Cyclostationary signal extraction method based on multi-scale gating attention mechanism and bi-lstm

Abstract

This invention discloses a method for extracting ring-stationary signals based on a multi-scale gated attention mechanism and BiLSTM, belonging to the field of mechanical fault diagnosis and signal processing technology. It aims to solve the problem of existing technologies struggling to accurately extract weak fault features under strong background noise and non-Gaussian anomalous interference environments. The method includes constructing a deep neural network model, capturing local waveforms of different receptive fields through parallel convolutional branches, and generating channel masks using an attention mechanism to achieve hierarchical adaptive suppression of background noise; embedding a BiLSTM in the bottleneck layer, utilizing its bidirectional memory units to perform full-sequence temporal modeling of compressed features, and accurately eliminating random anomalous pulses that do not conform to periodic logic through the long-term evolution law of sequence features; finally, using a decoder to perform nonlinear shaping and detail restoration on the upsampled features after transposed convolution. This method can efficiently filter out Gaussian white noise and non-Gaussian outliers, and reconstruct a pure ring-stationary fault impact signal with high fidelity.

Description

Technical Field

[0001] This invention belongs to the field of mechanical fault diagnosis and signal processing technology, specifically involving a method for extracting loop stationary signals based on multi-scale gated attention mechanism and BiLSTM. Background Technology

[0002] With the development of modern industry towards intelligence and large-scale operations, the health status monitoring of rotating machinery, such as aircraft engines, wind turbines, and high-speed train wheelsets, which serve as core power and transmission equipment, is crucial for ensuring the safe operation of the system. Rolling bearings and gearboxes, as the most vulnerable components in rotating machinery, have always presented a challenge in early fault diagnosis in the engineering field. Under actual operating conditions, fault signals generated by mechanical equipment typically exhibit significant second-order loop stationary characteristics, meaning their statistical characteristics change periodically over time. However, the impact signals generated by early faults are often extremely weak and heavily influenced by two types of interference: strong broadband background noise from the equipment's operating environment and non-Gaussian anomalous pulses originating from sensor loosening, electromagnetic interference, or transmission errors. Loop stationary signals, which are signals whose statistical characteristics change periodically over time, are commonly used in rotating machinery fault diagnosis to extract fault features.

[0003] Currently, existing technologies for enhancing and extracting fault features from stationary signals can be broadly categorized into two types: traditional signal processing methods and deep learning-based methods. Traditional signal processing methods are mainly represented by minimum entropy deconvolution, multi-point optimal minimum entropy deconvolution, and variational mode decomposition (EMD). Deconvolution methods aim to maximize the kurtosis of the signal by finding an inverse filter, thereby recovering the impulse component. However, these methods essentially pursue the maximization of impulse characteristics, making them highly susceptible to being misled by non-Gaussian anomalous impulses, often misinterpreting them as fault features for enhancement, leading to misdiagnosis. While adaptive decomposition methods overcome the mode aliasing problem of EMD, their performance is highly dependent on the selection of penalty factors and the number of modes. In noisy and prior-unknown conditions, parameter optimization is extremely difficult, making adaptive signal separation challenging. With the development of generative AI, existing deep learning methods, represented by generative adversarial networks (GANs) and Transformers, are beginning to be applied to signal enhancement and reconstruction. GAN-based methods reconstruct clean signals through a game between the generator and the discriminator.

[0004] However, the training process of GANs suffers from significant instability (Nash equilibrium is difficult to achieve) and is prone to mode collapse, which can lead to the generation of false artifact features and compromise the temporal authenticity of the original fault signal. While Transformer-based methods address long-distance dependencies using self-attention mechanisms, their global attention computational complexity is quadratic with the sequence length, resulting in extremely high computational resource consumption when processing high-frequency, long-sequence vibration signals. Furthermore, Transformers lack inductive biases for local translation invariance, and their ability to capture weak fault features with small samples is often inferior to that of specifically designed convolutional recurrent networks. These existing models often lack explicit multi-scale weighting mechanisms when handling "signal enhancement" and "feature extraction" tasks. The models typically process the input data in a "black box" manner, lacking a "gating mechanism" similar to human vision to actively suppress broadband noise and sudden outliers in the background. This results in reconstructed signals often retaining a large amount of noise or losing crucial loop-stationary modulation information under extremely low signal-to-noise ratio conditions.

