Operating personnel physiological signal denoising method for multiple electric power scenes

Abstract

The invention discloses an operator physiological signal denoising method for multiple electric power scenes, relates to the field of electric power signal denoising, and forms a collaborative mechanism of multi-scale feature extraction and dynamic feature focusing through the design and combination of a context contrast module CCB and a CBAM module, thereby realizing accurate processing for different noise types. Specifically, the CCB module provides stronger feature modeling capability through multi-scale context fusion in a bottleneck layer; and the CBAM module performs dynamic weighting on the features by using channels and space attention, and optimizes the processing capability of the model on a key signal area. The invention provides a set of efficient de-noising technical scheme aiming at complex characteristics such as multi-noise superposition and high dynamic interference in an electric power working environment.

Description

[0001] Methodology Area

[0002] The present invention relates to the field of power signal denoising, and in particular to a method for denoising physiological signals of operators in multiple power scenarios.

[0003] Background Methods

[0004] In recent years, deep learning technology has made significant progress in the field of physiological signal denoising. The denoising autoencoder (DAE) achieves good noise reduction effects by learning the potential structure of the signal. However, this type of method is prone to losing key detail features in the signal, resulting in partial information loss. In addition, most deep learning models are usually designed for a single type of signal or a single type of noise, and it is difficult to adapt to the complex situation of multiple signals and noise superposition in power operation scenarios. For example, in high-altitude operations, high-frequency power frequency noise, electromyographic noise, and baseline drift often exist at the same time, which further increases the difficulty of signal denoising. Therefore, the development of more efficient multi-signal and multi-noise denoising methods for complex power operation scenarios has become a key issue that needs to be urgently addressed in the field of physiological signal processing. The existing research has the following main deficiencies:

[0005] (1) There is a lack of adaptive denoising methods for multiple noise superposition, which makes it difficult to maintain efficient denoising performance in complex scenes;

[0006] (2) Existing deep learning methods are not capable of jointly denoising multiple signals and multiple noises, and cannot meet the requirements of high dynamics and high real-time performance in power operation scenarios.

[0007] Therefore, a method for denoising physiological signals of operators in multiple power scenarios is provided to solve the above problems. Summary of the invention

[0008] The purpose of the present invention is to provide a method for denoising physiological signals of operators in multiple power scenarios. By introducing a channel-spatial attention module CBAM, a context comparison module CCB and a multi-scale convolution structure, combined with an improved U-Net architecture, the denoising effect is significantly improved.

[0009] To achieve the above objectives, the present invention provides a method for denoising physiological signals of operators in multiple power scenarios. By introducing a channel-spatial attention module CBAM, a context comparison module CCB and a multi-scale convolution structure, a denoising model CBAM-UNet based on an improved U-Net architecture is designed in combination with an improved U-Net architecture. The method specifically includes the following steps:

[0010] Step S1: The input signal is normalized and filtered through the preprocessing module to reduce the impact of baseline drift and high-frequency noise on the model;

[0011] Step S2: The preprocessed signal enters the encoder part of CBAM-UNet, and the signal features are extracted using convolution kernels of different scales;

[0012] Step S3: After feature extraction is completed, the output of the encoder is sent to the context comparison module to enhance the feature representation capability, using the comparison of local and global information to emphasize key features while suppressing redundant information;

[0013] Step S4: After being enhanced by the context contrast module in step S3, the features are further sent to the channel-spatial attention module CBAM. CBAM includes channel attention CA and spatial attention SA, which are used to enhance the important information of the signal in the channel dimension and time dimension respectively;

[0014] Step S5: The feature map enhanced by the CBAM module is sent to the decoder, and the upsampled features are concatenated with the first-layer features of the encoder to generate a feature map containing detailed information of the input signal. The concatenated feature map is processed by a one-dimensional convolution operation to generate the final output signal.

[0015] Preferably, in step S1, the input signal is represented as Where T is the number of time steps of the signal and C is the number of channels of the signal.

[0016] Preferably, in step S2, the encoder adopts a multi-scale convolution design, which specifically includes three stages:

[0017] The first stage: Use a one-dimensional convolution kernel of size 1×25 to perform a convolution operation on the input signal to extract local detail features:

[0018] F1=σ(BN(W1*X+b1))

[0019] Where W1 is the weight of the first layer convolution kernel, * represents the convolution operation, b1 is the bias term, BN(·) represents batch normalization, and σ(·) is the ReLU activation function. The time dimension is compressed by maximum pooling, so that the time resolution of the feature map is reduced to 1 / 5 of the original. The pooling window size is set to 1×5 and the step size is set to 5.

