation decreases from 0.48% with 20 samples to 0.01%, and is the lowest or near lowest among the comparison methods in scenarios with 40 to 100 samples, demonstrating better stability in repeated training. The overall accuracy of PCA-SVM is relatively low in various scenarios, indicating that the features are not enough to fully describe the coupling of impact, resonance and modulation. Although CNN-BiGRU can model temporal relationships, it only achieves 72.15% accuracy with 20 samples, indicating that high-parameter recurrent structures are more prone to overfitting with very few samples.

[0092] The above description is merely an optional embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the content of the present invention under the concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for diagnosing rolling bearing faults based on a three-path adaptive time-frequency fusion network using a small number of samples, characterized in that, Includes the following steps: S1. Collect one-dimensional acceleration vibration signals of rolling bearings under operating conditions and construct samples; S2. During the training phase of the three-path adaptive time-frequency fusion network, noise segments within a set amplitude range are randomly selected with a preset probability to perform pulse erasure or masking for data augmentation. S3. Perform dynamic initialization driven by physical priors; S4. Construct three parallel depthwise separable one-dimensional convolutional paths with different functions, and unify the temporal length of the output features of the three paths; S5. Perform singular value decomposition on the feature matrix of each path, divide the signal subspace and noise subspace according to the singular value ratio, and reconstruct the adaptive denoising features. S6. Calculate the global description for the three denoised features respectively, construct a third-order physical coupling matrix, and realize the recalibration of cross features between paths through physical adjacency initialization and sparse regularization constraints. S7. Weighted fusion of three-way recalibrated features, the fused features are input into multi-layer dilated causal convolution to model long temporal dependencies in parallel, and then the temporal feature channels are reweighted through ECA channel attention; S8. Global average pooling of the attention-weighted temporal features yields a fixed-dimensional feature vector, which is then passed through a fully connected layer and the SoftMax function to output the predicted probabilities of each health state and fault type of the rolling bearing.

2. The method for diagnosing rolling bearing faults based on a three-path adaptive time-frequency fusion network according to claim 1, characterized in that, In S1, vibration sensors are arranged on the bearing housing or adjacent structures to collect one-dimensional acceleration signals during the operation of the rolling bearing. The continuous acceleration signal is segmented into batch samples according to a sliding window. Where B is the batch size and L is the sample length.

3. The method for diagnosing rolling bearing faults based on a three-path adaptive time-frequency fusion network according to claim 1, characterized in that, In S2, during the training phase of the three-path adaptive time-frequency fusion network, the data augmentation representation is as follows: ， ； In the formula, To enhance the amplitude of the i-th sample at time t; To enhance the preceding amplitude; A binary mask that takes the values ​​0 or 1; t is the sampling point number within the window; The erase probability has a value range of [0,1]. This is a small-amplitude threshold, and its unit is consistent with the input signal. This indicates the probability of an event occurring.

4. The method for diagnosing rolling bearing faults based on a three-path adaptive time-frequency fusion network according to claim 1, characterized in that, S3 includes the following sub-steps: S3-1. Quickly calculate the power spectral density for each batch of input samples and extract at least three main peak frequencies; S3-2. Calculate the center frequency of each path and generate the corresponding parameterized basis filter; S3-3. Generate adaptive convolution kernels based on dynamic hybrid basis filters of global sample descriptors.