[0005] Based on this, the present invention proposes a loop stationary signal extraction method based on multi-scale gated attention mechanism and BiLSTM to solve the problem that the existing technology is difficult to accurately extract weak fault features in environments with strong background noise and non-Gaussian anomalous interference (such as sensor glitches and electromagnetic pulses). Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a loop stationary signal extraction method based on a multi-scale gated attention mechanism and BiLSTM, which solves the problem that existing technologies struggle to accurately extract weak fault features under strong background noise and non-Gaussian anomaly interference environments.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for extracting loop-stationary signals based on multi-scale gated attention mechanism and BiLSTM includes:

[0009] Step 1: Construct a deep neural network model;

[0010] The deep neural network model adopts an "encoder-bottleneck layer-decoder" architecture, which includes an encoder, a bottleneck layer, and a decoder. The encoder is used to map the original one-dimensional vibration signal to a low-dimensional feature space. The bottleneck layer is located between the encoder and the decoder and is used to perform global temporal modeling of the encoded features. The decoder has a structure symmetrical to the encoder and is used to restore the temporal features to a time-domain waveform.

[0011] Step 2: Construct a sample database of stationary loop signals;

[0012] The database of stationary signal samples includes a test set and a training set;

[0013] Step 3: Train the deep neural network model established in Step 1 using the training set of the loop stationary signal sample database constructed in Step 2;

[0014] Step 4: Input the test set of the loop stationary signal sample database constructed in Step 2 into the deep neural network model trained in Step 3, and extract the denoised pure loop stationary signal.

[0015] In a preferred embodiment of the present invention, the encoder includes a plurality of multi-scale gated attention modules connected in series and a max pooling layer, with each multi-scale gated attention module followed by a pooling layer.

[0016] The bottleneck layer includes a one-dimensional convolutional adaptation layer, a ReLU activation layer, a BiLSTM, and a random deactivation layer with a dropout rate of 0.2;

[0017] The decoder consists of several transposed convolutional layers connected in series and a multi-scale gated attention module. The transposed convolutional layers are used to upsample and restore the signal resolution, and finally the reconstructed clean signal is output through the regression layer at the output end of the decoder.

[0018] In a preferred embodiment of the present invention, the process by which the encoder maps the original one-dimensional vibration signal to a low-dimensional feature space includes:

[0019] In the first stage, a multi-scale gated attention module is constructed, which relies on the classic Inception parallel structure to complete the hierarchical extraction and fusion of multi-scale features; a Sigmoid gate mechanism is introduced to perform adaptive weight allocation and feature calibration for feature mappings of different scales, and then combined with a max pooling layer to complete feature dimensionality reduction and spatial information aggregation, thereby compressing the length of the input one-dimensional vibration signal.

[0020] In the second stage, the same network structure as in the first stage is adopted, but the number of convolutional kernel groups is increased. This structure is used to further extract deep abstract features of the data and simultaneously complete feature dimensionality reduction.

[0021] In a preferred embodiment of the present invention, the process of the bottleneck layer performing global temporal modeling of the encoded features includes:

[0022] After the feature sequence passes through the adaptation convolutional layer and ReLU activation layer, the number of channels in the feature sequence is uniformly adjusted through the convolutional layer and ReLU activation layer.

[0023] In the input core BiLSTM, the BiLSTM includes a preset number of hidden units and is set to sequence output mode. By simultaneously utilizing forward and backward information to scan the feature sequence, it captures the global periodic evolution law of the loop stationary signal.

