CN-122019972-A - Incremental sparse network vibration signal denoising method based on differential attention

Abstract

The invention discloses a differential attention-based incremental sparse network vibration signal denoising method, and belongs to vibration signal denoising technologies. The method comprises the steps of collecting an original bearing vibration signal, adding composite noise into the preprocessed signal, constructing a target denoising network by combining a sparse connection neural network and a differential attention mechanism, constructing a multi-target loss function, performing end-to-end training on the target denoising network, inputting the vibration signal to be denoised into the trained denoising network to obtain a high-fidelity denoised signal, and performing incremental updating on the denoising network based on a new sample when a new noise type or working condition occurs, and adjusting network parameters through elastic weight consolidation and knowledge distillation technology to achieve self-adaptive optimization and performance improvement of a model. According to the invention, the denoising parameters are not required to be independently regulated for different signals, so that the denoising of the vibration signals in a multi-task scene can be effectively completed, and the follow-up spectrum analysis and the accurate implementation of intelligent fault diagnosis are facilitated.

Inventors

• QIAN WEIWEI
• YANG YUTING

Assignees

• 南京信息工程大学

Dates

Publication Date

20260512

Application Date

20260128

Claims (6)

1. 1. The incremental sparse network vibration signal denoising method based on differential attention is characterized by comprising the following steps of: collecting an original bearing vibration signal, carrying out standardization and segmentation pretreatment on the original bearing vibration signal, and simulating a complex noise environment in a real working condition by adding composite noise into the pretreated signal to generate a noise-clean pairing sample which is used for constructing a signal sample set for model training and verification; Constructing a target denoising network by combining the sparse connection neural network and a differential attention mechanism; constructing a multi-objective loss function by combining the difference loss, the sparse constraint loss and the topology regularization loss, and performing end-to-end training on the objective denoising network to simultaneously optimize the denoising effect and the model efficiency and obtain a trained denoising network, wherein the input item of the objective denoising network is a noisy signal, and the output item is a denoising signal; inputting the vibration signal to be denoised into a trained denoising network to obtain a denoised signal with high fidelity; when a new noise type or working condition occurs, the denoising network is updated in an increment mode based on a new sample, and network parameters are adjusted through elastic weight consolidation and knowledge distillation technology, so that self-adaptive optimization and performance improvement of the model are realized.
2. 2. The method for denoising an incremental sparse network vibration signal based on differential attention of claim 1, wherein the step of end-to-end training the target denoising network comprises: Let the noisy signal be Wherein In order to be of the size of the batch, The signal length is that the noise signal is formed by the clean signal after pretreatment By linearly superimposing composite noise Obtaining a noise-carrying signal, inputting the noise-carrying signal into a sparse connection neural network to obtain a noise-removing signal Expressed as: , Wherein, the Is a sparse dictionary intended to capture typical waveforms in the bearing vibration signal; The sparse coefficient is used to determine the degree of the sparse coefficient, Corresponds to the presence of a significant fault component in the signal.
3. 3. The differential attention-based incremental sparse network vibration signal denoising method according to claim 2, wherein the sparse dictionary is updated by updating the sparse dictionary respectively Sparse coefficient Solving the value by the way of (a), and the process comprises the following steps: Firstly, fixing a sparse dictionary, and solving sparse coefficients by using a differential attention mechanism; Then, the sparse coefficient is fixed, the sparse dictionary is updated through gradient descent, and the updated sparse dictionary is expressed as: , In the formula, Represents the learning rate, is used to control the update step size, The gradient operator is represented by a gradient operator, Representing the input noisy signal.
4. 4. The differential attention-based incremental sparse network vibration signal denoising method of claim 3, wherein solving the sparse coefficients using a differential attention mechanism comprises: For the first Sparse coefficient of wheel Fixed sparse dictionary Calculating the gradient descent step to obtain an intermediate characteristic variable, wherein the formula is as follows: , Projecting intermediate feature variables as queries Key and key Sum value Expressed as: , , , In the formula, 、 、 Is an adaptive parameter of the differential attention model, Calculating a differential attention operator, wherein the formula is as follows: , In the formula, Representing a soft threshold function, The coefficients for adjusting the step characteristics of the attention weights are expressed as: , In the formula, , , , In order for the vector to be a learnable vector, To be used for initialization Is controlled between 0 and 1; According to the differential attention operator, the sparse coefficient under the current round is obtained through calculation, and the expression is as follows: In the formula, Which represents the Hadamard product of the two, Representing regularization parameters.
5. 5. The differential attention-based incremental sparse network vibration signal denoising method of claim 1, wherein the composite noise Gaussian noise including pink noise characteristics Interference noise of power frequency Baseline drift noise Impulse noise ; The composite noise is expressed as: , In the formula, Representing the original signal, the signal is then processed, Is a weight coefficient corresponding to noise.
6. 6. The differential attention-based incremental sparse network vibration signal denoising method according to claim 1, wherein after incremental updating the denoising network, the final multi-objective loss function of the denoising network is expressed as: , In the formula, As a function of the difference loss, In order to constrain the loss function, In order for the topology to be regularized and lost, In order to consolidate the loss of the elastic weight, Is distillation loss; 、 、 representing the weights of the corresponding loss functions, respectively.

