PPG signal adaptive noise reduction method based on wavelet transform and auto-encoder

Abstract

The invention relates to the technical field of signal adaptive noise reduction, and provides a PPG signal adaptive noise reduction method based on wavelet transform and an auto-encoder. The method combines adaptive variational mode decomposition, stationary wavelet denoising, multi-dimensional consistency optimization, a synchronous updating technology and a Lagrangian multiplier constraint adjustment technology to remove high-frequency noise of PPG signals, and can better adapt to individual differences of drivers and complexity of the signals; the total loss function of the auto-encoder is optimized by adopting contrast loss, weighted reconstruction error and sparsity regularization, so that the sensitivity and the judgment capability of an abnormal health state are enhanced; according to the method, the PPG signals can be accurately restored and evaluated in multiple dimensions, reliable technical support is provided for continuous monitoring of the health state of the driver, high stability and expandability are achieved, and important guarantee can be provided for safe driving of vehicles and health management of the driver.

Description

Technical Field

[0001] The present invention relates to the technical field of signal adaptive noise reduction, and in particular to a PPG signal adaptive noise reduction method based on wavelet transform and autoencoder. Background Art

[0002] The adaptive noise reduction method of PPG signal (photoplethysmography signal, PPG signal is a biomedical signal that uses optical principles to detect changes in blood volume under the skin) based on wavelet transform and autoencoder is a technology that combines wavelet transform with deep learning autoencoder to process noise in PPG signals. Wavelet transform decomposes the signal in the time and frequency domains and can effectively separate the noise components. PPG is a non-invasive biosignal measurement technology that uses optical methods to detect changes in human peripheral blood volume. PPG signals reflect the changes in blood volume in human blood vessels as the heart pumps blood. PPG signals usually contain a DC component and an AC component, and are widely used in health monitoring, including heart rate monitoring, blood oxygen saturation measurement, and blood pressure estimation. In driver health monitoring, PPG signals are an important basis for evaluating the driver's physiological state. The signal quality of PPG signals is crucial to the accuracy of driver health status assessment. However, due to the complexity of signal noise in dynamic driving environments, traditional wavelet transform denoising methods are difficult to cope with the multidimensional characteristics of PPG signals. Traditional wavelet transform denoising methods cannot adaptively balance signal and noise components, resulting in signal distortion and poor noise reduction effects, and cannot provide high-quality data support for driver health status assessment. At the same time, traditional autoencoders are difficult to fully capture the deep features in multidimensional physiological signals, thereby affecting the accuracy of anomaly detection. Therefore, there is an urgent need for a technical solution that can adaptively remove noise from multidimensional PPG signals and accurately evaluate the driver's health status, providing strong support for driver health management and early warning. Summary of the invention

[0003] The present invention realizes denoising of multidimensional PPG signals by constructing MAVMD-SWT denoising algorithm (adaptive variational mode decomposition and stationary wavelet transform denoising algorithm). The MAVMD-SWT denoising algorithm first uses adaptive variational mode decomposition to preliminarily decompose the multidimensional PPG signal, and then uses synchronous update technology and Lagrange multiplier constraint adjustment technology to ensure the balance and consistency of noise and signal components in multidimensional space. Finally, denoising is performed through stationary wavelet transform to provide high-quality data support for subsequent health status assessment. In order to further improve the data quality after signal denoising, an enhanced sparse autoencoder is designed, which optimizes the total loss function of the autoencoder by combining contrast loss, weighted reconstruction error, weight attenuation term and sparsity regularization term to achieve deep feature learning of denoised data and classification of health status, so as to accurately judge the health status of the driver. The invention not only improves the validity of data, but also provides reliable safety protection in the field of driving safety and health monitoring.

