A method for diagnosing a fault of a rolling bearing based on acoustic framing

Abstract

The application discloses a rolling bearing fault diagnosis method based on acoustic framing. Firstly, the collected acoustic signal of the rolling bearing is subjected to Hilbert transform to construct a double-channel analytic representation containing instantaneous amplitude and instantaneous phase information. Then, the analytic signal is divided into short-time frame segments by the acoustic framing embedding module according to a preset frame length and step, and a uniform dimension frame-level embedding vector is generated by linear mapping, so as to highlight the local transient impact feature and effectively compress the sequence length. An improved lightweight Transformer encoder is adopted for timing feature extraction, the encoder combines a low-dimensional multi-head self-attention mechanism and a feedforward network based on SiLU activation and deep convolution, so as to simultaneously model the cross-frame global dependence and the short-range local structure, and finally, the intelligent identification of the multi-class faults of the rolling bearing is completed through the global feature vector of the class token.

Description

TECHNICAL FIELD

[0001] The present application relates to a mechanical equipment state monitoring and intelligent fault diagnosis method, in particular to a rolling bearing fault diagnosis method based on acoustic frame division. BACKGROUND

[0002] Rolling bearings are the core components in rotating machinery, and their running state directly affects the stability and reliability of the equipment. In order to realize early fault identification, traditional methods are mostly based on vibration signals for monitoring, but require contact installation, limited layout position, and are difficult to deploy under special working conditions. Acoustic signals have the advantages of non-contact acquisition, convenient installation, sensitivity to external structure, etc., and are suitable for rolling bearing state perception under complex working conditions. However, acoustic signals are easily disturbed by environmental noise, structure reflection and background noise during air propagation, showing obvious non-stationarity and weak fault pulse characteristics, and the amplitude and phase are easily disturbed, resulting in short-time, weak and scattered fault information, making it challenging to effectively extract acoustic fault features.

[0003] Existing acoustic fault diagnosis methods mostly rely on manually constructed time-domain, frequency-domain or time-frequency map features, which are difficult to stably characterize the short-time structure of fault pulses in noisy and significantly varying working conditions. Although deep learning methods can automatically extract features, convolutional neural networks mainly focus on local patterns when processing acoustic signals, and have limited modeling of long-range dependencies. In recent years, Transformer has shown good performance in time series modeling, but direct application to continuous acoustic waves is affected by factors such as large sequence length, weak transient impact and strong noise interference, making it difficult to simultaneously consider short-time local structure and long-time sequence correlation. The above shortcomings make it difficult to fully extract the fault features contained in the acoustic signal, and the diagnosis performance is unstable in multi-noise and multi-working condition environments. SUMMARY

[0004] The present application aims to solve the problems of weak fault pulse extraction, strong non-stationarity of signals, local transient structure easily covered by noise, and difficulty in simultaneously modeling local impact features and global time dependence in acoustic signal-based rolling bearing fault diagnosis. In addition, existing deep learning-based diagnosis models generally have complex structure, large number of parameters, and high computational overhead, making it difficult to deploy on resource-constrained edge devices in actual industrial scenarios, limiting real-time online monitoring and rapid inference. Therefore, there is an urgent need for an acoustic fault identification method that can maintain high diagnostic accuracy while being lightweight, low complexity and robust.

[0005] To solve the above technical problems, the application provides a rolling bearing fault diagnosis method based on acoustic framing, which is designed for the weak fault pulse, non-stationary and noise sensitivity characteristics of acoustic signals, and realizes high-precision identification of multiple fault states while keeping the model lightweight through analytic domain double-channel construction, short-time framing embedding and lightweight Transformer feature extraction.

[0006] A rolling bearing fault diagnosis method based on acoustic framing, comprising the following steps:

[0007] Acoustic signals of a rolling bearing in a normal state and multiple fault states during operation are collected;

[0008] A Hilbert transform is applied to the acoustic signals to construct analytic signals, obtaining real and imaginary parts corresponding to acoustic instantaneous energy changes and phase evolution, respectively, and forming a double-channel analytic acoustic signal;

[0009] The double-channel analytic acoustic signal is divided into multiple acoustic frame segments that can simultaneously represent local amplitude, phase and transient impact characteristics through sliding window segmentation;

[0010] Each acoustic frame segment is unfolded through vectorization and mapped to a fixed-dimensional embedding vector through linear projection, and all embedding vectors are superimposed with learnable position encoding, and a category token for aggregating global information is added at the head of the embedding vector, so as to form the input sequence of the Transformer;

[0011] The input sequence is sent to a Transformer encoder composed of a stack of multiple acoustic feature extraction modules, the acoustic feature extraction module includes a multi-head self-attention layer and a feedforward network layer, and residual connection and layer normalization are respectively arranged between the multi-head self-attention layer and the feedforward network layer;

[0012] Finally, the feature vector corresponding to the category token in the output sequence of the Transformer encoder is taken, and the rolling bearing fault category is output through the classifier after mapping.