[0024] In a preferred embodiment of the invention, the ends of the BiLSTM are connected to a randomly deactivated layer with a set drop rate.

[0025] In a preferred embodiment of the present invention, the process by which the decoder recovers the timing features into a time-domain waveform includes:

[0026] In the first decoding stage, the feature sequence is upsampled through transposed convolutional layers for initial recovery, and the features are refined by a multi-scale gated attention module.

[0027] In the second decoding stage, the sequence is further upsampled to restore its original length through transposed convolution, and then connected to a multi-scale gating attention module to complete the restoration of deep features.

[0028] In a preferred embodiment of the present invention, the multi-scale gated attention module includes a parallel gated attention branch and a feature aggregation branch; the feature aggregation branch adopts the architecture of the classic Inception parallel structure, including multiple parallel one-dimensional convolution branches; the gated attention branch and the feature aggregation branch are cascaded to receive multi-scale feature maps as input, calculate the attention weights of the feature channels through multi-scale parallel convolution and channel-adaptive convolution operations of the gated attention branch, and finally generate gated coefficients through the Sigmoid activation function;

[0029] The gating coefficient has a numerical range of [0,1].

[0030] In a preferred embodiment of the present invention, the gated attention branch includes four parallel convolutional branches;

[0031] The first branch is a one-dimensional convolution with a kernel size of 1;

[0032] The second branch consists of one-dimensional convolutional convolution with a kernel size of 1 and one-dimensional dilated convolution with a kernel size of 3 and an inflation factor of 2.

[0033] The third branch consists of one-dimensional convolutional convolution with a kernel size of 1, and one-dimensional dilated convolution with a kernel size of 3 and an inflation factor of 4.

[0034] The fourth branch is a one-dimensional convolution with a max pooling layer concatenated convolution kernel size of 1.

[0035] In a preferred embodiment of the present invention, the gated attention branch receives the multi-scale feature map output by the feature aggregation branch, and generates a weight mask with the same size as the input feature map through the operation path of "dimensionality reduction convolution - ReLU activation - dimensionality increase convolution - Sigmoid activation".

[0036] In a preferred embodiment of the present invention, the weight mask has a numerical range of [0,1].

[0037] Compared with existing technologies, this invention provides a method for extracting loop stationary signals based on multi-scale gated attention mechanism and BiLSTM, which has the following beneficial effects:

[0038] This method introduces an innovative multi-scale gating attention mechanism, enabling the network to automatically "focus" on the effective frequency band where the fault occurs, just like human vision. It also actively suppresses background noise and randomly occurring non-Gaussian anomalous spikes through Sigmoid gating, solving the problem that traditional methods and ordinary CNNs cannot distinguish between fault impacts and interference pulses.

[0039] This method combines the advantages of the Inception module's multiple receptive fields (utilizing convolutional kernels with different dilation factors) with the long-term memory advantage of BiLSTM. The Inception module is responsible for capturing local details and multi-band features, while BiLSTM is responsible for capturing the periodic evolution of fault signals. The two complement each other, significantly improving the completeness of feature extraction.

[0040] This method eliminates the need for manually setting sensitive parameters such as mode number, penalty factor, or filter bandwidth, achieving automated end-to-end extraction from raw data to clean waveforms and reducing reliance on expert experience. It demonstrates good applicability and widespread application value in complex non-stationary data scenarios such as rotating machinery vibration signal processing. Attached Figure Description

[0041] Figure 1 This is a network structure diagram of the deep neural network model in this invention.

[0042] Figure 2 This is a network structure diagram of the multi-scale gating attention module of the present invention.

[0043] Figure 3 This is a network structure diagram of the classic Inception parallel architecture.

[0044] Figure 4 This is a diagram of the pure signal waveform in an embodiment of the present invention.