[0020] The second stage: The output feature map F1 of the first stage is used as input. F1 is further convolved with a 1×15 one-dimensional convolution kernel to extract deeper features:

[0021] F2=σ(BN(W2*F1+b2))

[0022] The time dimension is further compressed to 1 / 10 of the original through maximum pooling again, the pooling window size is set to 1×2, and the step size is set to 2;

[0023] The third stage: The input feature F2 is passed through a 1×5 convolution kernel to extract deep features and global features:

[0024] F3=σ(BN(W3*F2+b3))

[0025] Max pooling is performed again, the temporal resolution is reduced to 1 / 20, the pooling window size is set to 1×2, the step size is set to 2, and finally the feature representation F3 is obtained for subsequent context modeling.

[0026] Preferably, in step S3, the context comparison module first calculates the comparison score between the local feature and the global feature, and then weights the features, which is specifically expressed as follows:

[0027] F CCB =α·F local +β·F global

[0028] Among them, F local is the local feature, F global is a global feature, and α and β are adaptively adjusted weight parameters.

[0029] Preferably, in step S4, in terms of channel attention, CBAM obtains the saliency information of a single channel by performing average pooling and maximum pooling operations on a single channel, and then reweights each channel through a fully connected layer to improve the feature expression capability of the signal, which is specifically expressed as follows:

[0030] M c (F)=σ(MLP(AvgPoll(F))+MLP(MaxPool(F)))

[0031] Among them, M c (F) is the channel attention weight, σ is the Sigmoid activation function, and MLP is the multi-layer perceptron;

[0032] In terms of spatial attention, CBAM combines the features of average pooling and maximum pooling to capture the important areas of the signal in the spatial dimension. Spatial attention helps the model focus on the most critical time-space areas of the signal while retaining the details of the signal. Specifically, it is expressed as:

[0033] M s (F')=σ(Conv2D([AvgPool(F');MaxPool(F')],w s ));

[0034] Among them, M s (F') is the spatial attention weight, w s is the convolution kernel weight.

[0035] Preferably, in step S5, the decoder is responsible for recovering the original structure of the signal from the features extracted by the encoder. The decoder gradually recovers the spatial resolution of the signal through upsampling operations, and combines the low-level features of the encoder with the high-level features through skip connections to retain the detailed information of the signal. The upsampling and skip connection formulas are:

[0036]

[0037] in is the input feature map of the decoder at stage i, The feature map after upsampling, Upsample(·) is the upsampling operation, s is the upsampling ratio, is the feature map of the i-th layer of the encoder, The feature map after upsampling the decoder Feature map corresponding to the encoder stage Splicing in the channel dimension;

[0038] In the first stage of the decoder, the input features come from the context-enhanced features of the bottleneck layer of the encoder. This feature includes high-level context information and is extracted from the third convolutional stage of the encoder. As part of a skip connection;

[0039] In the second stage of the decoder, the output features of the first stage are As input, it introduces the features F2∈R from the second convolutional stage of the encoder T / 10×16 Used for skip connections to fuse mid-level detail information;

[0040] In the third stage of the decoder, the output features of the second stage are As input, it introduces features from the first convolutional stage of the encoder As part of the skip connection.

[0041] Preferably, in the first stage of the decoder, the initial resolution is restored from the bottleneck layer features and combined with the high-level features of the encoder to enhance the global information, specifically including the following operations:

[0042] Bottleneck layer feature F CCB The temporal resolution is restored from T / 20 to T / 10 through upsampling. It is concatenated with the third-layer feature F3 of the encoder in the channel dimension to form a feature map that combines detail features and global information The concatenated feature maps are processed through one-dimensional convolution operations to extract important features and generate the output features of the first stage. Provide high-quality intermediate features for subsequent stages.

[0043] Preferably, in the second stage of the decoder, the resolution is further restored, and the expression ability of the context is improved by combining the middle-level features, which specifically includes the following steps: the output features of the first stage are The temporal resolution is restored from T / 10 to T / 5 through upsampling operation, and then the upsampled features Concatenated with the second-layer feature F2 of the encoder to form a feature map that integrates details and mid-level context information Finally, the concatenated feature map is processed using a one-dimensional convolution operation to generate the output features of the second stage. This feature further enhances the resolution and detail information.

[0044] Preferably, in the third stage of the decoder, the original resolution of the input signal is restored, while the detailed information of the signal is retained in combination with the low-level features, which specifically includes the following operations: Through the upsampling operation, the time resolution is restored from T / 5 to T consistent with the input signal. Then, the upsampled feature Concatenate with the first layer feature F1 of the encoder to generate a feature map containing detailed information of the input signal The concatenated feature map is processed through a one-dimensional convolution operation to generate the final output signal Its resolution is consistent with the input signal.

[0045] Therefore, the present invention adopts a method for denoising physiological signals of operators in multiple power scenarios using the above structure, which has the following beneficial effects:

[0046] (1) The present invention constructs a multi-physiological signal dataset adapted to power operation scenarios, which provides unified experimental data for studying and evaluating multi-signal and multi-noise noise reduction methods.

[0047] (2) The present invention designs a targeted improved U-Net model which can simultaneously remove multiple noises and retain the key features of the signal by optimizing the network structure and introducing a multi-scale feature extraction mechanism, thereby improving the noise reduction effect while ensuring the high fidelity of the signal.