[0013] Preferably, the Hilbert transform is applied to the acoustic signal to construct the analytic signal, and the real and imaginary parts corresponding to the acoustic instantaneous energy changes and phase evolution are obtained, and the double-channel analytic acoustic signal is formed, which specifically includes:

[0014] Let the collected real-valued acoustic signal be , and the Hilbert transform is applied to obtain the orthogonal component :

[0015] ;

[0016] wherein, is the Hilbert operator, denotes the Cauchy principal value, is the integral variable, is the time variable;

[0017] In the frequency domain, we have:

[0018] ;

[0019] wherein, denotes the Fourier transform, is the frequency, is the imaginary unit, is the spectrum of , is the sign function, and the real and imaginary parts form the complex analytic signal:

[0020] ;

[0021] wherein the instantaneous amplitude is:

[0022] ;

[0023] the instantaneous phase is:

[0024] ;

[0025] The acoustic signal is represented in the form of a vector of two-channel analytic acoustic signals:

[0026] ;

[0027] wherein, and denote the in-phase component and the quadrature component, respectively, denotes the transpose of a matrix or vector, is the time length of a single sample, i.e. the number of sampling points, denotes a real matrix with a feature dimension of 2 and a time length of .

[0028] Preferably, the two-channel analytic acoustic signal is divided by a sliding window to form a plurality of acoustic frame segments capable of simultaneously representing local amplitude, phase, and transient impact features, specifically including:

[0029] The analytic signal is divided into frames by a sliding window with a length of and a step size of . The th acoustic frame segment is represented as:

[0030] ;

[0031] wherein, is the starting sampling point, and

[0032] ;

[0033] is the frame length, is the step length, denotes the floor symbol.

[0034] Preferably, each acoustic frame segment is unfolded by vectorization and mapped to a fixed-dimension embedding vector through linear projection, specifically including:

[0035] for each acoustic frame a flattening operation is performed to obtain a vector:

[0036] ;

[0037] and mapped to a dimension embedding space through a linear layer:

[0038] ;

[0039] wherein, , are learnable projection parameters; each i.e. corresponds to a frame-level token, encoding the amplitude-phase local structure within the frame;

[0040] All frame embeddings are spliced to form a time-ordered token sequence:

[0041] .

[0042] Preferably, a learnable position encoding is superimposed on all embedding vectors, and a class token for aggregating global information is added at the head of the embedding vector to make it constitute the input sequence of the Transformer, specifically including:

[0043] A learnable class token vector is added at the head of the token sequence to obtain the final input sequence:

[0044] ;

[0045] A position encoding matrix of the same dimension as is constructed and element-wise added to the extended sequence to obtain the final input sequence ;

[0046] ;

[0047] wherein, The input sequence is defined as a positionally encoded input sequence.

[0048] Preferably, the input sequence is fed into a Transformer encoder composed of a stack of multi-layer acoustic feature extraction modules, which include multi-head self-attention layers and feed-forward network layers, and residual connections and layer normalization are respectively arranged between the multi-head self-attention layers and the feed-forward network layers; specifically including:

[0049] The input sequence is defined as a positionally encoded input sequence. The input of the first layer of the input sequence is linearly projected to obtain query, key and value matrices:

[0050] ;

[0051] wherein , , is the number of attention heads, is the low-dimensional subspace dimension of each head;

[0052] The scaled dot-product attention is calculated for each head respectively, and the combination is written as:

[0053] ;

[0054] wherein is the normalized attention weight matrix, representing the correlation strength between different frame-level tokens in the acoustic sequence; is the output matrix of the attention mechanism; The multi-head output is connected and restored to the model dimension; The context representation of the global inter-frame correlation is fused.

