Gated jump connection network denoising method fusing multi-scale space-time attention

Abstract

The invention discloses a gated jump connection network denoising method fusing multi-scale space-time attention. According to the method, a coding-decoding network fusing a multi-scale parallel convolution structure, a gating jump connection mechanism and a BiLSTM time sequence modeling unit is constructed, so that adaptive denoising and feature reconstruction of a noisy one-dimensional time sequence signal are realized. Wherein the encoder captures local details and long-period characteristics of a signal by using multi-scale parallel convolution; in the decoding stage, a gating jump connection mechanism is embedded, and channel splicing is carried out on shallow layer features of an encoder and deep layer features of a decoder; a driver attention sub-network is embedded to generate a gating mask, the mask is used to carry out adaptive weighted filtering on shallow layer features, and background noise is blocked from spreading to a decoding end before fusion. According to the method, multi-scale representation considering local transient change and long-term evolution trend can be obtained, the input quality of subsequent feature extraction, pattern recognition and life evaluation algorithms is remarkably improved, and the accuracy and reliability of a fault diagnosis process are improved.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, specifically relating to a denoising method for gated jump-connection networks that integrates multi-scale spatiotemporal attention. Background Technology

[0002] In the operation of complex industrial equipment and electromechanical systems, many critical components are subjected to long-term variable loads, high noise, and strong interference environments. Their operating status is often indirectly characterized by one-dimensional time-series signals such as vibration, current, acoustics, or pressure. Because the characteristic signals generated in the early stages of a fault have small amplitudes and short durations, and are often superimposed with background noise, operating condition fluctuations, and structural resonance, effective information is difficult to directly separate from the original measurement signals, posing a significant challenge to condition monitoring and fault diagnosis. Therefore, accurately extracting key signal components related to equipment health status against a strong noise background has always been a key research focus in the fields of digital signal processing and intelligent fault diagnosis.

[0003] Traditional one-dimensional signal analysis methods often rely on manually designed filters, time-frequency analysis, or statistical feature extraction techniques, such as bandpass filtering and empirical wavelet decomposition. These methods typically require manual parameter adjustment based on prior experience and are susceptible to problems such as noise leakage, mode aliasing, or resolution limitations when dealing with non-stationary or loop-stationary signals, making it difficult to maintain stable feature extraction results under different operating conditions. Especially in the early stages of fault evolution, weak impulse signals are often masked by broadband noise, limiting the ability of traditional methods to identify weak faults.

[0004] With the development of artificial intelligence technology, deep learning-based signal processing methods have been gradually introduced into the field of fault diagnosis. These methods automatically complete feature extraction and pattern recognition through end-to-end learning, reducing reliance on manual feature design to some extent. However, existing deep learning methods still have several shortcomings in one-dimensional time-series signal processing: on the one hand, single-scale convolutional structures have limited adaptability to features at different time scales, making it difficult to simultaneously address transient impacts and long-term evolutionary features; on the other hand, some network structures, when using skip connections for feature fusion, often perform unselective superposition of features from lower layers, easily introducing noise components into higher-level feature representations, thus affecting denoising and diagnostic performance. Furthermore, some models lack the ability to model temporal dependencies, making it difficult to effectively characterize the periodicity and long-term correlation features implicit in stationary loop signals.

[0005] Therefore, there is an urgent need for a signal processing and fault diagnosis method that can be optimized in multi-scale feature extraction, cross-layer feature fusion and time-dependent modeling for the characteristics of one-dimensional time-series signals, so as to achieve accurate extraction of effective signal components in complex noise environments and provide a reliable data foundation for subsequent condition monitoring and intelligent fault diagnosis. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a denoising method using a gated skip connection network that integrates multi-scale spatiotemporal attention. This method constructs an encoder-decoder network that integrates a multi-scale parallel convolutional structure, a gated skip connection mechanism, and a BiLSTM temporal modeling unit to achieve adaptive denoising and feature reconstruction of noisy one-dimensional temporal signals. Specifically, the encoder uses multi-scale parallel convolution to capture local details and long-period features of the signal; the decoding stage embeds a gated skip connection mechanism to concatenate shallow features from the encoder with deep features from the decoder; and a driving attention sub-network is embedded to generate a gated mask, which is used to adaptively weight and filter shallow features, blocking the propagation of background noise to the decoding end before fusion. This invention can obtain a multi-scale representation that takes into account both local transient changes and long-term evolution trends, significantly improving the input quality of subsequent feature extraction, pattern recognition, and lifetime assessment algorithms, and enhancing the accuracy and reliability of the entire fault diagnosis process at the system level.