[0004] The present invention provides a PPG signal adaptive denoising method based on wavelet transform and autoencoder, the method comprising the following steps: Step T1: Data collection: Collect the driver's heart rate data, heart rate variability data, blood oxygen saturation data and pulse waveform data, and fuse them to obtain multi-dimensional comprehensive health data; Step T2: Data preprocessing: performing signal synchronization, DC offset removal, filtering and standardization on the multi-dimensional comprehensive health data to obtain pre-processed multi-dimensional comprehensive health data; Step T3: Data denoising: Combining adaptive variational mode decomposition, stationary wavelet denoising, multidimensional consistency optimization, synchronous update technology and Lagrange multiplier constraint adjustment technology, the MAVMD-SWT denoising algorithm is obtained, and the pre-processed multidimensional comprehensive health data is denoised by the MAVMD-SWT denoising algorithm to obtain denoised multidimensional comprehensive health data; the MAVMD-SWT denoising algorithm specifically includes a decomposition layer initialization unit, a multidimensional VMD decomposition unit (multidimensional variational mode decomposition unit), a key value calculation unit, a key value judgment unit, a multidimensional stationary wavelet denoising unit and a multidimensional signal reconstruction unit; Step T4: Anomaly detection: Combine contrast loss, weighted reconstruction error, weight decay term and sparsity regularization term to optimize the total loss function of the autoencoder to obtain an enhanced sparse autoencoder. Input the denoised multi-dimensional comprehensive health data into the enhanced sparse autoencoder to perform driver anomaly detection and judge the driver's health status.

[0005] Furthermore, in step T3, the decomposition layer initialization unit executes the steps including setting the initial VMD decomposition layer to 2.

[0006] Furthermore, in step T3, the multidimensional VMD decomposition unit executes the steps including performing VMD decomposition on the preprocessed multidimensional comprehensive health data through multiple channels, and decomposing the preprocessed multidimensional comprehensive health data into multidimensional IMF components (multidimensional intrinsic mode function components) according to the number of initial VMD decomposition layers.

[0007] Furthermore, in step T3, the key value calculation unit executes the steps of respectively calculating the kurtosis and correlation coefficient of each dimension of the multidimensional IMF component, and taking the product of the kurtosis and the correlation coefficient as the key value.

[0008] Further, in step T3, the key value judgment unit executes the following steps specifically: checking the minimum key value of the IMF component of each dimension in the multidimensional IMF component; if the minimum key value is ≤1, marking the IMF component as a high-noise component and entering stationary wavelet denoising; if the minimum key value is >1, increasing the number of decomposition layers and re-performing multidimensional VMD decomposition until a multidimensional high-noise IMF component and a multidimensional noise-free IMF component are marked; obtaining a multidimensional consistency optimization algorithm through synchronous update technology and Lagrange multiplier constraint adjustment technology; and using the multidimensional consistency optimization algorithm to further select and synchronize the multidimensional high-noise IMF components to obtain optimized multidimensional high-noise IMF components.

[0009] Furthermore, in step T3, the multidimensional stationary wavelet denoising unit performs the following steps: selecting a wavelet basis for each dimension of the optimized multidimensional high-noise IMF component, applying stationary wavelet transform to remove high-frequency noise, and obtaining a denoised multidimensional IMF component.

[0010] Furthermore, in step T3, the steps performed by the multidimensional signal reconstruction unit specifically include recombining the denoised multidimensional IMF components and the multidimensional noise-free IMF components, performing multidimensional reconstruction, and forming denoised multidimensional comprehensive health data.

[0011] Further, in the key value judgment unit, the multidimensional high-noise IMF component is further selected and synchronized using the multidimensional consistency optimization algorithm to obtain a process of optimizing the multidimensional high-noise IMF component, which specifically includes the following steps: Step S1: Construct objective function: Construct a multi-dimensional joint optimization objective function. The formula used is as follows: ; in, represents the IMF component index, Indicates IMF components, Indicates The center frequency of the IMF component; represents the multi-dimensional joint optimization objective function, Represents the total number of dimensions, represents the dimension index, Indicates The weight coefficient of the dimension, represents the total number of IMF components, Indicates time The derivative of represents the imaginary unit, represents a complex exponential function with a center frequency, represents the square of the norm; Step S2: IMF component update: Based on the multi-dimensional joint optimization objective function, the multi-dimensional high-noise IMF components are synchronously updated to gradually reduce the noise in each dimension. The formula used is as follows: ; in, and Both represent IMF component indexes. represents a frequency variable, The index representing the number of iterations, Indicates The IMF component is The spectrum update value in the iteration, represents the spectrum of the original signal, Indicates The representation of the IMF components in the frequency domain is: Indicates that except The sum of the spectra of the other IMF components except the IMF components, represents the Lagrange multiplier, represents the smoothing parameter; Step S3: Lagrange multiplier update: After each synchronization update, Lagrange multipliers are introduced to perform constraint adjustment to ensure the frequency consistency of the multidimensional high-noise IMF component, iteratively update the parameters and Lagrange multipliers of the multidimensional high-noise IMF component, and finally output the optimized multidimensional high-noise IMF component. The formula used is as follows: ; in, The index representing the number of iterations, Indicates The updated value of the Lagrange multiplier at the iteration, Indicates The Lagrange multiplier of the iteration, represents the update step coefficient, Indicates The IMF component is The updated value at the iteration, represents the sum of the spectra of the updated IMF components.