[0055] Preferably, residual connections and layer normalization are respectively arranged between the multi-head self-attention layers and the feed-forward network layers, specifically including:

[0056] ;

[0057] wherein LN represents the layer normalization operation; is the normalized sequence;

[0058] First, the input is expanded to the dimension through a linear layer: ;

[0059] wherein

[0060] ​​​​​​to extend the projection matrix; denotes concatenation along the channel dimension; followed by gated linear unit with SiLU activation SC-GLU:

[0061] ;

[0062] where, denotes element-wise multiplication, is a smooth non-linear function;

[0063] for applying depth-wise separable one-dimensional convolution along the sequence dimension:

[0064] ;

[0065] where, denotes depth-wise convolution operation using independent convolution kernels for each channel.

[0066] compress the convolved features back to dimension through a linear layer, and add the input residual, normalize:

[0067] ;

[0068] where, is a linear projection matrix, is a layer normalization function;

[0069] by stacking encoder layers described above, we finally obtain a high-dimensional feature sequence containing rich global-local temporal information; the class token at the beginning of the sequence constantly interacts with all frame tokens across layers, and its final output vector serves as the global representation of the acoustic signal.

[0070] Preferably, the classifier specifically comprises:

[0071] input the to a linear layer and bias to obtain logits for each fault class:

[0072] ;

[0073] where, is a classifier weight, is a classifier bias, is the number of fault classes;

[0074] obtain the class probability distribution through the softmax function:

[0075] ;

[0076] wherein, is the prediction probability of the class fault, is the th component of the logical output vector , is an exponential function;

[0077] The cross-entropy loss function is adopted to calculate the loss with the real label and the prediction probability , and the network parameters are updated by back propagation.

[0078] An electronic device comprising: one or more processors; a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0079] A computer-readable storage medium having stored thereon executable instructions that, when executed by a processor, cause the processor to implement the method.

[0080] The beneficial effects of the present application are: by constructing a double-channel parsed acoustic representation, adopting a frame-embedded manner to retain local transient impact structure, and combining a lightweight Transformer encoder to realize global correlation modeling, the acoustic fault feature separability and integrity are significantly improved. This method can maintain stable diagnostic performance in multi-noise and multi-working-condition environments, has higher sensitivity to weak fault pulses, and has the advantages of low computational complexity, strong generalization ability, wide application range, etc., which can effectively improve the intelligent diagnosis precision and reliability of acoustic monitoring of rolling bearings. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 is a schematic diagram of the overall architecture of a rolling bearing fault diagnosis method based on acoustic frame provided by an embodiment of the present application;

[0082] Figure 2 is a schematic diagram of a double-channel parsed acoustic signal constructed in an embodiment of the present application;

[0083] Figure 3 is a structural schematic diagram of an acoustic frame-embedded module in an embodiment of the present application;

[0084] Figure 4 is a structural schematic diagram of a lightweight Transformer encoder in an embodiment of the present application;

[0085] Figure 5 is a schematic diagram of the training loss and accuracy curves of the present application on multiple comparative models;

[0086] Figure 6A comparison diagram of multi-model fault diagnosis accuracy under different signal-to-noise ratio conditions is shown in the figure.

[0087] Figure 7 A comparison diagram of T-SNE dimensionality reduction visualization of each model in the feature space is shown in the figure. DETAILED DESCRIPTION

[0088] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other. In order to achieve the above purpose, the technical scheme is as follows.

[0089] The present application provides a rolling bearing fault diagnosis method based on acoustic framing, comprising the following steps:

[0090] First, the acoustic signals of the rolling bearing in the normal state and in various fault states during operation are collected, and the collection mode can be a non-contact microphone or an acoustic sensor. In order to improve the stability of subsequent feature extraction, the original acoustic signal is preprocessed, including removing DC bias, suppressing outliers and amplitude normalization, so that the signal remains consistent in amplitude scale and statistical characteristics, thereby providing a reliable input basis for constructing the analytic domain representation.

[0091] After the preprocessing is completed, Hilbert transform is applied to the acoustic signal to generate a virtual signal that is 90° different from the original sound wave; the virtual signal and the original real signal are combined to form a double-channel analytic acoustic signal containing instantaneous amplitude variation and phase evolution information. The double-channel structure can more comprehensively represent the weak pulse and local phase disturbance caused by the interaction between the rolling body and the defect in the sound wave propagation process.