[0045] Figure 5 A waveform diagram of a noisy signal is constructed for an embodiment of the present invention.

[0046] Figure 6 This is a diagram showing the enhanced result of an embodiment of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] As per the instruction manual Figure 1 -Appendix Figure 6 As shown, this invention proposes a method for extracting ring-stationary signals based on a multi-scale gated attention mechanism and BiLSTM. The method specifically includes the following steps:

[0049] Step 1: Construct a deep neural network model (MGI-BiLSTM-DAE, i.e., multi-scale gated Inception-BiLSTM denoising autoencoder).

[0050] The deep neural network model constructed in this embodiment adopts an "encoder-bottleneck layer-decoder" architecture, with the specific hierarchical structure mainly including an encoder, a bottleneck layer, and a decoder. Wherein:

[0051] The encoder is used to map the raw one-dimensional vibration signal to a low-dimensional feature space. The encoder consists of several multi-scale gated attention modules connected in series with a max-pooling layer. Each multi-scale gated attention module is followed by a pooling layer for downsampling to compress the data length and extract abstract features.

[0052] The bottleneck layer, located between the encoder and decoder, is used for global temporal modeling of the encoded features. BiLSTM comprises a one-dimensional convolutional adaptation layer, a BiLSTM layer, and a random deactivation layer. The BiLSTM layer utilizes two state vectors, forward and backward, to scan the local feature sequence, capturing the long-range dependencies of stationary loop signals along the time axis.

[0053] The decoder has a structure symmetrical to the encoder and is used to recover temporal features into temporal waveforms. The decoder consists of several cascaded transposed convolutional layers and a multi-scale gated attention module. The transposed convolutional layers are used to upsample and restore signal resolution, and finally, a regression layer outputs the reconstructed clean signal.

[0054] The working principles of the encoder, bottleneck layer, and decoder mentioned above include:

[0055] Step 1.1: First, the encoder module is responsible for feature extraction and dimensionality reduction of the standardized one-dimensional vibration signal (input sample) with a length of 1024. This encoder module includes two cascaded encoding stages:

[0056] Step 1.1.1: In the first stage, a multi-scale gated attention module containing 8 convolutional kernel groups is constructed. This relies on the classic Inception parallel architecture (a core architecture for multi-scale feature extraction in deep learning; its core design philosophy is to simultaneously capture local details, mesoscale features, and global information of the input data by deploying convolutional kernels and pooling operations with different receptive fields in parallel, avoiding the limitations of single-scale convolution for feature extraction, and ultimately achieving a more comprehensive representation through feature fusion). Its structure diagram is shown below. Figure 3 As shown, multi-scale feature extraction and fusion are completed. The structure is transformed into a one-dimensional form, including four branches: the first branch uses a 1x1 convolution kernel (kernel size 1, stride 1) for dimensionality reduction; the second branch uses a 1x3 convolution kernel (kernel size 3, stride 1) to extract small-scale features; the third branch uses a 1x5 convolution kernel (kernel size 5, stride 1) to extract medium-scale features; the fourth branch uses a max pooling layer concatenated with a 1x1 convolution to capture global information; a Sigmoid gating mechanism is introduced to adaptively allocate weights and calibrate features at different scales, and then combined with a max pooling layer with a stride of 2 to complete feature dimensionality reduction and spatial information aggregation. Through the above multi-scale feature interaction and downsampling operations, the length of the input standardized one-dimensional vibration signal with a length of 1024 is compressed to 512.