[0012] Furthermore, in step T4, the process of inputting the denoised multi-dimensional comprehensive health data into the enhanced sparse autoencoder to perform driver abnormality detection and determine the driver's health status specifically includes the following steps: Step T41: Data input: The enhanced sparse autoencoder includes an encoding layer, a hidden layer, and a decoding layer. The weights of the enhanced sparse autoencoder are initialized, and the denoised multi-dimensional comprehensive health data is input into the encoding layer; Step T42: forward propagation encoding: convert the denoised multi-dimensional comprehensive health data into a low-dimensional sparse representation through the encoding layer, and set the sparsity parameters to control the activity of the hidden layer neurons; Step T43: Calculate the sparsity loss: Calculate the sparsity regularization term through KL divergence to measure the deviation between the actual average activation value of the hidden layer neurons and the sparsity parameter. The formula used is as follows: ; in, represents the sparsity regularization term, represents the neuron index of the hidden layer, represents the total number of neurons in the hidden layer, represents the summation symbol, represents the sparsity parameter, Indicates The actual average activation value of the hidden layer neurons, represents the KL divergence, The first part of the KL divergence formula measures The deviation between the actual average activation value of the hidden layer neurons and the sparsity parameter in the activated state, Represents the second part of the KL divergence formula, measuring the The inactive state bias of hidden layer neurons; Step T44: Back propagation optimization: Introduce contrast loss, weighted reconstruction error and weight decay terms, and combine the sparsity regularization term to construct the total loss function. Minimize the total loss function through back propagation and adjust the weights of the enhanced sparse autoencoder. The formula used is as follows: ; in, represents the index of the sample, represents the total number of samples, represents the index of the feature, represents the total number of features, Express The sum of samples, Indicates that for each sample The sum of feature dimensions, Indicates The weight of the feature, Indicates The sample The original input value of the feature, Indicates The sample The reconstructed input value of features, express The sample The squared error on the feature dimension; represents the weight of the sparsity regularization term, Represents the control coefficient of the weight decay term; and represents the neuron index, Indicates that from Neuron to The connection weights of neurons, represents the square of weight; represents the weight coefficient of contrast loss, represents contrast loss; Step T45: Health status assessment: According to the weights of the enhanced sparse autoencoder, the low-dimensional sparse representation is restored through the decoding layer to obtain reconstructed data, the reconstruction error is calculated based on the reconstructed data, and the driver's health status is judged based on the reconstruction error.