[0092] Subsequently, the double-channel analytic acoustic signal is divided into multiple short-time frame segments according to the preset frame length L and step length R. Since framing can maintain the approximate stationarity of the acoustic signal in the short time scale, the obtained frame segments not only contain both real and imaginary data, but also can retain local transient impact structure, amplitude fluctuation and phase disturbance, thereby avoiding feature ambiguity and detail loss caused by directly processing the entire sequence.

[0093] To convert the framed acoustic data into model-processable sequence input, each frame segment is vectorized and projected to a fixed-dimension embedding vector space through linear mapping. Then, learnable position encoding is added to all embedding vectors to characterize the temporal order of frame segments; meanwhile, a class token is added at the beginning of the sequence to enable the model to aggregate global information in the subsequent encoding stage. The resulting embedding sequence is input into the Transformer encoder for deep feature extraction.

[0094] In the feature extraction stage, the embedding sequence enters the lightweight Transformer encoder stacked by multiple acoustic feature extraction modules. Each module first models the long-range association between different frame segments through multi-head self-attention layers, enabling the model to perceive cross-frame repetitive structures, fault impact intervals, and global patterns affected by working condition changes. Subsequently, local continuity features are extracted through a feedforward network based on SiLU activation and depthwise separable convolution. This structure first generates two groups of feature channels through linear mapping, then enhances the model's response to local mutation points and weak pulses through activation gating, and strengthens the temporal correlation between adjacent frame segments using depth convolution. Finally, the dimension is restored through linear mapping. To ensure stable training of the deep network and effective fusion of local and global features, residual connections and layer normalization operations are set between the multi-head self-attention layers and the feedforward network layers.

[0095] After passing through multiple acoustic feature extraction modules, the feature vector corresponding to the class token output by the Transformer encoder is considered as the global acoustic representation. Finally, this feature vector is input into the fully connected layer and output through the softmax classifier to identify the fault type of the rolling bearing, achieving accurate identification of multiple fault states.

[0096] Embodiment one: overall structure of acoustic framing embedding Transformer

[0097] The present embodiment, as shown in Figure 1 The rolling bearing fault diagnosis system of the present application is composed of a Hilbert transform preprocessing layer, an acoustic framing embedding module (AFEM), a lightweight Transformer encoder, and a classification output unit in sequence.

[0098] Firstly, the acoustic signals of bearings in different states are collected by microphones, and the Hilbert transform is used to generate a dual-channel analytic representation containing real and imaginary parts; then the analytic signals are framed by sliding window with a set frame length and step, and the whole acoustic waveform is reorganized into a series of short-time frame segments, and the frame-level embedding sequence is constructed by linear mapping and position encoding; in the feature extraction stage, the embedding sequence is input into the lightweight Transformer encoder formed by stacking multi-head self-attention and SC-GLU, which realizes the joint modeling of local transient impact structure and cross-frame global timing dependence; finally, the global feature vector corresponding to the class token in the encoder output is extracted, and the fault class of rolling bearings is obtained through the full connection layer and softmax classifier, realizing the end-to-end acoustic intelligent diagnosis process.

[0099] Embodiment two: Hilbert transform preprocessing layer

[0100] As shown in Figure 2 , this embodiment uses Hilbert transform to construct a dual-channel analytic acoustic signal to explicitly represent the evolution of instantaneous amplitude and phase.

[0101] Let the collected real-valued acoustic signal be , and the Hilbert transform is applied to obtain the orthogonal components :

[0102] ;

[0103] where is the Hilbert operator, represents the Cauchy principal value, is the integral variable, is the time variable.

[0104] In the frequency domain, we have:

[0105] ;

[0106] where denotes the Fourier transform, is the frequency, is the imaginary unit, is the spectrum of , and is the sign function (1 for positive frequency and -1 for negative frequency). The real and imaginary parts are combined into a complex analytic signal:

[0107] ;

[0108] where the instantaneous amplitude is:

[0109] ;

[0110] instantaneous phase:

[0111] ;

[0112] The acoustic signal is represented in a vector form of a two-channel analytic acoustic signal:

[0113] ;

[0114] wherein, and represent the in-phase component and the quadrature component, respectively, denotes the transpose of a matrix or vector, is the time length of a single sample (i.e. the number of sampling points), represents a real matrix with a feature dimension of 2 and a time length of .

[0115] By this pre-processing, the single-channel acoustic signal originally only containing amplitude information is promoted to a two-channel analytic representation containing both instantaneous amplitude and phase structure, providing a more complete physical information basis for subsequent framing and feature learning.

