Search

CN-121996911-A - Gating jump network denoising method integrating multi-scale space-time attention

CN121996911ACN 121996911 ACN121996911 ACN 121996911ACN-121996911-A

Abstract

The invention discloses a gating jump network denoising method integrating multi-scale space-time attention. The method realizes the self-adaptive denoising and characteristic reconstruction of the noisy one-dimensional time sequence signal by constructing an encoding-decoding network which fuses a multi-scale parallel convolution structure, a gate control jump connection mechanism and BiLSTM time sequence modeling units. The method comprises the steps of capturing local details and long period characteristics of a signal by utilizing multi-scale parallel convolution, embedding a gating jump connection mechanism in a decoding stage, carrying out channel splicing on shallow layer characteristics of the encoder and deep layer characteristics of a decoder, embedding a driving attention sub-network to generate a gating mask, carrying out self-adaptive weighted filtering on the shallow layer characteristics by utilizing the mask, and blocking the propagation of background noise to a decoding end before fusion. The invention can obtain multi-scale representation taking local transient change and long-term evolution trend into consideration, remarkably improve the input quality of subsequent feature extraction, pattern recognition and life assessment algorithms, and improve the accuracy and reliability of fault diagnosis flow.

Inventors

  • LIU TAO
  • DI ZHICHENG
  • LI YONGBO
  • WANG TENG

Assignees

  • 西北工业大学

Dates

Publication Date
20260508
Application Date
20251229

Claims (9)

  1. 1. A gating jump network denoising method integrating multi-scale space-time attention is characterized by comprising the following steps: step 1, collecting data and constructing a data set; Firstly, collecting and storing a data set of mechanical vibration signals according to requirements, finally generating a training set and a testing set, and completing normalization pretreatment of the data; step 2, constructing an encoder based on multi-scale gating Inception; Firstly, receiving a one-dimensional time domain signal through a sequence input layer, and then connecting two multi-scale gating Inception MGI coding modules and a maximum pooling layer in series; Four feature extraction branches are arranged in parallel in the MGI coding module, namely a 1X 1 convolution branch, a 3X 3 convolution branch with expansion rate of 2, a 3X 3 convolution branch with expansion rate of 4 and a maximum pooling branch, and are used for capturing local details of different scales and long period features in signals at the same time; Step 3, constructing a time sequence characteristic bottleneck layer based on BiLSTM; Firstly, a characteristic diagram output by an encoder is adjusted to an adaptation dimension through a1 multiplied by 1 convolution layer, then BiLSTM layers containing 64 hidden units are accessed, the BiLSTM layers judge signal properties by utilizing waveform information before and after the current moment by utilizing a bidirectional circulation mechanism, and finally, model overfitting is prevented by a discarding layer; Step 4, constructing a decoder comprising a gate control jump connection; The method comprises the steps of carrying out channel splicing on shallow features of an encoder and current deep features of a decoder, sending the shallow features and the current deep features of the decoder into an attention generation sub-network, generating a gating mask with a value range of 0-1 after convolution and Sigmoid activation, carrying out weighted filtering on the shallow features of the encoder by using the gating mask, allowing only components related to fault features to pass through so as to block the propagation of background noise to a decoding end, merging the filtered features with the features of the decoder, and sending the merged features into a subsequent MGI decoding module, wherein the shallow features refer to features containing details but accompanying noise; step 5, constructing an output layer; the output end of the network is provided with a convolution layer with a convolution kernel size of 1 to compress the multi-channel characteristics into single-channel output, and the multi-channel characteristics are connected with an MGI decoding module to reconstruct the final time domain waveform; Step 6, configuring network training super parameters and executing model training; And (3) configuring an Adam optimizer as an updating algorithm of network parameters, inputting the constructed training set into a network, and iteratively updating the network weight by minimizing regression errors between the predicted waveform and the pure labels until the loss function converges.
  2. 2. The method for denoising the gating skip network with the integration of the multi-scale space-time attention as claimed in claim 1, wherein the signal sampling rate is set to 10000Hz in the step 1, the signal length of a single sample is 1024 data points, and the fault characteristic frequency randomly fluctuates between 80Hz and 120 Hz.
  3. 3. The method of claim 1, wherein the training set comprises 2000 sets of samples and the test set comprises 200 sets of samples.
  4. 4. The method for denoising a gated jump network with integrated multiscale spatio-temporal attention according to claim 1, wherein the discard rate of the discard layer is 0.2.
  5. 5. The method for denoising the gating skip network with the integration of the multi-scale space-time attention as claimed in claim 1, wherein in the step 5, an initial learning rate is set to be 0.001, a segmented attenuation strategy is adopted, the learning rate is reduced to be 10% of the original learning rate after 15 training rounds, the maximum training round is set to be 30 rounds, and the small-batch sample size is set to be 64.
  6. 6. An electronic device comprising a processor and a memory, the memory for storing a computer program, the processor for executing the computer program stored by the memory to cause the electronic device to perform the method of any one of claims 1 to 5.
  7. 7. A computer readable storage medium, on which a computer program is stored, which computer program, when being executed by a processor, implements the method according to any one of claims 1 to 5.
  8. 8. A chip comprising a processor for calling and running a computer program from a memory, causing a device on which the chip is mounted to perform the method of any one of claims 1 to 5.
  9. 9. A computer program product comprising a computer storage medium storing a computer program comprising instructions executable by at least one processor, the instructions when executed by the at least one processor implementing the method of any one of claims 1 to 5.