[0013] By adopting the above scheme, the beneficial effects achieved by the present invention are as follows: The present invention realizes adaptive noise reduction processing of multidimensional PPG signals by constructing the MAVMD-SWT denoising algorithm, and effectively improves the quality and reliability of driver health monitoring data; the algorithm first uses the adaptive variational mode decomposition technology to preliminarily decompose the multidimensional signal, and then uses the synchronous update and Lagrange multiplier constraint adjustment technology to ensure the balance and consistency of each signal component in the multidimensional space, and finally combines the stationary wavelet transform to further remove the noise; the adaptive denoising process overcomes the limitation that traditional filtering and fixed mode decomposition methods are difficult to cope with the complex noise in the dynamic driving environment, significantly improves the accuracy of signal denoising, provides high-quality input data support for subsequent health status assessment, and ensures that clear and reliable physiological data can be obtained in different environments; At the same time, the present invention further improves the accuracy of feature learning and health status assessment of denoised data by designing an enhanced sparse autoencoder; the enhanced sparse autoencoder optimizes the loss function by combining contrast loss, weighted reconstruction error, weight attenuation term and sparsity regularization term, effectively extracts the deep features of PPG signal data, and realizes the classification and anomaly detection of driver health status on this basis; compared with traditional autoencoders, the enhanced sparse autoencoder of the present invention can accurately identify fatigue, hypoxia and abnormal heart rate, effectively enhances the abnormal warning capability of the health monitoring system, and improves the reliability and accuracy of health assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A schematic diagram of a flow chart of a PPG signal adaptive denoising method based on wavelet transform and autoencoder proposed in the present invention; Figure 2 It is a flow chart of the multi-dimensional consistency optimization algorithm in the key value judgment unit proposed by the present invention; Figure 3 This is a schematic diagram of the process of the enhanced sparse autoencoder in step T4 proposed in the present invention. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0016] Embodiment 1, according to Figure 1 The present invention provides a PPG signal adaptive denoising method based on wavelet transform and autoencoder, the method comprising the following steps: Step T1: Data collection: Collect the driver's heart rate data, heart rate variability data, blood oxygen saturation data and pulse waveform data, and fuse them to obtain multi-dimensional comprehensive health data; Step T2: Data preprocessing: performing signal synchronization, DC offset removal, filtering and standardization on the multi-dimensional comprehensive health data to obtain pre-processed multi-dimensional comprehensive health data; Step T3: Data denoising: Combining adaptive variational mode decomposition, stationary wavelet denoising, multidimensional consistency optimization, synchronous update technology and Lagrange multiplier constraint adjustment technology, the MAVMD-SWT denoising algorithm is obtained, and the pre-processed multidimensional comprehensive health data is denoised by the MAVMD-SWT denoising algorithm to obtain denoised multidimensional comprehensive health data; the MAVMD-SWT denoising algorithm specifically includes a decomposition layer initialization unit, a multidimensional VMD decomposition unit, a key value calculation unit, a key value judgment unit, a multidimensional stationary wavelet denoising unit and a multidimensional signal reconstruction unit; Step T4: Anomaly detection: Combine contrast loss, weighted reconstruction error, weight decay term and sparsity regularization term to optimize the total loss function of the autoencoder to obtain an enhanced sparse autoencoder. Input the denoised multi-dimensional comprehensive health data into the enhanced sparse autoencoder to perform driver anomaly detection and judge the driver's health status.

[0017] Embodiment 2: This embodiment is based on the above embodiment. In step T3, the decomposition layer initialization unit executes the steps including setting the initial decomposition layer number of VMD to 2.

[0018] Embodiment 3: This embodiment is based on the above embodiment. In step T3, the multidimensional VMD decomposition unit executes the following steps: performing VMD decomposition on the preprocessed multidimensional comprehensive health data through multiple channels, and decomposing the preprocessed multidimensional comprehensive health data to obtain multidimensional IMF components according to the number of initial VMD decomposition layers.

[0019] Embodiment 4: This embodiment is based on the above embodiment. In step T3, the key value calculation unit executes the following steps: calculating the kurtosis and correlation coefficient of each dimension of the multidimensional IMF component, and taking the product of the kurtosis and the correlation coefficient as the key value.

[0020] Embodiment 5. This embodiment is based on the above embodiment. In step T3, the key value judgment unit executes the following steps: checking the minimum key value of the IMF component of each dimension in the multidimensional IMF component. If the minimum key value is ≤1, the IMF component is marked as a high-noise component and enters the stationary wavelet denoising; if the minimum key value is >1, the number of decomposition layers is increased and the multidimensional VMD decomposition is performed again until a multidimensional high-noise IMF component and a multidimensional noise-free IMF component are marked; a multidimensional consistency optimization algorithm is obtained through synchronous update technology and Lagrange multiplier constraint adjustment technology; and the multidimensional consistency optimization algorithm is used to further select and synchronize the multidimensional high-noise IMF components to obtain optimized multidimensional high-noise IMF components.

[0021] Embodiment 6. This embodiment is based on the above embodiment. In step T3, the multidimensional stationary wavelet denoising unit executes the following steps: selecting a wavelet basis for each dimension of the optimized multidimensional high-noise IMF component, applying a stationary wavelet transform to remove high-frequency noise, and obtaining a denoised multidimensional IMF component.

[0022] Embodiment 7: This embodiment is based on the above embodiment. In step T3, the steps performed by the multidimensional signal reconstruction unit specifically include recombining the denoised multidimensional IMF components and the multidimensional noise-free IMF components, performing multidimensional reconstruction, and forming denoised multidimensional comprehensive health data.