[0007] The technical solution adopted by this invention to solve its technical problem is as follows: Step 1: Collect data and build a dataset; First, based on the requirements, a dataset of mechanical vibration signals is collected and stored, and finally, training and testing sets are generated, and the data is normalized and preprocessed. Step 2: Build an encoder based on multi-scale gated Inception; First, a one-dimensional time-domain signal is received through a sequence input layer, and then two multi-scale gated InceptionMGI coding modules and a max pooling layer are connected in series. The MGI encoding module internally employs four parallel feature extraction branches: a 1×1 convolution branch, a 3×3 convolution branch with a dilation rate of 2, a 3×3 convolution branch with a dilation rate of 4, and a max-pooling branch. These branches are used to simultaneously capture local details and long-period features at different scales in the signal. The outputs of the four feature extraction branches are depthwise concatenated and then pass through an internal gated branch consisting of a convolutional layer and a sigmoid activation function to generate an attention mask. This mask is then multiplied element-wise with the original concatenated features, thereby achieving adaptive suppression of background noise in the initial stage of feature extraction. The feature dimensionality is further reduced by a subsequent max-pooling layer. Step 3: Construct a time-series feature bottleneck layer based on BiLSTM; First, a 1×1 convolutional layer is used to adjust the feature map output by the encoder to the appropriate dimension, and then a BiLSTM layer containing 64 hidden units is connected. The BiLSTM layer uses a bidirectional recurrent mechanism to use waveform information before and after the current time step to determine the signal properties. Finally, a dropout layer is used to prevent the model from overfitting. Step 4: Construct a decoder containing gated hop connections; The shallow features of the encoder and the current deep features of the decoder are concatenated and fed into an attention generation subnetwork. After convolution and sigmoid activation, a gating mask with a value between 0 and 1 is generated. This gating mask is used to perform weighted filtering on the shallow features of the encoder, allowing only components related to fault features to pass through, thereby blocking the propagation of background noise to the decoder. The filtered features are then fused with the decoder features and fed into the subsequent MGI decoding module. The shallow features refer to features containing details but accompanied by noise; the deep features refer to features containing pure semantics. Step 5: Output layer construction; The network output is set with a convolutional layer with a kernel size of 1 to compress multi-channel features into a single-channel output, and then connected to the MGI decoding module to reconstruct the final temporal waveform; Step 6: Configure network training hyperparameters and execute model training; Configure the Adam optimizer as the network parameter update algorithm, input the constructed training set into the network, and iteratively update the network weights by minimizing the regression error between the predicted waveform and the pure label until the loss function converges.

[0008] Preferably, in step 1, the signal sampling rate is set to 10000Hz, the length of a single sample signal is 1024 data points, and the fault characteristic frequency fluctuates randomly between 80Hz and 120Hz.

[0009] Preferably, the training set includes 2000 samples and the test set includes 200 samples.

[0010] Preferably, the discard rate of the discard layer is 0.2.

[0011] Preferably, in step 5, the initial learning rate is set to 0.001, and a segmented decay strategy is adopted, reducing the learning rate to 10% of the original rate every 15 rounds of training; the maximum number of training rounds is set to 30 rounds, and the small batch sample size is 64.

[0012] An electronic device includes: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the above-described gating jump network denoising method.

[0013] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described gating jumper network denoising method.