[0116] Embodiment Three: Acoustic Framing Embedding Module (AFEM)

[0117] As shown in FIG. 3, the present embodiment reorganizes the analytic acoustic waveform into a series of short-time frame-level tokens by AFEM, thereby balancing local transient structure preservation and computational complexity control. Figure 3 The analytic signal

[0118] is framed using a sliding window with a length of and a step size of . The th acoustic frame segment is represented as:

[0119] ;

[0120] wherein, is the starting sampling point, and the frame number is:

[0121] ;

[0122] ​In the specific implementation of the present application, the length of a single acoustic signal sample is 1024 sampling points, which can cover multiple impact cycles generated by the rolling body contacting the defect while maintaining a high time resolution. In order to extract short-time local structures and reduce the overall sequence length, the analyzed acoustic signal is divided into frames with a length L = 16 sampling points, and a sliding manner with a step R = L / 2 = 8 is used, i.e. an overlap rate of 50%. Under this configuration, the 1024-point signal can be divided into about 127 acoustic frame segments with short-time stability, each frame maintaining quasi-stationary characteristics in a local time range, facilitating the concentrated description of the instantaneous impact pattern generated by the rolling body-defect interaction, and effectively improving the time sequence expression capability of the subsequent Transformer modeling.

[0123] The flattening operation is performed on each acoustic frame to obtain a vector:

[0124] ;

[0125] and mapped to a dimensional embedding space through a linear layer:

[0126] ;

[0127] wherein , are learnable projection parameters; each i.e. corresponds to a frame-level token, encoding the amplitude-phase local structure within the frame;

[0128] All frame embeddings are spliced to form a time-ordered token sequence:

[0129] .

[0130] A learnable class token vector is added at the beginning of the token sequence to obtain the final input sequence:

[0131] ;

[0132] A position encoding matrix of the same dimension as is constructed and element-wise added to the extended sequence to obtain the final input sequence .

[0133] ;

[0134] wherein is defined as the complete input sequence (including the class token and frame tokens) enhanced by position encoding, which is the direct input of the subsequent Transformer encoder.

[0135] Since the computational complexity of self-attention is quadratic with respect to the sequence length, if the original sequence with length is directly modeled, the complexity is about . After AFEM framing, the sequence length is reduced to , and the complexity is about

[0136] ;

[0137] where is the original signal length, is the frame length, R is the step size, is the feature dimension, is the sequence length after framing. When , the computational complexity of the attention part can be theoretically reduced to about of the original waveform modeling, significantly improving the overall operation efficiency.

[0138] Embodiment Four: Improved Lightweight Transformer Encoder

[0139] As shown in Figure 4 , this embodiment is designed based on the standard Transformer encoder, including low-dimensional multi-head self-attention and SiLU-convolutional gated linear units.

[0140] The input of the input sequence from the acoustic framing embedding module is linearly projected to obtain the query , key and value matrices:

[0141] ;

[0142] where , , is the number of attention heads, is the low-dimensional subspace dimension of each head;

[0143] The scaled dot-product attention is calculated for each head respectively, and the combination is written as:

[0144] ;

[0145] where is the normalized attention weight matrix, representing the correlation strength between different frame-level tokens in the acoustic sequence; is the output matrix of the attention mechanism; After connecting the multi-head outputs, it is restored to the model dimension.​ Contextual representation fusing global inter-frame correlation.

[0146] To enhance the stability of deep network training, the "attention-residual-normalization" structure is adopted:

[0147] ;

[0148] where LN denotes the layer normalization operation; is the normalized sequence;

[0149] First, the input is expanded to dimensions through a linear layer:

[0150] ;

[0151] where, is the expansion projection matrix, used to expand the input feature dimension from D to the hidden layer dimension 2F; denotes channel-wise concatenation; then the gated linear unit SC-GLU with SiLU activation is adopted:

[0152] ;

[0153] where denotes element-wise multiplication, is a smooth nonlinear function;

[0154] For , a depth separable one-dimensional convolution is applied along the sequence dimension:

[0155] ;

[0156] where, denotes a depth convolution operation using independent convolution kernels for each channel.

[0157] The convolved features are compressed back to dimensions through a linear layer, and added to the input residual, normalized:

[0158] ;

[0159] where, is the linear projection matrix, is the layer normalization function;

[0160] By stacking encoder layers described above, a high-dimensional feature sequence containing rich global-local temporal information is finally obtained; the class token located at the beginning of the sequence constantly interacts with all frame tokens between layers, and its final output vector is the global representation of the acoustic sample.