[0057] Specifically, the "8 convolutional kernel groups" here correspond to the multi-size parallel convolutional kernels of the Inception structure (such as 1×1, 3×3, 5×5, 7×7 + dilated convolutional combinations). These 8 kernel groups can achieve finer-grained multi-scale feature capture. The output weight values ​​of "Sigmoid-gated weighting" are ∈ [0, 1], assigning high weights to effective feature branches and low weights to ineffective feature branches. "Max pooling with a stride of 2" typically uses a 2×2 pooling kernel with a stride of 2 = kernel size. This non-overlapping pooling efficiently compresses sequence length while avoiding feature information loss, making it the optimal choice for dimensionality reduction of one-dimensional sequence features. Compressing the length of the one-dimensional vibration signal to 512 can adapt to the temporal feature input requirements of subsequent BiLSTM, perfectly connecting the overall architecture of the multi-scale gated Inception bidirectional long short-term memory denoising autoencoder (MGI-BiLSTM-DAE).

[0058] Step 1.1.2: In the second stage, the same multi-scale gated attention module structure as in the first stage is adopted, and the number of convolutional kernel groups is expanded to 16. Relying on the isomorphic Inception parallel convolutional structure and Sigmoid gated weighting mechanism, the deep abstract features of the data are further mined and extracted. At the same time, the downsampling dimensionality reduction is completed by combining the max pooling layer with a stride of 2, and the sequence length of the feature signal is further compressed to 256.

[0059] Step 1.2: The bottleneck layer, as the key hub connecting encoding and decoding, is mainly used for long-term time-series dependency modeling;

[0060] Step 1.2.1: The feature sequence first passes through an adaptive convolutional layer with a kernel size of 1 and a ReLU activation layer to uniformly adjust the number of channels to 64;

[0061] Step 1.2.2: The input is then fed into the core BiLSTM, which includes 64 hidden units and is set to sequence output mode. By simultaneously utilizing forward and backward information to scan the feature sequence, it effectively captures the global periodic evolution law of the loop stationary fault signal. The end connection is a random deactivation layer with a dropout rate of 0.2 to prevent overfitting.

[0062] Step 1.3: Again, the decoder module adopts a symmetrical structure design with the encoder, aiming to achieve high-precision signal reconstruction; the encoder module includes two cascaded decoding stages:

[0063] Step 1.3.1: In the first decoding stage, the feature sequence is upsampled and restored to a length of 512 (64 channels) by a transposed convolutional layer with a kernel size of 2 and a stride of 2, and the features are refined and shaped using a multi-scale gated attention module with 16 convolutional kernel groups.

[0064] Step 1.3.2: In the second decoding stage, the sequence is further upsampled to the original length of 1024 (32 channels) by transposing convolution, and then connected to a multi-scale gated attention module including 8 convolutional kernel groups to complete the restoration of deep features;

[0065] Among them, such as Figure 2 As shown, the multi-scale gated attention module in the decoder includes parallel gated attention branches and feature aggregation branches. The feature aggregation branch adopts the architecture of the classic Inception parallel structure, including multiple parallel one-dimensional convolutional branches, which extract multi-receptive field features using convolutional kernels of different sizes and dilation factors, and then perform depth concatenation of the outputs of each branch in the channel dimension to form a multi-scale feature map;

[0066] The gated attention branch and the feature aggregation branch are cascaded to receive multi-scale feature maps as input. The multi-scale parallel convolution of the gated attention branch (corresponding to the four parallel convolution branches of the gated attention branch (1×1 convolution, 1×1+3×3 dilated convolution (dilation factor 2), 1×1+3×3 dilated convolution (dilation factor 4), max pooling + 1×1 convolution) is responsible for extracting multi-scale features and providing a basis for subsequent weight calculation) and channel-adaptive convolution operations (corresponding to the "dimensionality reduction convolution-ReLU activation-dimensionality increase convolution" operation path, the core function of which is to compress and restore the channel dimension of the multi-scale feature map, and then calculate the attention weight of each channel (i.e., the median value of the gate coefficient)) calculate the attention weight of the feature channels. Finally, the gate coefficients with values ​​in the range [0,1] are generated by the Sigmoid activation function.