[0014] A chip includes a processor for retrieving and running a computer program from a memory, causing a device equipped with the chip to perform the above-described gating jumper network denoising method.

[0015] A computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the above-described gating jump-connection network denoising method.

[0016] The beneficial effects of this invention are as follows: (1) Enhance signal reconstruction capability under strong noise background. This invention uses a multi-scale parallel feature extraction structure to simultaneously analyze one-dimensional time-series signals under different time receptive fields, effectively covering multiple feature forms such as transient impact, local changes and long-term evolution, so that the network can still retain key structural information under strong noise conditions, significantly improving the integrity and readability of the denoised signal.

[0017] (2) Suppressing noise propagation in cross-layer feature fusion. By introducing a gated skip connection mechanism between the encoding and decoding stages, this invention adaptively filters the features in the encoding stage before cross-layer feature transmission, avoiding the problem of non-selective superposition of noise features in traditional skip connections. This reduces the interference of noise transmission on the signal reconstruction result from a structural perspective and improves the stability of the denoising process.

[0018] (3) Enhance the ability to preserve weak fault features and loop stationary characteristics. Introduce a time-series modeling unit into the network to perform forward and backward joint modeling of the feature sequence, so that the model can more fully characterize the periodic structure and long-term dependency in the one-dimensional time-series signal, which helps to preserve weak impact features and early fault information during the denoising process and improve the ability to perceive weak anomalies.

[0019] (4) Reduce reliance on manual parameters and prior experience. This invention adopts an end-to-end supervised learning approach to achieve automatic mapping from the original noisy signal to the denoised signal. There is no need to manually design filter parameters or manually select feature forms, which reduces the reliance of traditional signal processing methods on expert experience and improves the adaptability of the method under different equipment types and operating conditions.

[0020] (5) Applicable to various one-dimensional time-series signals and fault diagnosis scenarios. The method does not impose specific restrictions on signal type and noise form, and can be widely applied to the preprocessing of various one-dimensional time-series signals such as vibration, current, and acoustics. It has good versatility in condition monitoring and intelligent fault diagnosis of various types of equipment such as rotating machinery, hydraulic systems, and motor drive devices.

[0021] By improving the signal-to-noise ratio and structural clarity of the denoised signal, this invention can significantly improve the input quality of subsequent feature extraction, pattern recognition, and lifetime assessment algorithms, thereby enhancing the accuracy and reliability of the entire fault diagnosis process at the system level. Attached Figure Description

[0022] Figure 1 The present invention provides a structural diagram of the neural network proposed in the patent. Figure 2 Here is a structural diagram of the MGI encoding module; Figure 3 The pure signal in the example; Figure 4 To construct a noisy signal; Figure 5 The signal is filtered by a neural network. Figure 6 This is the envelope spectrum of the denoised signal. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] To address the challenges of effectively separating weak features from one-dimensional time-series signals in noisy environments, the strong parameter dependence of traditional digital signal processing methods, and the tendency of existing deep learning models to introduce noise during multi-scale feature representation and cross-layer feature fusion, this invention aims to provide an intelligent denoising and feature extraction method for one-dimensional time-series signals under complex operating conditions. By introducing multi-scale feature extraction, gated cross-layer feature filtering, and time-dependent modeling mechanisms into a deep network structure, adaptive enhancement of effective signal components and suppression of noise components are achieved, thereby improving signal reconstruction quality and the reliability and stability of subsequent fault diagnosis.

[0025] This invention provides a one-dimensional time-series signal processing method based on deep neural networks. This method constructs an encoder-decoder network that integrates multi-scale parallel convolutional structures, gated skip connections, and time-series modeling units to achieve adaptive denoising and feature reconstruction of noisy one-dimensional time-series signals. The network as a whole adopts a structure combining hierarchical feature compression and stepwise feature recovery. In the encoding stage, a multi-scale feature extraction unit is set up. Signal features are extracted from different receptive fields at different times through multiple parallel convolutional branches, and the features at each scale are fused to obtain a multi-scale representation that takes into account both local transient changes and long-term evolution trends.