Description

Incremental sparse network vibration signal denoising method based on differential attention Technical Field The invention relates to a vibration signal denoising technology, in particular to a differential attention-based incremental sparse network vibration signal denoising method. Background Accurate fault diagnosis in rotating machinery is critical to ensure operational safety and to prevent costly downtime in modern industrial systems. However, extracting reliable fault signatures from bearing vibration signals remains a significant challenge. These signals generated under complex operating conditions are extremely susceptible to noise contamination. This inherent noise severely masks critical fault features, preventing accurate diagnosis and prediction. To address this challenge, efficient signal denoising techniques are an indispensable preprocessing step. Although various denoising methods exist, many methods have difficulty in adequately preserving subtle, transient time characteristics indicative of early failure while effectively suppressing complex noise patterns in real-world scenarios. In the field of vibration signal processing, a common denoising method mainly comprises a filtering technology, wavelet transformation, time-frequency analysis and the like. The filtering method is characterized in that interference components in a specific frequency range in signals are removed, wavelet transformation is particularly suitable for processing non-stationary vibration signals, useful information and noise parts can be effectively separated, and the time-frequency analysis method such as Empirical Mode Decomposition (EMD), variational Mode Decomposition (VMD) and improved algorithm thereof is more suitable for signals with strong nonlinearity and non-stationarity, and can provide more detailed characteristic characterization. The most appropriate denoising strategy should be selected for the vibration characteristics, noise type and signal-to-noise ratio level of different devices, and this process usually requires expert knowledge. In order to further improve the self-adaptive performance of the denoising method, researchers often introduce evaluation bases such as kurtosis indexes and the like, and a better noise suppression effect is realized by combining a genetic algorithm and other parameter optimization algorithms. Although the denoising technology based on signal processing can effectively inhibit noise, expert experience is still relatively dependent when dealing with complex noise scenes, and the adaptability to various noise in a variable environment is still limited. In contrast, the deep learning method can directly extract features from the original signals and output denoising results in an end-to-end mode, and can effectively model uncertain factors in the environment by fusing information such as characteristics and forms of noise, so that the robustness and reliability of a denoising system are enhanced. Therefore, the traditional denoising strategy is embedded into the deep learning model, and the advantages of the traditional denoising strategy and the deep learning model are integrated, so that the method plays an important role in improving the denoising precision and the self-adaptive capacity. For example, some studies have employed a CNN-based denoising model to extract features in a signal by constructing multiple convolution layers and reconstruct the denoised signal using a deconvolution layer or an upsampling layer. In addition, there have been some studies on the application of RNN or long-short-term memory network (LSTM) to vibration signal denoising. The network structures can capture time sequence information in the signals, and have good effect on processing continuous vibration signals. The existing vibration signal denoising method still has obvious limitations when dealing with unstable, strong noise interference and variable working conditions, and particularly has defects in the aspects of distinguishing capability of signal feature extraction, model generalization and calculation efficiency. The traditional deep learning denoising model often depends on a fixed structure, is difficult to adaptively focus on the difference between key components and noise, and has large calculation load generally, so that the wide application of the model in a real-time monitoring system is limited. Disclosure of Invention Aiming at the problems, the invention aims to provide an incremental sparse network vibration signal denoising method based on differential attention, which is characterized in that a sparse network structure is introduced to effectively separate signals from noise, a differential attention mechanism is introduced at the same time, the distinguishing capability of useful components in vibration signals and noise is enhanced, and an incremental learning mechanism and a sparse network structure are combined, so that the denoising precision is improved, the con