[0067] It should be noted that the gated attention branch includes four parallel convolutional branches; the first branch uses a one-dimensional convolution with a kernel size of 1 to extract point-level features between channels; the second branch uses a one-dimensional convolution with a kernel size of 1 followed by a one-dimensional dilated convolution with a kernel size of 3 and a dilation factor of 2 to extract short-term local features; the third branch uses a one-dimensional convolution with a kernel size of 1 followed by a one-dimensional dilated convolution with a kernel size of 3 and a dilation factor of 4 to extract features with a large receptive field; and the fourth branch uses a max-pooling layer followed by a one-dimensional convolution with a kernel size of 1 to extract salient features.

[0068] In use, the outputs of the one-dimensional convolutional branches parallel to the gated attention branch are concatenated along the channel dimension to form a multi-scale feature map. Specifically: the gated attention branch receives the multi-scale feature map output from the feature aggregation branch and generates a weight mask with the exact same size as the input feature map through the operation path of "dimensionality reduction convolution - ReLU activation - dimensionality increase convolution - Sigmoid activation". The numerical range of this mask is [0,1]. The multi-scale feature map output from the feature aggregation branch is multiplied element-wise with the weight mask generated by the gated attention branch. Through weight modulation, the channel features corresponding to the effective signal are enhanced, and the channel features corresponding to noise are suppressed, outputting a weighted feature map.

[0069] Step 1.4: Finally, the output layer is responsible for mapping the high-dimensional features back to the time-domain waveform;

[0070] BiLSTM employs a one-dimensional reconstruction convolutional layer with a kernel size of 1 to fuse and map the decoded multi-channel features into a single-channel output. Finally, it outputs a clean, stable fault pulse waveform after removing background noise and abnormal interference through a regression layer, thus achieving end-to-end signal extraction.

[0071] Step 2: Construct a sample database of stationary loop signals;

[0072] The purpose of constructing a loop-stationary signal sample database is to enable the deep neural network model established in step 1 to extract fault features in a noisy environment. This embodiment constructs a loop-stationary signal sample database including strong background noise and abnormal impulse interference as network input, and uses the corresponding pure loop-stationary impulse signals as labels. Specifically, it includes:

[0073] Step 2.1: Signal parameter settings;

[0074] Set the signal sampling length to 1024 and the sampling frequency to 10000Hz;

[0075] Step 2.2: Signal generation model;

[0076] The clean signal (label) simulates the periodic fault impact of rotating machinery. The fault frequency is set to fluctuate randomly between [80, 120] Hz, and the resonant frequency is between [2000, 5000] Hz. To simulate the real working conditions, random phase jitter is introduced at the moment of fault impact to obtain the signal generation model.

[0077] Among them, the noisy signal (input) is superimposed with two types of interference on the basis of the clean signal: one is to adjust the signal-to-noise ratio of broadband Gaussian white noise to simulate strong background noise; the other is to randomly generate random pulses with amplitude values ​​of [0,3] (about 4 times the signal amplitude) to simulate outliers caused by sensor failure or electromagnetic interference.

[0078] Step 2.3: Sample size;

[0079] 1000 pairs of "noisy signal - clean signal" were generated as the training set, and 200 pairs were used as the test set. At the same time, real bearing datasets such as the CWRU bearing fault dataset were introduced, including fault signals of the inner ring, outer ring and rolling elements, totaling 500 pairs, for additional verification of the model's generalization ability.

[0080] Step 3: Use the training set of the ring stationary signal sample database constructed in Step 2 to train the deep neural network model established in Step 1; at the same time, during the training process, use the gated attention mechanism to automatically learn the channel weights of multi-scale features, and optimize the weight coefficients through backpropagation to complete the training of the deep neural network model and achieve accurate focusing on the fault pulse components.

[0081] The training set of the circular stationary signal sample database is input into the deep neural network model for supervised regression learning. By optimizing the reconstruction loss function, the network parameters are adaptively updated, enabling it to master the nonlinear mapping ability to remove coupled noise from high-noise backgrounds and identify transient impact components.