[0023] Embodiment 8, according to Figure 2 This embodiment is based on the above embodiment. In the key value judgment unit, the multidimensional high-noise IMF component is further selected and synchronized using the multidimensional consistency optimization algorithm to obtain a process of optimizing the multidimensional high-noise IMF component, which specifically includes the following steps: Step S1: Construct objective function: Construct a multi-dimensional joint optimization objective function. The formula used is as follows: ; in, represents the IMF component index, Indicates IMF components, Indicates The center frequency of the IMF component; represents the multi-dimensional joint optimization objective function, Represents the total number of dimensions, represents the dimension index, Indicates The weight coefficient of the dimension, represents the total number of IMF components, Indicates time The derivative of represents the imaginary unit, represents a complex exponential function with a center frequency, represents the square of the norm; Step S2: IMF component update: Based on the multi-dimensional joint optimization objective function, the multi-dimensional high-noise IMF components are synchronously updated to gradually reduce the noise in each dimension. The formula used is as follows: ; in, and Both represent IMF component indexes. represents a frequency variable, The index representing the number of iterations, Indicates The IMF component is The spectrum update value in the iteration, represents the spectrum of the original signal, Indicates The representation of the IMF components in the frequency domain is: Indicates that except The sum of the spectra of the other IMF components except the IMF components, represents the Lagrange multiplier, represents the smoothing parameter; Step S3: Lagrange multiplier update: After each synchronization update, Lagrange multipliers are introduced to perform constraint adjustment to ensure the frequency consistency of the multidimensional high-noise IMF component, iteratively update the parameters and Lagrange multipliers of the multidimensional high-noise IMF component, and finally output the optimized multidimensional high-noise IMF component. The formula used is as follows: ; in, The index representing the number of iterations, Indicates The updated value of the Lagrange multiplier at the iteration, Indicates The Lagrange multiplier of the iteration, represents the update step coefficient, Indicates The IMF component is The updated value at the iteration, represents the sum of the spectra of the updated IMF components; The algorithm first uses adaptive variational mode decomposition technology to perform a preliminary decomposition of the multidimensional signal, and then ensures the balance and consistency of each signal component in the multidimensional space through synchronous update and Lagrange multiplier constraint adjustment technology. Finally, it combines the stationary wavelet transform to further remove noise. This adaptive denoising process overcomes the limitations of traditional filtering and fixed mode decomposition methods that are difficult to cope with the complex noise in dynamic driving environments, significantly improves the accuracy of signal denoising, provides high-quality input data support for subsequent health status assessment, and ensures that clear and reliable physiological data can be obtained in different environments.

[0024] Embodiment 9, according to Figure 3 This embodiment is based on the above embodiment. In step T4, the process of inputting the denoised multi-dimensional comprehensive health data into the enhanced sparse autoencoder to perform driver abnormality detection and judge the driver's health status specifically includes the following steps: Step T41: Data input: The enhanced sparse autoencoder includes an encoding layer, a hidden layer, and a decoding layer. The weights of the enhanced sparse autoencoder are initialized, and the denoised multi-dimensional comprehensive health data is input into the encoding layer; Step T42: forward propagation encoding: convert the denoised multi-dimensional comprehensive health data into a low-dimensional sparse representation through the encoding layer, and set the sparsity parameters to control the activity of the hidden layer neurons; Step T43: Calculate the sparsity loss: Calculate the sparsity regularization term through KL divergence to measure the deviation between the actual average activation value of the hidden layer neurons and the sparsity parameter. The formula used is as follows: ; in, represents the sparsity regularization term, represents the neuron index of the hidden layer, represents the total number of neurons in the hidden layer, represents the summation symbol, represents the sparsity parameter, Indicates The actual average activation value of the hidden layer neurons, represents the KL divergence, The first part of the KL divergence formula measures The deviation between the actual average activation value of the hidden layer neurons and the sparsity parameter in the activated state, Represents the second part of the KL divergence formula, measuring the The inactive state bias of hidden layer neurons; Step T44: Back propagation optimization: Introduce contrast loss, weighted reconstruction error and weight decay terms, and combine the sparsity regularization term to construct the total loss function. Minimize the total loss function through back propagation and adjust the weights of the enhanced sparse autoencoder. The formula used is as follows: ; in, represents the index of the sample, represents the total number of samples, represents the index of the feature, represents the total number of features, Express The sum of samples, Indicates that for each sample The sum of feature dimensions, Indicates The weight of the feature, Indicates The sample The original input value of the feature, Indicates The sample The reconstructed input value of features, Indicates The sample The squared error on the feature dimension; represents the weight of the sparsity regularization term, Represents the control coefficient of the weight decay term; and represents the neuron index, Indicates that from Neuron to The connection weights of neurons, represents the square of weight; represents the weight coefficient of contrast loss, represents contrast loss; Step T45: Health status assessment: according to the weight of the enhanced sparse autoencoder, the low-dimensional sparse representation is restored through the decoding layer to obtain reconstructed data, the reconstruction error is calculated according to the reconstructed data, and the driver's health status is judged according to the reconstruction error; The present invention further improves the accuracy of feature learning and health status assessment of denoised data by designing an enhanced sparse autoencoder; the enhanced sparse autoencoder optimizes the loss function by combining contrast loss, weighted reconstruction error, weight attenuation term and sparsity regularization term, effectively extracts the deep features of PPG signal data, and realizes the classification and anomaly detection of driver health status on this basis; compared with traditional autoencoders, the enhanced sparse autoencoder of the present invention can accurately identify fatigue, hypoxia and abnormal heart rate, effectively enhances the abnormal warning capability of the health monitoring system, and improves the reliability and accuracy of health assessment.