[0161] Example 5: Fault diagnosis process based on category token

[0162] As shown in Figure 1 and Figure 7 , this embodiment constructs a fault classifier based on the category token output by the encoder described above. Specifically, it includes:

[0163] The original long sequence is divided into fixed-length samples by a sliding window to form a training set, a validation set and a test set (as shown in Figure 7 );

[0164] is input into the linear layer and bias to obtain logits for each fault category:

[0165] ;

[0166] wherein, is the classifier weight, is the classifier bias, is the number of fault categories;

[0167] The category probability distribution is obtained by the softmax function:

[0168] ;

[0169] wherein, is the predicted probability of the th fault category, is the th component of the logical output vector , and is the exponential function;

[0170] The cross-entropy loss function is used to calculate the loss between the true label and the predicted probability , and the network parameters are updated by backpropagation.

[0171] After the training is completed, the test samples are input into the trained model to obtain the fault category prediction results of each sample. At the same time, the training / validation loss and accuracy curves of different models can be plotted (as shown in Figure 5 ), and the diagnostic accuracy changes under various signal-to-noise ratios (as shown in Figure 6 ) and feature space T-SNE distribution (as shown in Figure 7 ) are calculated, which are used to further demonstrate the convergence characteristics, noise robustness and feature separability of the method of the present application.

[0172] Example 6: Effect verification and explanation ​

[0173] The present embodiment verifies the method of the present application using a rolling bearing multi-working condition acoustic data set. The data set contains 10 typical bearing operating states, covering normal state and various structural fault types, specifically including: normal working condition Normal (NOR); inner ring fault Inner Race Fault (INC: inner ring crack, INP: inner ring pitting, INW: inner ring wear); rolling element fault Ball Fault (BAC: rolling element crack, BAP: rolling element pitting, BAW: rolling element wear); and outer ring fault Outer Race Fault (ORC: outer ring crack, ORP: outer ring pitting, ORW: outer ring wear). Each type of acoustic data is collected by a non-contact microphone at a fixed speed, different loads and different defect positions, which can truly reflect the influence of multi-working condition background noise, structural reflection interference and different damage mechanisms on acoustic response, thereby constructing a highly realistic industrial scene diagnosis environment in the experiment.

[0174] To ensure the fairness of model training and performance evaluation, the present embodiment divides each type of sample into a training set, a validation set and a test set in a fixed proportion, as shown in the following table. Each type of working condition contains 300 training samples, 100 validation samples and 200 test samples, totaling 6000 training samples, 1000 validation samples and 2000 test samples, achieving a strictly balanced class distribution. All data are uniformly preprocessed before entering the model, including amplitude normalization, simple filtering and fixed-length slicing, to ensure signal scale consistency between different fault types. Based on this multi-working condition, multi-fault type acoustic data set, the experiment can comprehensively verify the robustness and generalization ability of the method of the present application under complex noise environment, different fault forms and multiple operating conditions.

[0175]

[0176] To verify the effectiveness of the method of the present application, the rolling bearing fault diagnosis method based on acoustic frame embedding Transformer proposed by the present application is compared with various typical models in the prior art, including standard Transformer, ViT, CLFormer, MCSwinT, Convformer-NSE, LiConvFormer and TCAC-Transformer. Each model is trained and tested under the same data set, same training strategy and same data preprocessing conditions. The comparison indicators include accuracy, precision, recall, F1 score, model parameter size, FLOPs and training time, and the results are shown in the following table.

[0177]

[0178] As can be seen from the overall performance table, the traditional Transformer is at the lowest level in each index, and the existing models (such as CLFomer, LiConvFormer, etc.) with convolution structure or lightweight module have improved performance, but still have problems of insufficient local transient structure extraction or high parameter quantity. In contrast, the accuracy, precision, recall and F1 score of the method of the application are all 99.75%, which is the highest among all comparison models, and only requires 0.09M parameters and 13.66M FLOPs, and the training time is also kept at a low level, which reflects a good balance between high precision and lightweight.