[0026] A cross-layer feature transfer structure is set up between the encoding and decoding stages, and a gated skip connection mechanism is introduced during the cross-layer feature transfer process. This mechanism dynamically filters multi-scale features from the encoding stage based on the feature information of the current decoding stage. Only feature components related to the current reconstruction target are introduced into the decoding process, thereby effectively suppressing the propagation of noise features in cross-layer fusion and avoiding interference from low-level noise to the high-level reconstruction results.

[0027] A timing modeling unit is set in the intermediate layer of the encoder-decoder structure. By performing forward and backward joint modeling of the feature sequence, the network's ability to express long-term dependencies, periodic structures, and loop stationary features in one-dimensional time-series signals is enhanced, and the ability to perceive and retain weak fault features is improved.

[0028] During the network training phase, a training sample set consisting of noisy time-series signals and their corresponding reference signals is constructed. Supervised learning is employed for end-to-end network training. By minimizing the reconstruction error between the network output and the reference signal, the network parameters are jointly optimized, enabling the network to gradually learn a nonlinear mapping relationship from noisy input to low-noise output.

[0029] After network training is complete, the one-dimensional time-series signal to be analyzed is input into the network model. The denoised signal result, processed by multi-scale feature filtering, gated cross-layer fusion, and time-series modeling, is directly output and can be used as input data for subsequent fault diagnosis, condition assessment, or health monitoring.

[0030] Example: This embodiment takes one-dimensional temporal feature enhancement as the application background and constructs a deep neural network that integrates multi-scale feature extraction, gated cross-layer feature screening, and temporal dependency modeling to recover the target signal from a noisy background.

[0031] Step 1: Construct a mechanical fault simulation dataset under strong noise background.

[0032] This embodiment first establishes a generation model for mechanical vibration signals, setting the signal sampling rate to 10000Hz and the length of a single sample signal to 1024 data points. Mechanical fault impact is simulated through a cyclic stationary process, where the fault characteristic frequency fluctuates randomly between 80Hz and 120Hz, and random time jitter is introduced to approximate real-world working conditions. Subsequently, Gaussian white noise is superimposed on the clean signal, and large outlier anomalies are randomly introduced with a certain probability. Finally, a training set containing 2000 samples and a test set containing 200 samples are generated, completing the data normalization preprocessing.

[0033] Step 2: Build an encoder network based on multi-scale gating Inception.

[0034] In the encoding stage of constructing the deep neural network, a one-dimensional time-domain signal is first received through a sequence input layer, followed by concatenation of two multi-scale gated Inception (MGI) encoding modules and a max-pooling layer. The MGI encoding module internally sets up four parallel feature extraction branches: a 1×1 convolution branch, a 3×3 convolution branch with a dilation rate of 2, a 3×3 convolution branch with a dilation rate of 4, and a max-pooling branch. These are used to simultaneously capture local details and long-period features at different scales in the signal. The outputs of the four branches are depthwise concatenated and then pass through an internal gated branch consisting of a convolutional layer and a sigmoid activation function to generate an attention mask. This mask is then multiplied element-wise with the original concatenated features, thus achieving adaptive suppression of background noise in the initial stage of feature extraction. The subsequent max-pooling layer further reduces the feature dimensionality.

[0035] Step 3: Construct a time-series feature bottleneck layer based on BiLSTM.

[0036] A temporal modeling layer is embedded between the encoder and decoder to handle the time dependence of mechanical signals. First, a 1×1 convolutional layer is used to adjust the feature map output by the encoder to the adaptation dimension, followed by a BiLSTM layer containing 64 hidden units. This layer uses a bidirectional recurrent mechanism to comprehensively utilize waveform information before and after the current time to determine the signal properties, effectively enhancing the model's logical reasoning ability for continuous fault impact sequences. Finally, a dropout layer with a dropout rate of 0.2 is used to prevent the model from overfitting.