Description

Gating jump network denoising method integrating multi-scale space-time attention Technical Field The invention belongs to the technical field of signal processing, and particularly relates to a gating jump network denoising method integrating multi-scale space-time attention. Background In the operation process of complex industrial equipment and electromechanical systems, a large number of key components are in working environments of variable load, high noise and strong interference for a long time, and the operation state of the key components is often indirectly represented by one-dimensional time sequence signals such as vibration, current, acoustics or pressure. Because the characteristic signals generated in the early stage of the fault have smaller amplitude and short duration and are often overlapped with background noise, working condition fluctuation and structural resonance, effective information is difficult to directly separate from an original measurement signal, and a remarkable challenge is brought to state monitoring and fault diagnosis. Therefore, how to accurately extract key signal components related to the health state of equipment in a strong noise background is always the research focus in the fields of digital signal processing and intelligent fault diagnosis. The traditional one-dimensional signal analysis method mostly depends on manually designed filters, time-frequency analysis or statistical feature extraction means, such as band-pass filtering, empirical wavelet decomposition and the like. The method generally needs to rely on prior experience to manually adjust parameters, is easily affected by problems such as noise leakage, modal aliasing or limited resolution when facing non-stationary or ring stationary signals, and is difficult to maintain stable feature extraction effects under different working conditions. Especially in the early stages of fault evolution, weak impact signals are often covered by broadband noise, so that the identification capability of the traditional method on weak faults is limited. With the development of artificial intelligence technology, a signal processing method based on deep learning is gradually introduced into the field of fault diagnosis, and feature extraction and pattern recognition are automatically completed in an end-to-end learning mode, so that the dependence of artificial feature design is reduced to a certain extent. However, the existing deep learning method still has a plurality of defects in one-dimensional time sequence signal processing, namely, on one hand, the adaptability of a single-scale convolution structure to different time scale features is limited, transient impact and long-term evolution features are difficult to be simultaneously considered, and on the other hand, when part of network structures adopt jump connection to perform feature fusion, non-selection superposition is often performed on features from low layers, noise components are easy to be introduced into high-layer feature representation, and therefore denoising and diagnosis performance are influenced. In addition, the modeling capability of the partial model on the time sequence dependency relationship is insufficient, and it is difficult to effectively characterize the implicit periodicity and long-term related features in the ring stationary signal. Therefore, a signal processing and fault diagnosis method capable of performing collaborative optimization in aspects of multi-scale feature extraction, cross-layer feature fusion, time-dependent modeling and the like is needed to be oriented to the characteristics of one-dimensional time sequence signals, so that accurate extraction of effective signal components is realized in a complex noise environment, and a reliable data basis is provided for subsequent state monitoring and intelligent fault diagnosis. Disclosure of Invention In order to overcome the defects of the prior art, the invention provides a gating jump network denoising method integrating multi-scale space-time attention. The method realizes the self-adaptive denoising and characteristic reconstruction of the noisy one-dimensional time sequence signal by constructing an encoding-decoding network which fuses a multi-scale parallel convolution structure, a gate control jump connection mechanism and BiLSTM time sequence modeling units. The method comprises the steps of capturing local details and long period characteristics of a signal by utilizing multi-scale parallel convolution, embedding a gating jump connection mechanism in a decoding stage, carrying out channel splicing on shallow layer characteristics of the encoder and deep layer characteristics of a decoder, embedding a driving attention sub-network to generate a gating mask, carrying out self-adaptive weighted filtering on the shallow layer characteristics by utilizing the mask, and blocking the propagation of background noise to a decoding end before fusion. The invention can obt