[0048] (3) The improved model of the present invention has superior performance in complex and noisy environments, and is applied to health status monitoring and safety warning for power workers, providing technical support for improving the safety management of power operations.

[0049] (4) The present invention realizes dynamic weighting of key feature areas by introducing the channel-spatial attention mechanism (CBAM), significantly enhancing the model's ability to extract effective signals under complex noise backgrounds, thereby improving the signal-to-noise ratio (SNR) and root mean square error (RMSE) indicators. Secondly, combined with the context comparison module (CCB), the present invention realizes the comparison optimization of feature information under multiple noise interferences, enabling the model to effectively capture the difference between noise and target signals in power operation scenarios, and enhancing the removal effect of multiple noises.

[0050] (5) The multi-scale convolution structure designed by the present invention can simultaneously extract feature information at different scales, thereby taking into account the balance between global features and local details in complex dynamic environments, and effectively improving the robustness of the model to various types of noise. In addition, the present invention, in view of the particularity of power operation scenes, significantly enhances the generalization ability of the denoising method in multiple scenarios by comprehensively modeling and simulating the noise characteristics in different operating environments (such as high-altitude operations, emergency maintenance, etc.).

[0051] The method scheme of the present invention is further described in detail below through the drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 CBAM-UNet framework diagram in an embodiment of the present invention;

[0053] Figure 2 is a context comparison module in an embodiment of the present invention;

[0054] Figure 3 It is a flow chart of a method for denoising physiological signals of operators in multiple power scenarios according to the present invention;

[0055] Figure 4 It is a flow chart of the construction of physiological signals of electric personnel during operation in an embodiment of the present invention;

[0056] Figure 5 This is a comparison chart of EEG signal denoising performance under different models in the comparative example of the present invention. DETAILED DESCRIPTION

[0057] The method scheme of the present invention is further described below through drawings and embodiments.

[0058] Unless otherwise defined, method terms or scientific terms used in the present invention shall have the common meanings understood by one of ordinary skill in the art to which the present invention belongs.

[0059] The words "include" or "comprises" and the like used in the present invention mean that the elements before the word include the elements listed after the word, and do not exclude the possibility of also including other elements. The orientation or position relationship indicated by the terms "inside", "outside", "upper", "lower", etc. is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation of the present invention. When the absolute position of the described object changes, the relative position relationship may also change accordingly. In the present invention, unless otherwise clearly specified and limited, the terms "attachment" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral body; it can be directly connected, or indirectly connected through an intermediate medium, and it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0060] Example

[0061] like Figure 1-Figure 3 As shown, the present invention provides a method for denoising physiological signals of operators in multiple power scenarios. By introducing a channel-spatial attention module CBAM, a context comparison module CCB and a multi-scale convolution structure, the CCB module processes local features and global context information respectively by designing a coarse context branch and a fine context branch, so as to achieve accurate modeling of multi-scale noise. A denoising model CBAM-UNet based on an improved U-Net architecture is designed in combination with an improved U-Net architecture. The global features are first weighted by channel attention to improve the expression ability of important channel features and suppress the interference of irrelevant channels. Subsequently, the spatial attention captures the important spatial areas in the feature map and strengthens the features of the key areas by weighting. Specifically, the CBAM module can dynamically optimize multiple physiological signals in the power operation environment. For example, in the denoising process of ECG signals, the model can focus on the abnormal areas of power frequency interference and baseline drift, rather than global features. This selective processing greatly improves the denoising accuracy. At the same time, in the EDA signal processing process, the CBAM module can adaptively optimize the calculation path according to the signal distribution, significantly improving the real-time performance. This dynamic feature weighting mechanism enables the model to accurately process complex signal features.

[0062] The specific steps include:

[0063] Step S1: The input signal is normalized and filtered through the preprocessing module to reduce the impact of baseline drift and high-frequency noise on the model;

[0064] In step S1, the input signal is expressed as Where T is the number of time steps of the signal and C is the number of channels of the signal.

[0065] Step S2: The preprocessed signal enters the encoder part of CBAM-UNet, and the signal features are extracted using convolution kernels of different scales;

[0066] In step S2, the encoder adopts a multi-scale convolution design, which includes three stages:

[0067] The first stage: Use a one-dimensional convolution kernel of size 1×25 to perform a convolution operation on the input signal to extract local detail features:

[0068] F1=σ(BN(W1*X+b1))

[0069] Where W1 is the weight of the first layer convolution kernel, * represents the convolution operation, b1 is the bias term, BN(·) represents batch normalization, and σ(·) is the ReLU activation function. The time dimension is compressed by maximum pooling, so that the time resolution of the feature map is reduced to 1 / 5 of the original. The pooling window size is set to 1×5 and the step size is set to 5.