5. The method for diagnosing rolling bearing faults based on a three-path adaptive time-frequency fusion network according to claim 4, characterized in that, In S3-1, the power spectral density is calculated using the following formula: ； ； In the formula, The discrete Fourier complex coefficients of the i-th sample at the k-th frequency grid point; This corresponds to the power spectral density; Let be the t-th time-domain sampled value; t is the time-domain sample point index; k is the frequency grid point index; L is the sample length; j is the imaginary unit, satisfying... ; The formula for calculating the center frequency of the m-th path in S3-2 is as follows: ； In the formula, Let m be the center frequency of the m-th path; The frequency of the main spectrum assigned to the m-th path; The initial value is 0, which is the learnable frequency offset. , These are the lower and upper limits of the prior frequency interval, respectively; m is the path number, which can be 1, 2, or 3; Clip() means to truncate the value within the specified interval; The parameterized basis filter is expressed as: ； In the formula, Let be the impulse response of the parameterized basis filter for the m-th path at time t; This is the set of parameters for the filter; The amplitude coefficient; The scale parameters for controlling pulse width and equivalent bandwidth are expressed in terms of available sampling points or seconds. is the center frequency; t is the discrete-time index or the converted time. The center frequency and scaling parameters of the fundamental filter in S3-3 participate in end-to-end backpropagation. The gated network generates dynamic coefficients based on the input global descriptor, as shown below: ； ； ； In the formula, α is a vector composed of the mixed weights of each basis filter; W and b are the weight matrix and bias of the gated network, respectively; z is the global description vector of the input features; The non-negative normalized weights of the r-th basis filter; The kernel is a dynamically generated convolution kernel based on the samples; M is the number of basis filters. Let r be the r-th parameterized basis filter; r is the basis filter index; SoftMax is used to make all... The sum is 1.

6. The method for diagnosing rolling bearing faults based on a three-path adaptive time-frequency fusion network according to claim 1, characterized in that, S4 consists of three paths: impulse sensing path P1, frequency domain modulation path P2, and time-frequency joint path P3. Each path includes channel-wise depthwise convolution and pointwise convolution, and their operations are represented as follows: ； ； In the formula, This is the output of the c-th channel after depthwise convolution; The characteristics of the o-th output channel; Features of the c-th input channel; The q-th coefficient of the convolution kernel of the c-th channel; These are the pointwise convolution weights that map from input channel c to output channel o; t is the time index; q is the kernel position index; K is the kernel length; d is the dilation rate; c and o are the input and output channel indices, respectively; C is the number of input channels; The formulas for calculating the output length of each path are as follows: ； In the formula, is the length of the output sequence after one-dimensional convolution; L is the length of the input sequence; P is the padding amount used on both sides of the input; d is the dilation rate; K is the kernel length; s is the convolution stride; This indicates rounding down to the nearest integer.

7. The method for diagnosing rolling bearing faults based on a three-path adaptive time-frequency fusion network according to claim 1, characterized in that, In S5, singular value decomposition is performed on the feature map of each path, and the boundary is determined based on the position of the singular spectrum change. After dividing the features into signal subspace and noise subspace, soft gating is used to obtain denoised features.

8. The method for diagnosing rolling bearing faults based on a three-path adaptive time-frequency fusion network according to claim 1, characterized in that, The inter-path crossover feature recalibration in S6 is represented as follows: ； In the formula, This is the output after cross-recalibration of the m-th path; , These are the input features for the m-th and n-th paths, respectively. , This is the channel modulation vector for the corresponding path; The weights are the cross-path residuals from path n to path m. m and n are path numbers and n ≠ m; This indicates element-wise multiplication.

9. The method for diagnosing rolling bearing faults based on a three-path adaptive time-frequency fusion network according to claim 1, characterized in that, The formula for calculating the prediction probability in S8 is as follows: ； ； In the formula, The scalar value of the c-th channel after global average pooling; L represents the attention feature; L′ represents the time length. Let be the predicted probability that the sample belongs to the j-th class, with a value of [0,1]. and These are the weight vector and bias of the j-th classifier, respectively; g is the weight vector and bias of all... The feature vector is composed of C, which is the total number of fault categories; j and r are the category indices. , denoted as the weight vector and bias of the r-th classifier, respectively.

10. The method for diagnosing rolling bearing faults based on a three-path adaptive time-frequency fusion network according to claim 1, characterized in that, When training the three-path adaptive time-frequency fusion network, a composite loss function is used, which includes cross-entropy classification loss, supervised contrastive prototype loss, and physical coupling regularization loss. Backpropagation updates all network parameters synchronously.

Images