[0082] Specifically, given the long training time of deep recurrent networks, this embodiment adopts a parallel computing strategy: the Adam optimizer is used with an initial learning rate of 0.001 and a segmented decay strategy is adopted (decaying by 10% every 10 rounds); the training data is distributed to each computing core, the gradient is calculated synchronously and the network weights are updated until the model converges (the maximum number of iterations is set to 100).

[0083] Step 4: Input the test set of the loop-stationary signal sample database constructed in Step 2 (containing signals with strong noise (SNR=-8dB) and abnormal spikes) into the trained deep neural network model to extract the denoised pure loop-stationary signal (pure fault pulse waveform after removing noise and abnormal interference). The test results are as follows. Figures 4-6 As shown. Wherein:

[0084] Figure 4 The ideal ring-stationary fault signal is shown, characterized by a series of periodic impulsive pulses with steep rising edges and exponentially decaying tails, with amplitudes stable between 1.0, representing the "true value" target under undisturbed conditions.

[0085] Figure 5 The diagram illustrates network input data under simulated harsh operating conditions. The original periodic pulses are almost completely submerged by large-amplitude broadband Gaussian white noise, resulting in an extremely low signal-to-noise ratio. Furthermore, a non-Gaussian anomalous spike with an amplitude as high as -4.5 exists at approximately 630 on the horizontal axis, simulating a sudden outlier caused by sensor failure or electromagnetic interference. Figure 5 As shown, aperiodic large-amplitude abnormal pulses in the input signal often mislead traditional deconvolution algorithms.

[0086] exist Figure 6 In the output waveform, the abnormal spikes at the corresponding positions are completely removed, leaving no artifacts.

[0087] The above results verify the effectiveness of the core gating attention mechanism of this invention: this mechanism, through global context awareness, identifies the spike as an isolated non-cyclic stationary event, and thus automatically generates a weight mask close to 0 through the Sigmoid activation unit, actively blocking and filtering the abnormal channel. Figure 6 The reconstructed output and Figure 4 Comparison reveals that the output signal not only perfectly aligns with the real pulse in phase with no time offset, but also perfectly reproduces the asymmetric physical characteristics of the fault impact—"rapid rise, slow decay"—in terms of waveform morphology. This is attributed to the BiLSTM structure in the bottleneck layer, which establishes a long-distance dependency through bidirectional forward and backward time scanning. Based on the global periodicity of the signal, it can accurately infer and recover the weak fault impact component from a noisy background, achieving high-fidelity end-to-end feature extraction.

[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for extracting stationary loop signals based on multi-scale gated attention mechanism and BiLSTM, characterized in that, include: Step 1: Construct a deep neural network model; The deep neural network model adopts an "encoder-bottleneck layer-decoder" architecture, which includes an encoder, a bottleneck layer, and a decoder; the encoder is used to map the original one-dimensional vibration signal to a low-dimensional feature space. The bottleneck layer is located between the encoder and the decoder and is used for global temporal modeling of the encoded features; The decoder has a structure symmetrical to the encoder and is used to recover temporal features into time-domain waveforms; The encoder consists of several multi-scale gated attention modules connected in series and a max pooling layer, with each multi-scale gated attention module followed by a pooling layer. The bottleneck layer includes a one-dimensional convolutional adaptation layer, a ReLU activation layer, a BiLSTM, and a random deactivation layer with a dropout rate of 0.2; The decoder consists of several transposed convolutional layers connected in series and a multi-scale gated attention module. The transposed convolutional layers are used to upsample and restore the signal resolution, and finally the reconstructed clean signal is output through the regression layer at the output end of the decoder. The multi-scale gated attention module includes parallel gated attention branches and feature aggregation branches; The feature aggregation branch adopts the classic Inception parallel architecture, which includes multiple parallel one-dimensional convolution branches. The gated attention branch is cascaded with the feature aggregation branch and is used to receive multi-scale feature maps as input. The attention weights of the feature channels are calculated through multi-scale parallel convolution and channel-adaptive convolution operations of the gated attention branch. Finally, the gated coefficients are generated through the Sigmoid activation function. The gating coefficient has a numerical range of [0,1]. The gated attention branch receives the multi-scale feature map output by the feature aggregation branch and generates a weight mask with the same size as the input feature map through the operation path of "dimensionality reduction convolution - ReLU activation - dimension increase convolution - Sigmoid activation". Step 2: Construct a sample database of stationary loop signals; The database of stationary signal samples includes a test set and a training set; Step 3: Train the deep neural network model established in Step 1 using the training set of the loop stationary signal sample database constructed in Step 2; Step 4: Input the test set of the loop stationary signal sample database constructed in Step 2 into the deep neural network model trained in Step 3, and extract the denoised pure loop stationary signal.