[0070] The second stage: The output feature map F1 of the first stage is used as input. F1 is further convolved with a 1×15 one-dimensional convolution kernel to extract deeper features:

[0071] F2=σ(BN(W2*F1+b2))

[0072] The time dimension is further compressed to 1 / 10 of the original through maximum pooling again, the pooling window size is set to 1×2, and the step size is set to 2;

[0073] The third stage: The input feature F2 is passed through a 1×5 convolution kernel to extract deep features and global features:

[0074] F3=σ(BN(W3*F2+b3))

[0075] Max pooling is performed again, the temporal resolution is reduced to 1 / 20, the pooling window size is set to 1×2, the step size is set to 2, and finally the feature representation F3 is obtained for subsequent context modeling.

[0076] Step S3: After feature extraction is completed, the output of the encoder is sent to the context comparison module to enhance the feature representation capability, using the comparison of local and global information to emphasize key features while suppressing redundant information;

[0077] In step S3, the context comparison module first calculates the comparison score between the local features and the global features, and then weights the features, which is specifically expressed as follows:

[0078] F CCB=α·F local +β·F global

[0079] Among them, F local is the local feature, F global is a global feature, and α and β are adaptively adjusted weight parameters.

[0080] Step S4: After being enhanced by the context contrast module in step S3, the features are further sent to the channel-spatial attention module CBAM. CBAM includes channel attention CA and spatial attention SA, which are used to enhance the important information of the signal in the channel dimension and time dimension respectively;

[0081] In step S4, in terms of channel attention, CBAM obtains the saliency information of a single channel by performing average pooling and maximum pooling operations on a single channel, and then reweights each channel through a fully connected layer to improve the feature expression ability of the signal, which is specifically expressed as:

[0082] M c (F)=σ(MLP(AvgPoll(F))+MLP(MaxPool(F)))

[0083] Among them, M c (F) is the channel attention weight, σ is the Sigmoid activation function, and MLP is the multi-layer perceptron;

[0084] In terms of spatial attention, CBAM combines the features of average pooling and maximum pooling to capture the important areas of the signal in the spatial dimension. Spatial attention helps the model focus on the most critical time-space areas of the signal while retaining the details of the signal. Specifically, it is expressed as:

[0085] M s (F')=σ(Conv2D([AvgPool(F');MaxPool(F')],w s ));

[0086] Among them, M s (F') is the spatial attention weight, w s is the convolution kernel weight.

[0087] Step S5: The feature map enhanced by the CBAM module is sent to the decoder, and the upsampled features are concatenated with the first-layer features of the encoder to generate a feature map containing the detailed information of the input signal. The concatenated feature map is processed by a one-dimensional convolution operation to generate the final output signal. In step S5, the decoder is responsible for recovering the original structure of the signal from the features extracted by the encoder. The decoder gradually restores the spatial resolution of the signal through upsampling operations, and combines the low-level features of the encoder with the high-level features through jump connections to retain the detailed information of the signal. The upsampling and jump connection formulas are:

[0088]

[0089] in is the input feature map of the decoder at stage i, The feature map after upsampling, Upsample(·) is the upsampling operation, s is the upsampling ratio, is the feature map of the i-th layer of the encoder, The feature map after upsampling the decoder Feature map corresponding to the encoder stage Splicing in the channel dimension;

[0090] In the first stage of the decoder, the input features come from the context-enhanced features of the bottleneck layer of the encoder. This feature includes high-level context information and is extracted from the third convolutional stage of the encoder. As part of a skip connection;

[0091] In the second stage of the decoder, the output features of the first stage are As input, it introduces the features F2∈R from the second convolutional stage of the encoder T / 10×16 Used for skip connections to fuse mid-level detail information;

[0092] In the third stage of the decoder, the output features of the second stage are As input, it introduces features from the first convolutional stage of the encoder As part of the skip connection.

[0093] In the first stage of the decoder, the initial resolution is restored from the bottleneck layer features and combined with the high-level features of the encoder to enhance the global information. The specific operations include the following:

[0094] Bottleneck layer feature F CCB The temporal resolution is restored from T / 20 to T / 10 through upsampling. It is concatenated with the third-layer feature F3 of the encoder in the channel dimension to form a feature map that combines detail features and global information The concatenated feature maps are processed through one-dimensional convolution operations to extract important features and generate the output features of the first stage. Provide high-quality intermediate features for subsequent stages.

[0095] In the second stage of the decoder, the resolution is further restored and the mid-level features are combined to improve the expression of the context. The following steps are specifically included: The temporal resolution is restored from T / 10 to T / 5 through upsampling operation, and then the upsampled features Concatenated with the second-layer feature F2 of the encoder to form a feature map that integrates details and mid-level context information Finally, the concatenated feature map is processed using a one-dimensional convolution operation to generate the output features of the second stage. This feature further enhances the resolution and detail information.