[0037] Step 4: Construct a decoder network containing gated jump connections.

[0038] The decoding stage gradually restores the temporal resolution of the signal through transposed convolutional layers and introduces an innovative gated skip connection mechanism to replace the traditional direct splicing. The specific implementation of this mechanism is as follows: the shallow features of the corresponding layer of the encoder (containing rich details but accompanied by noise) are concatenated with the current deep features of the decoder (containing pure semantics), and fed into an attention generation sub-network. After convolution and sigmoid activation, a gated mask with a value between 0 and 1 is generated. This mask is used to perform weighted filtering of the shallow features of the encoder, allowing only components related to fault features to pass through, thereby blocking the propagation of background noise to the decoding end. The filtered features are then fused with the decoder features and fed into the subsequent MGI decoding module.

[0039] Step 5: Output layer construction.

[0040] The network output is set with a convolutional layer with a kernel size of 1 to compress multi-channel features into a single-channel output, and then connected to a regression layer to reconstruct the final temporal waveform.

[0041] Step 6: Configure network training hyperparameters and execute model training.

[0042] The Adam optimizer is configured as the algorithm for updating network parameters. The initial learning rate is set to 0.001, and a piecewise decay strategy is adopted, reducing the learning rate to 10% of the original rate every 15 training epochs. The maximum number of training epochs is set to 30, and the mini-batch sample size is 64. The constructed training set is input into the network, and the network weights are iteratively updated by minimizing the regression error between the predicted waveform and the pure label until the loss function converges.

[0043] Step 7: Verify the noise reduction effect.

[0044] After model training, samples with a signal-to-noise ratio of -12dB are extracted from the test set for validation. Noisy signals are input into the trained network, and the reconstructed waveform output by the network effectively removes Gaussian white noise and anomalous interference, clearly restoring the periodic fault impact components submerged by noise, thus achieving blind denoising of mechanical signals in noisy environments. The overall structure of the neural network is as follows: Figure 1 As shown, the MGI module is as follows Figure 2 As shown.

[0045] Figure 3 The ideal stationary fault signal is presented. It can be seen that the signal exhibits a distinct periodic transient impact structure throughout the entire sampling sequence, with each impact event repeating at approximately constant time intervals on the time axis. Within each period, the signal exhibits a damped oscillatory response following the impact excitation, with its amplitude gradually decaying and stabilizing over time, demonstrating typical transient response characteristics of a damped system.

[0046] Figure 4 The waveform in the time domain is a noisy signal with broadband noise and random abnormal interference superimposed on a pure, stationary signal. It can be observed that the originally regular periodic impact structure is significantly submerged in the strong noise background, and the overall signal exhibits approximately random fluctuation characteristics. Simultaneously, the signal also contains a small number of random pulse interferences with significantly larger amplitudes, making it difficult to directly distinguish transient impacts from noise in the time domain. The periodic characteristics are no longer intuitively discernible, reflecting typical characteristics of signals acquired under complex operating conditions.

[0047] Figure 5 This is the time-domain result after filtering the noisy signal. Figure 4 In contrast, broadband random noise in the filtered signal is significantly suppressed, and the periodic transient impact structure reappears. The damped oscillation response within each period gradually recovers, and the impact locations exhibit a more regular distribution characteristic on the time axis. Although some residual noise still exists, the overall structure of the signal is clearer, and the periodic boundaries are more defined, providing a more favorable time-domain basis for subsequent feature analysis.

[0048] Figure 6 for Figure 5 The envelope spectrum results show that distinct characteristic frequency peaks appear in the envelope spectrum, with the main peak corresponding to the fault characteristic frequency of the signal. The integer harmonics (such as the 2nd and 3rd harmonics) also exhibit prominent spectral peaks. This spectral distribution reflects the modulation characteristics of the periodic impulse components in the signal, indicating that the periodic structure contained in the signal is effectively preserved and enhanced after filtering, which is beneficial for the identification and analysis of potential fault features in loop-stationary signals.