[0179] Further, in order to evaluate the stability of different models under noise conditions, the test accuracy of each model under no noise and different signal-to-noise ratios (5 dB, 3 dB, 0 dB, -3 dB, -5 dB) is compared, and the results are shown in the following table. With the decrease of signal-to-noise ratio, the diagnostic performance of all models shows a downward trend, but the decline amplitude is obviously different. Among them, the accuracy of the traditional Transformer is only 79.48% under -5 dB, and its anti-noise performance is weak; the performance of ViT, CLFormer and TCAC-Transformer also decreases obviously under medium and low noise conditions. The method of the application maintains high level of fault recognition ability under each noise level, especially under the extremely low signal-to-noise ratio of -5 dB, it still reaches 96.85% of accuracy, which is the highest among all models, and its average performance also reaches 98.33%, which is significantly better than other comparison methods.

[0180]

[0181] The above results show that the method of the application can more fully extract the short-time impact features and global temporal correlation in acoustic signals by jointly modeling the design of domain dual-channel representation, acoustic frame-embedding structure and lightweight Transformer, and has significantly better classification performance and robustness than the prior art under multiple noise and multiple working conditions. The experiment fully proves the effectiveness and advancement of the method proposed by the application.

[0182] Figure 5 The training loss, training accuracy, validation loss and validation accuracy curves of the method of the application and various comparison models on Case 1 dataset are shown. It can be seen that the method of the application can converge in about 10 epochs, and has the lowest validation loss and the highest validation accuracy, which shows that it has good fitting ability and generalization performance while maintaining a lightweight structure.

[0183] Figure 6The diagnostic accuracy comparison curves under different signal-to-noise ratio conditions are given. The method of the application can maintain an accuracy higher than 96% at SNR=-5dB, while the accuracy of the standard Transformer and part of the comparison model significantly decreases under this condition, indicating that the dual-channel analytic representation, frame embedding and SC-GLU structure jointly enhance the noise interference suppression capability.

[0184] Figure 7 The T-SNE visualization results of different models under strong noise conditions are shown. The clusters of each fault category obtained by the method of the application are compact and the boundaries between classes are clear, which is significantly better than the standard Transformer and part of the lightweight variants, indicating that the proposed architecture has stronger class separation capability and diagnostic interpretability in the feature space.

[0185] The above embodiments show that the application not only strengthens the joint modeling capability of the bearing fault induced acoustic short-time impact and long-range dependence, but also significantly reduces the model calculation complexity and parameter size, and obtains excellent diagnostic effect under multiple data sets and multiple noise conditions.

[0186] The above description of the specific embodiments is only used to help understand the technical concept and core idea of the application. Although the technical solutions of the application are described and explained in the specification in combination with specific preferred embodiments, these embodiments should not be understood as any limitation on the protection scope of the application. Those skilled in the art can make various changes, equivalent replacements or improvements to the embodiments in form, structure or details without departing from the technical concept of the application. These obvious changes based on the concept of the application should fall within the protection scope of the application.

Claims

1. A method for diagnosing rolling bearing faults based on acoustic framing, characterized in that, Includes the following steps: Acoustic signals of rolling bearings during operation under normal conditions and various fault conditions were collected; Apply Hilbert transform to the acoustic signal to construct an analytical signal, obtain the real and imaginary components corresponding to the instantaneous acoustic energy change and phase evolution, and combine them into a dual-channel analytical acoustic signal; The dual-channel analytical acoustic signal is divided into multiple acoustic frame segments by a sliding window to simultaneously characterize local amplitude, phase, and transient impact features. Each acoustic frame segment is vectorized, expanded, and mapped to a fixed-dimensional embedding vector via linear projection. Learnable positional codes are superimposed on all embedding vectors, and a category token for aggregating global information is added to the beginning of the embedding vectors to form the input sequence of the Transformer. The input sequence is fed into a Transformer encoder consisting of stacked multi-layer acoustic feature extraction modules, which include a multi-head self-attention layer and a feedforward network layer, and residual connections and layer normalization are respectively set between the multi-head self-attention layer and the feedforward network layer. Finally, the feature vector corresponding to the category token in the output sequence of the Transformer encoder is taken, mapped, and output as the fault category of the rolling bearing by the classifier.