[0096] In the third stage of the decoder, the original resolution of the input signal is restored, while the signal details are retained by combining the low-level features. The specific operations include: Through the upsampling operation, the time resolution is restored from T / 5 to T consistent with the input signal. Then, the upsampled feature Concatenate with the first layer feature F1 of the encoder to generate a feature map containing detailed information of the input signal The concatenated feature map is processed through a one-dimensional convolution operation to generate the final output signal Its resolution is consistent with the input signal. The goal of denoising is to make the output signal To approximate the noise-free signal, the loss function uses the mean square error (MSE). The loss function is optimized through back propagation, so that the model can effectively remove noise, improve the signal-to-noise ratio (SNR) and reduce the root mean square error (RMSE). The entire denoising process covers signal input, multi-scale feature extraction, attention enhancement, decoding recovery and loss optimization, ensuring that the model can still efficiently extract noise-free physiological signals in complex power operation environments.

[0097] In order to study the denoising problem of physiological signals of power workers in multiple scenarios that are disturbed by complex environments, a multimodal physiological signal dataset covering electroencephalogram (EEG), electrocardiogram (ECG), electrodermal activity (EDA) and photoplethysmography (PPG) signals was constructed based on the public FatigueSet, MIT-BIH Arrhythmia Database and PPG-DaLia datasets. The FatigueSet dataset contains EEG records related to mental fatigue, which is suitable for simulating EEG changes in long-term high-intensity work; the MIT-BIH dataset provides a variety of arrhythmia ECG signals, which can be used to reflect the heart rhythm changes of power workers under high load; the PPG-DaLia dataset contains multimodal data related to heart rate estimation, which supports the study of EDA and PPG signals to simulate psychological stress and dynamic changes of pulse waves. In the original signal processing, this paper standardized various types of data and unified the format into a four-dimensional tensor form to adapt to the input requirements of subsequent deep learning models.

[0098] It mainly targets three typical scenarios: ground operations of power workers (Group1), aerial operations of power workers (Group2), and maintenance operations of power workers (Group3). These three scenarios correspond to different main noise characteristics: Group1 uses power frequency interference generated by equipment operation as the main noise source; Group2 is significantly affected by electromyographic noise due to the intense movement during aerial operations; Group3 is mainly affected by the combined effect of baseline drift and power frequency interference. These noise characteristics fully reflect the complex and changeable interference patterns in the power operation environment. In response to the typical noise in power operation scenarios, this solution designs three simulation noise models, such as Figure 4 shown.

[0099] In the process of data set construction, this scheme first superimposes the simulated noise on the original physiological signal at a set ratio according to the noise model to generate signal data with scene-specific noise. Subsequently, the data is divided into training set, validation set and test set with a ratio of 60%, 20% and 20% respectively. The final generated data set covers the complex noise characteristics related to multimodal physiological signals and scenes, providing a rich experimental basis for subsequent algorithm research.

[0100] To verify the quality of the data set, this solution uses the signal-to-noise ratio (SNR) and root mean square error (RMSE) to quantitatively evaluate the generated signal, ensuring the authenticity and reliability of the data from the two aspects of noise intensity and signal discreteness. The evaluation found that the generated data set has a high degree of consistency in terms of the degree of agreement between the noise characteristics and the actual operation scene.

[0101] Experimental environment

[0102] The experimental configuration of this study uses a reliable computing environment and runs on a computer with Ubuntu 18.04.6 operating system installed. Python 3.8.10 is used as the development language in the experiment, combined with the PyTorch 1.13.0 deep learning framework to complete the algorithm development and model training. The computing resources include an NVIDIA Tesla T4 GPU, which is supported by CUDA 11.6 technology and cuDNN 8.3 deep learning acceleration library to ensure the efficiency and stability of model training and reasoning.

[0103] During the model training process, the dataset is divided into training set, validation set and test set, with a ratio of 70%, 15% and 15%. The training set is used to train the model, the validation set is used to monitor the model performance during the training process and adjust the hyperparameters when necessary; the test set is used to finally evaluate the denoising effect of the model.

[0104] During the training process, the cross-validation method was used, and the hyperparameters of the model were adjusted according to the performance on the validation set. In order to ensure the stability of the training, the Early Stopping strategy was used to prevent the model from overfitting during the training process.

[0105] Evaluation indicators

[0106] The performance of the denoising model is measured by two evaluation indicators, including signal-to-noise ratio (SNR) and root mean square error (RMSE). Signal-to-noise ratio (SNR): used to measure the ratio of the clarity of the denoised signal to the noise. A higher SNR value indicates a better denoising effect. Root mean square error (RMSE): used to measure the error between the denoised signal and the true signal. A lower RMSE value indicates that the denoising effect of the model is more accurate. The calculation formula is as follows:

[0107]

[0108] Where xg(n) is the original signal, the nth sampling point, xr(n) is the denoised signal, and N is the total number of samples of the signal.