2. The method for extracting ring stationary signals based on multi-scale gated attention mechanism and BiLSTM as described in claim 1, characterized in that, The process by which an encoder maps a raw one-dimensional vibration signal to a low-dimensional feature space includes: In the first stage, a multi-scale gated attention module is constructed, which relies on the classic Inception parallel structure to complete the hierarchical extraction and fusion of multi-scale features; a sigmoid gate mechanism is introduced to perform adaptive weight allocation and feature calibration for feature mappings of different scales, and then combined with a max pooling layer to complete feature dimensionality reduction and spatial information aggregation, thereby compressing the length of the input one-dimensional vibration signal. In the second stage, the same network structure as in the first stage is adopted, but the number of convolutional kernel groups is increased. This structure is used to further extract deep abstract features of the data and simultaneously complete feature dimensionality reduction.

3. The method for extracting ring stationary signals based on multi-scale gated attention mechanism and BiLSTM as described in claim 1, characterized in that, The bottleneck layer's process of performing global temporal modeling of encoded features includes: After the feature sequence passes through the adaptation convolutional layer and ReLU activation layer, the number of channels in the feature sequence is uniformly adjusted through the convolutional layer and ReLU activation layer. In the input core BiLSTM, the BiLSTM includes a preset number of hidden units and is set to sequence output mode. By simultaneously utilizing forward and backward information to scan the feature sequence, it captures the global periodic evolution law of the loop stationary signal.

4. The method for extracting ring stationary signals based on multi-scale gated attention mechanism and BiLSTM as described in claim 3, characterized in that, BiLSTM's end connections set a random deactivation layer with a drop rate.

5. The method for extracting ring stationary signals based on multi-scale gated attention mechanism and BiLSTM as described in claim 1, characterized in that, The process by which the decoder recovers the time-domain waveform from the timing features includes: In the first decoding stage, the feature sequence is upsampled through transposed convolutional layers for initial recovery, and the features are refined by a multi-scale gated attention module. In the second decoding stage, the sequence is further upsampled to restore its original length through transposed convolution, and then connected to a multi-scale gating attention module to complete the restoration of deep features.

6. The method for extracting ring stationary signals based on multi-scale gated attention mechanism and BiLSTM as described in claim 1, characterized in that, The gated attention branch consists of four parallel convolutional branches; The first branch is a one-dimensional convolution with a kernel size of 1; The second branch consists of one-dimensional convolutional convolution with a kernel size of 1 and one-dimensional dilated convolution with a kernel size of 3 and an inflation factor of 2. The third branch consists of one-dimensional convolutional convolution with a kernel size of 1, and one-dimensional dilated convolution with a kernel size of 3 and an inflation factor of 4. The fourth branch is a one-dimensional convolution with a max pooling layer concatenated convolution kernel size of 1.

7. The method for extracting ring stationary signals based on multi-scale gated attention mechanism and BiLSTM as described in claim 1, characterized in that, The weight mask has a numerical range of [0,1].

Images