Claims

1. A denoising method for gated jump networks integrating multi-scale spatiotemporal attention, characterized in that, Includes the following steps: Step 1: Collect data and build a dataset; First, based on the requirements, a dataset of mechanical vibration signals is collected and stored, and finally, training and testing sets are generated, and the data is normalized and preprocessed. Step 2: Build an encoder based on multi-scale gated Inception; First, a one-dimensional time-domain signal is received through a sequence input layer, and then two multi-scale gated Inception MGI encoding modules are connected in series with a max pooling layer. The MGI encoding module internally employs four parallel feature extraction branches: a 1×1 convolution branch, a 3×3 convolution branch with a dilation rate of 2, a 3×3 convolution branch with a dilation rate of 4, and a max-pooling branch. These branches are used to simultaneously capture local details and long-period features at different scales in the signal. The outputs of the four feature extraction branches are depthwise concatenated and then pass through an internal gated branch consisting of a convolutional layer and a sigmoid activation function to generate an attention mask. This mask is then multiplied element-wise with the original concatenated features, thereby achieving adaptive suppression of background noise in the initial stage of feature extraction. The feature dimensionality is further reduced by a subsequent max-pooling layer. Step 3: Construct a time-series feature bottleneck layer based on BiLSTM; First, a 1×1 convolutional layer is used to adjust the feature map output by the encoder to the appropriate dimension, and then a BiLSTM layer containing 64 hidden units is connected. The BiLSTM layer uses a bidirectional recurrent mechanism to use waveform information before and after the current time step to determine the signal properties. Finally, a dropout layer is used to prevent the model from overfitting. Step 4: Construct a decoder containing gated hop connections; The shallow features of the encoder and the current deep features of the decoder are concatenated and fed into an attention generation subnetwork. After convolution and sigmoid activation, a gating mask with a value between 0 and 1 is generated. This gating mask is used to perform weighted filtering on the shallow features of the encoder, allowing only components related to fault features to pass through, thereby blocking the propagation of background noise to the decoder. The filtered features are then fused with the decoder features and fed into the subsequent MGI decoding module. The shallow features refer to features containing details but accompanied by noise; the deep features refer to features containing pure semantics. Step 5: Output layer construction; The network output is set with a convolutional layer with a kernel size of 1 to compress multi-channel features into a single-channel output, and then connected to the MGI decoding module to reconstruct the final temporal waveform; Step 6: Configure network training hyperparameters and execute model training; Configure the Adam optimizer as the network parameter update algorithm, input the constructed training set into the network, and iteratively update the network weights by minimizing the regression error between the predicted waveform and the pure label until the loss function converges.

2. The denoising method for gated skip networks incorporating multi-scale spatiotemporal attention as described in claim 1, characterized in that, In step 1, the signal sampling rate is set to 10000Hz and the length of a single sample signal is 1024 data points. The fault characteristic frequency fluctuates randomly between 80Hz and 120Hz.

3. The denoising method for gated skip networks incorporating multi-scale spatiotemporal attention as described in claim 1, characterized in that, The training set includes 2000 samples, and the test set includes 200 samples.

4. The denoising method for gated skip networks incorporating multi-scale spatiotemporal attention as described in claim 1, characterized in that, The discard rate of the discard layer is 0.

2.

5. The denoising method for gated skip networks incorporating multi-scale spatiotemporal attention as described in claim 1, characterized in that, In step 5, the initial learning rate is set to 0.001, and a segmented decay strategy is adopted, reducing the learning rate to 10% of the original rate every 15 rounds of training. The maximum number of training rounds is set to 30, and the mini-batch sample size is 64.

6. An electronic device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5.

8. A chip, characterized in that, include: A processor for retrieving and running a computer program from memory, causing a device on which the chip is mounted to perform the method as described in any one of claims 1 to 5.

9. A computer program product, characterized in that, The computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the method as described in any one of claims 1 to 5.

Images