[0109] Based on the multimodal physiological signal data construction method proposed in this scheme, a high-quality dataset covering three typical power operation scenarios was constructed to simulate the characteristics and interference of physiological signals in real working environments. This method simulates typical noise characteristics such as power frequency interference, myoelectric noise and baseline drift in power operation scenarios by integrating the public MIT-BIH Arrhythmia Database, PPG-BP and FatigueSet datasets, and combining specific noise generation strategies. The final generated dataset includes four physiological signals: electrocardiogram (ECG), photoplethysmography (PPG), electrodermal activity (EDA) and electroencephalogram (EEG), and the average signal-to-noise ratio (SNR) and root mean square error (RMSE) of its test set are statistically analyzed, as shown in Table 1.

[0110] Table 1

[0111]

[0112] In the ground operation scenario, the data set contains 2157 signals, of which normal signals account for 51% and abnormal signals account for 49%. Abnormal signals simulate the impact of equipment noise and control pressure on signals by introducing arrhythmia and abnormal blood pressure characteristics. In the high-altitude operation scenario, the data set contains 1567 signals, of which normal signals and abnormal signals each account for 50%. This scenario focuses on simulating the interference of electromyographic noise on the signal, and abnormal signals cover physiological changes under fatigue and high-intensity physical activity. In the maintenance operation scenario, the data set contains 1247 signals, of which abnormal signals account for 64%, mainly simulating the impact of baseline drift and sudden environmental interference on the signal.

[0113] The final constructed dataset contains a total of 4967 multimodal physiological signals, of which 1990 are normal signals, accounting for 40% of the total; and 2977 are abnormal signals, accounting for 60% of the total. Abnormal signals are further subdivided into arrhythmia, abnormal blood pressure, and fatigue-related signals, which fully cover the potential interference factors and physiological state characteristics in power operation scenarios. Based on the data construction method proposed in this paper, the generated dataset can not only realistically simulate the multi-scenario noise characteristics in the power operation environment, but also provide a high-quality experimental basis and reference standard for the evaluation of physiological signal denoising algorithms, the development of anomaly detection models, and health status monitoring.

[0114] Comparative Example

[0115] The data sources of ECG, PPG, EDA, and EEG and the noise convolution construction method are as follows: Group 1 is the MA-dominated data set. For the noise convolution ECG, PPG, EDA, and EEG in the test set, the average SNRs are 14.81dB, 28.89dB, 20.71dB, and 19.18dB, and the average RMSEs are 8.48×10-2, 6.17×10-2, 9.71×10-2, and 5.94×10-2, respectively. Group 2 is the EM-dominated data set. For the noise convolution ECG, PPG, EDA, and EEG in the test set, the average SNRs are 13.86dB, 21.38dB, 8.36dB, and 17.91dB, and the average RMSEs are 9.38×10-2, 6.63×10-2, 52.19×10-2, and 6.78×10-2, respectively. Group 3 is the BW-dominated dataset. For the noisy convolution ECG, PPG, EDA, and EEG in the test set, the average SNRs are 6.62dB, 21.39dB, 19.02dB, and 13.79dB, respectively, and the average RMSEs are 21.36×10-2, 6.62×10-2, 19.21×10-2, and 11.73×10-2, respectively.

[0116] In order to verify the denoising performance of the multi-physiological signal denoising method of power workers based on CBAM-Unet proposed in this scheme in ground operations, we compared the signal denoising performance of the signal before denoising, the traditional U-Net, the signal without CCB module, the signal without CBAM module and the CBAM-Unet model in this paper in different operation scenarios. The specific indicators include signal-to-noise ratio (SNR) and root mean square error (RMSE). The experimental results are shown in Table 2.

[0117] Table 2

[0118]

[0119] Experimental data analysis

[0120] For Group 1 data, the improved U-Net model showed significant improvements in the signal-to-noise ratio (SNR) and root mean square error (RMSE) of all physiological signal types (ECG, PPG, EDA, EEG). Especially for PPG signals, the SNR increased from the original 28.89 to 38.72, while the RMSE decreased significantly from 0.061701 to 0.012456, showing the significant advantages of the model for PPG signal processing.

[0121] In order to verify the denoising performance of the multi-physiological signal denoising method of power workers based on CBAM-Unet proposed in this scheme in high-altitude operations, we compared the signal denoising performance of the pre-denoising signal, the traditional U-Net, the non-CCB module, the non-CBAM module and the CBAM-Unet model in this paper in different operation scenarios. The specific indicators include signal-to-noise ratio (SNR) and root mean square error (RMSE). The experimental results are shown in Table 3.

[0122] Table 3

[0123]

[0124] Experimental data analysis

[0125] In Group 2, the denoising effect of the EDA signal is particularly outstanding, with the SNR increasing from the original 8.36 to 20.72, and the RMSE decreasing from 0.521972 to 0.187156, reflecting that the improved U-Net model can effectively improve the signal quality in the complex noise environment with high EDA signals. In addition, the SNR of the PPG signal is also significantly improved, from 21.38 to 38.72.