2. The method as described in claim 1, characterized in that: Applying a Hilbert transform to the acoustic signal to construct an analytic signal yields the real and imaginary components corresponding to the instantaneous acoustic energy change and phase evolution, respectively. These components are then combined to form a dual-channel analytic acoustic signal, specifically including: Let the acquired real-valued acoustic signal be... Applying a Hilbert transform to it yields orthogonal components. : ； in, For Hilbert operators, Indicates Cauchy's principal value. For integration variables, It is a time variable; In the frequency domain, we have: ； in, Indicates Fourier transform, For frequency, It is an imaginary unit. for The spectrum, For a sign function, the real and imaginary parts are combined to form a complex analytic signal: ； Instantaneous amplitude: ； Instantaneous phase: ； Representing acoustic signals as vector forms of dual-channel analytic acoustic signals: ； in, and These represent the in-phase components and the quadrature components, respectively. Represents the transpose of a matrix or vector. The time length of a single sample is equivalent to the number of sampling points. This indicates a feature dimension of 2 and a time length of [missing information]. A real matrix.

3. The method as described in claim 2, characterized in that: The dual-channel analytical acoustic signal is segmented using a sliding window to form multiple acoustic frame segments that can simultaneously characterize local amplitude, phase, and transient impact features, specifically including: For dual-channel analytical acoustic signals Use a length of Step size is The sliding window is used for frame division; the first Each acoustic frame segment is represented as: ； in, Starting sampling point, number of frames: ； For frame length, Step size, This indicates the floor function.

4. The method as described in claim 3, characterized in that: Each acoustic frame segment is vectorized, unfolded, and mapped to a fixed-dimensional embedding vector via linear projection. Specifically, this includes: For each acoustic frame segment Flattening the vector yields: ； And mapped to through a linear layer Dimensional embedding space: ； in, , For learnable projection parameters; each That is, it corresponds to a frame-level token, which encodes the amplitude-phase local structure within that frame; Embed and concatenate all frames to form a time-ordered token sequence: 。 5. The method as described in claim 4, characterized in that: Learnable positional encodings are superimposed on all embedding vectors, and a category token for aggregating global information is added to the beginning of the embedding vectors to form the input sequence of the Transformer, specifically including: Add a learnable category token vector to the beginning of the token sequence. The final input sequence is obtained as follows: ； Construction and Same-dimensional position encoding matrix Then, add element-wise the input sequence to the extended sequence to obtain the final input sequence. ; ； in, Defined as the input sequence after positional encoding enhancement.

6. The method as described in claim 5, characterized in that: The input sequence is fed into a Transformer encoder composed of stacked multi-layer acoustic feature extraction modules. Each acoustic feature extraction module includes a multi-head self-attention layer and a feedforward network layer. Residual connections and layer normalization are respectively set between the multi-head self-attention layer and the feedforward network layer, specifically including: For the input sequence The The input of the layer is linearly projected to obtain the query. ,key Sum matrix: ； in, , , For the number of attention heads, The low-dimensional subspace dimension for each head; Calculate the scaled dot product attention for each head separately, and combine them as follows: ； in, This is the normalized attention weight matrix, representing the correlation strength between different frame-level tokens in the acoustic sequence; This is the output matrix of the attention mechanism; After connecting the multiple outputs, restore them to the model dimension; A contextual representation that integrates global inter-frame correlations.

7. The method as described in claim 6, characterized in that: Specifically, setting residual connections and layer normalization between the multi-head self-attention layer and the feedforward network layer includes: ； Wherein, LN represents the layer normalization operation; The sequence is normalized; First, input Extended through linear layers dimension: ； in, To extend the projection matrix; This indicates concatenation by channel; subsequently, a gated linear unit (SC-GLU) with SiLU activation is used: ； in This represents element-wise multiplication. It is a smooth nonlinear function; right Applying a depthwise separable one-dimensional convolution along the sequence dimension: ； in, This indicates a depthwise convolution operation that uses an independent kernel for each channel; The convolutional features are compressed back through a linear layer. Dimension, add to the input residual, and normalize: ； in, It is a linear projection matrix. For layer normalization function; By stacking The encoder layers ultimately produce a high-dimensional feature sequence containing rich global-local temporal information; the category token at the beginning of the sequence continuously interacts with all frame tokens between layers, and its final output vector That is, as a global representation of acoustic signals.

8. The method as described in claim 7, characterized in that: The classifier specifically includes: Will Input to linear layer and bias In the process, the logits for each fault category are obtained: ； in, For classifier weights, For classifier bias, Number of fault categories; The class probability distribution is obtained using the softmax function: ； in, For the first Predicted probability of class-specific faults For logical output vectors The One portion, It is an exponential function; Employing the cross-entropy loss function with the true label With predicted probability Calculate the loss and update the network parameters through backpropagation.

9. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1 to 8.

Images