[0126] In order to verify the denoising performance of the multi-physiological signal denoising method of power workers based on CBAM-Unet proposed in this scheme in maintenance operations, we compared the signal denoising performance of the signal before denoising, the traditional U-Net, the signal without CCB module, the signal without CBAM module and the CBAM-Unet model in this paper in different operation scenarios. The specific indicators include signal-to-noise ratio (SNR) and root mean square error (RMSE). The experimental results are shown in Table 4.

[0127] Table 4

[0128]

[0129]

[0130] Experimental results data analysis

[0131] The results of Group 3 further verified the applicability and effectiveness of the improved U-Net model in denoising a variety of physiological signals. In particular, for ECG signals, although starting from a lower baseline SNR (6.62), the improved model can increase the SNR to 17.91 while reducing the RMSE. PPG and EDA signals also showed similar significant improvements.

[0132] In order to verify the denoising effect of CBAM-Unet on physiological signals, we analyzed the changes in SNR and RMSE of physiological signals in different scenarios. The results are as follows: Figure 5 shown.

[0133] Through comparative analysis of the three sets of data, it can be seen that the improved U-Net model has a general improvement in denoising effect, especially for the two signals of PPG and EDA. The model not only significantly improves the SNR, but also achieves a good reduction in RMSE, indicating its wide adaptability and efficiency in processing different physiological signals. This performance improvement is attributed to the contextual comparison block and channel-spatial attention module introduced in the model. These designs help the model locate and suppress noise more accurately while retaining the important biological characteristics of the signal. These results provide strong evidence and reference for the future application of deep learning models in a wider range of biomedical signal processing fields.

[0134] This solution proposes a set of efficient denoising technology solutions for the complex characteristics of multi-noise superposition and high dynamic interference in the power operation environment. Based on the improved U-Net architecture, combined with the channel-spatial attention mechanism (CBAM), contextual contrast module (CCB) and multi-scale convolution structure, the denoising ability of various physiological signals (such as ECG, EEG, PPG, EDA) is significantly improved.

[0135] Therefore, the present invention adopts the above-mentioned method for denoising the physiological signals of operators in multiple power scenarios, and forms a synergistic mechanism of multi-scale feature extraction and dynamic feature focusing through the design and combination of the context comparison module CCB and the CBAM module, thereby achieving accurate processing for different noise types. Specifically, the CCB module provides stronger feature modeling capabilities through multi-scale context fusion at the bottleneck layer; the CBAM module uses channel and spatial attention to dynamically weight features, optimizing the model's processing capabilities for key signal areas.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the method scheme of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, ordinary method personnel in the field should understand that they can still modify or replace the method scheme of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified method scheme to deviate from the spirit and scope of the method scheme of the present invention.

Claims

1. A method for denoising physiological signals of operators in multiple power scenarios, characterized by: By introducing the channel-spatial attention module CBAM, the contextual contrast module CCB and the multi-scale convolution structure, a denoising model CBAM-UNet based on the improved U-Net architecture is designed in combination with the improved U-Net architecture, which specifically includes the following steps: Step S1: The input signal is normalized and filtered through the preprocessing module to reduce the impact of baseline drift and high-frequency noise on the model; Step S2: The preprocessed signal enters the encoder part of CBAM-UNet, and the signal features are extracted using convolution kernels of different scales; Step S3: After feature extraction is completed, the output of the encoder is sent to the context comparison module to enhance the feature representation capability, using the comparison of local and global information to emphasize key features while suppressing redundant information; Step S4: After being enhanced by the context contrast module in step S3, the features are further sent to the channel-spatial attention module CBAM. CBAM includes channel attention CA and spatial attention SA, which are used to enhance the important information of the signal in the channel dimension and time dimension respectively; Step S5: The feature map enhanced by the CBAM module is sent to the decoder, and the upsampled features are concatenated with the first-layer features of the encoder to generate a feature map containing detailed information of the input signal. The concatenated feature map is processed by a one-dimensional convolution operation to generate the final output signal.

2. A method for denoising physiological signals of operators in multiple power scenarios according to claim 1, characterized in that: In step S1, the input signal is expressed as Where T is the number of time steps of the signal and C is the number of channels of the signal.

3. A method for denoising physiological signals of operators in multiple power scenarios according to claim 2, characterized in that: In step S2, the encoder adopts a multi-scale convolution design, which includes three stages: The first stage: Use a one-dimensional convolution kernel of size 1×25 to perform a convolution operation on the input signal to extract local detail features: F1=σ(BN(W1*X+b1)) Where W1 is the weight of the first layer convolution kernel, * represents the convolution operation, b1 is the bias term, BN(·) represents batch normalization, and σ(·) is the ReLU activation function. The time dimension is compressed by maximum pooling, so that the time resolution of the feature map is reduced to 1 / 5 of the original. The pooling window size is set to 1×5 and the step size is set to 5. The second stage: The output feature map F1 of the first stage is used as input. F1 is further convolved with a 1×15 one-dimensional convolution kernel to extract deeper features: F2=σ(BN(W2*F1+b2)) The time dimension is further compressed to 1 / 10 of the original through maximum pooling again, the pooling window size is set to 1×2, and the step size is set to 2; The third stage: The input feature F2 is passed through a 1×5 convolution kernel to extract deep features and global features: F3=σ(BN(W3*F2+b3)) Max pooling is performed again, the temporal resolution is reduced to 1 / 20, the pooling window size is set to 1×2, the step size is set to 2, and finally the feature representation F3 is obtained for subsequent context modeling.

4. A method for denoising physiological signals of operators in multiple power scenarios according to claim 3, characterized in that: In step S3, the context comparison module first calculates the comparison score between the local features and the global features, and then weights the features, which is specifically expressed as follows: F CCB =α·F local +β·F global Among them, F local is the local feature, F global is a global feature, and α and β are adaptively adjusted weight parameters.

5. A method for denoising physiological signals of operators in multiple power scenarios according to claim 4, characterized in that: In step S4, in terms of channel attention, CBAM obtains the saliency information of a single channel by performing average pooling and maximum pooling operations on a single channel, and then reweights each channel through a fully connected layer to improve the feature expression ability of the signal, which is specifically expressed as: M c (F)=σ(MLP(AvgPoll(F))+MLP(MaxPool(F))) Among them, M c (F) is the channel attention weight, σ is the Sigmoid activation function, and MLP is the multi-layer perceptron; In terms of spatial attention, CBAM combines the features of average pooling and maximum pooling to capture the important areas of the signal in the spatial dimension. Spatial attention helps the model focus on the most critical time-space areas of the signal while retaining the details of the signal. Specifically, it is expressed as: M s (F')=σ(Conv2D([AvgPool(F');MaxPool(F')],w s )); Among them, M s (F') is the spatial attention weight, w s is the convolution kernel weight.

6. A method for denoising physiological signals of operators in multiple power scenarios according to claim 5, characterized in that: In step S5, the decoder is responsible for recovering the original structure of the signal from the features extracted by the encoder. The decoder gradually recovers the spatial resolution of the signal through upsampling operations, and combines the low-level features of the encoder with the high-level features through skip connections to retain the detailed information of the signal. The upsampling and skip connection formulas are: in is the input feature map of the decoder at stage i, The feature map after upsampling, Upsample(·) is the upsampling operation, s is the upsampling ratio, is the feature map of the i-th layer of the encoder, The feature map after the decoder is upsampled Feature map corresponding to the encoder stage Splicing in the channel dimension; In the first stage of the decoder, the input features come from the context-enhanced features of the bottleneck layer of the encoder. This feature includes high-level context information and is extracted from the third convolutional stage of the encoder. As part of a skip connection; In the second stage of the decoder, the output features of the first stage are As input, it introduces the features F2∈R from the second convolutional stage of the encoder T / 10×16 Used for skip connections to fuse mid-level detail information; In the third stage of the decoder, the output features of the second stage are As input, it introduces features from the first convolutional stage of the encoder As part of the skip connection.

7. A method for denoising physiological signals of operators in multiple power scenarios according to claim 6, characterized in that: In the first stage of the decoder, the initial resolution is restored from the bottleneck layer features and combined with the high-level features of the encoder to enhance the global information. The specific operations include the following: Bottleneck layer feature F CCB The temporal resolution is restored from T / 20 to T / 10 through upsampling. It is concatenated with the third-layer feature F3 of the encoder in the channel dimension to form a feature map that combines detail features and global information The concatenated feature maps are processed through one-dimensional convolution operations to extract important features and generate the output features of the first stage. Provide high-quality intermediate features for subsequent stages.

8. A method for denoising physiological signals of operators in multiple power scenarios according to claim 7, characterized in that: In the second stage of the decoder, the resolution is further restored and the mid-level features are combined to improve the expression of the context. The following steps are specifically included: The temporal resolution is restored from T / 10 to T / 5 through upsampling operation, and then the upsampled features Concatenated with the second-layer feature F2 of the encoder to form a feature map that integrates details and mid-level context information Finally, the concatenated feature map is processed using a one-dimensional convolution operation to generate the output features of the second stage. This feature further enhances the resolution and detail information.

9. A method for denoising physiological signals of operators in multiple power scenarios according to claim 8, characterized in that: In the third stage of the decoder, the original resolution of the input signal is restored, while the signal details are retained by combining the low-level features. The specific operations include: Through the upsampling operation, the time resolution is restored from T / 5 to T consistent with the input signal. Then, the upsampled feature Concatenate with the first layer feature F1 of the encoder to generate a feature map containing detailed information of the input signal The concatenated feature map is processed through a one-dimensional convolution operation to generate the final output signal Its resolution is consistent with the input signal.

Images