[0025] The present invention and its embodiments are described above, which is not restrictive. What is shown in the accompanying drawings is only one of the embodiments of the present invention, and the actual structure is not limited to this. In short, if ordinary technicians in this field are inspired by it and do not deviate from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all fall within the scope of protection of the present invention.

Claims

1. A PPG signal adaptive denoising method based on wavelet transform and autoencoder, characterized in that: The following steps are involved: Step T1: Data collection: Collect the driver's heart rate data, heart rate variability data, blood oxygen saturation data and pulse waveform data, and fuse them to obtain multi-dimensional comprehensive health data; Step T2: Data preprocessing: performing signal synchronization, DC offset removal, filtering and standardization on the multi-dimensional comprehensive health data to obtain pre-processed multi-dimensional comprehensive health data; Step T3: Data denoising: Combining adaptive variational mode decomposition, stationary wavelet denoising, multidimensional consistency optimization, synchronous update technology and Lagrange multiplier constraint adjustment technology, the MAVMD-SWT denoising algorithm is obtained, and the pre-processed multidimensional comprehensive health data is denoised by the MAVMD-SWT denoising algorithm to obtain denoised multidimensional comprehensive health data; the MAVMD-SWT denoising algorithm specifically includes a decomposition layer initialization unit, a multidimensional VMD decomposition unit, a key value calculation unit, a key value judgment unit, a multidimensional stationary wavelet denoising unit and a multidimensional signal reconstruction unit; Step T4: Anomaly detection: Combine contrast loss, weighted reconstruction error, weight decay term and sparsity regularization term to optimize the total loss function of the autoencoder to obtain an enhanced sparse autoencoder. Input the denoised multi-dimensional comprehensive health data into the enhanced sparse autoencoder to perform driver anomaly detection and judge the driver's health status.

2. The method for adaptive denoising of PPG signals based on wavelet transform and autoencoder according to claim 1, characterized in that: In step T3, the decomposition layer initialization unit executes the steps including setting the initial VMD decomposition layer to 2.

3. The method for adaptive denoising of PPG signals based on wavelet transform and autoencoder according to claim 2, characterized in that: In step T3, the multidimensional VMD decomposition unit executes the steps of performing VMD decomposition on the pre-processed multidimensional comprehensive health data through multiple channels, and decomposing the pre-processed multidimensional comprehensive health data into multidimensional IMF components according to the number of initial VMD decomposition layers.

4. The method for adaptive denoising of PPG signals based on wavelet transform and autoencoder according to claim 3, characterized in that: In step T3, the key value calculation unit executes the steps of calculating the kurtosis and correlation coefficient of each dimension of the multidimensional IMF component, and taking the product of the kurtosis and the correlation coefficient as the key value.

5. The method for adaptive denoising of PPG signals based on wavelet transform and autoencoder according to claim 4, characterized in that: In step T3, the key value judgment unit executes the following steps specifically: checking the minimum key value of the IMF component of each dimension in the multidimensional IMF component; if the minimum key value is ≤1, marking the IMF component as a high-noise component and entering stationary wavelet denoising; if the minimum key value is >1, increasing the number of decomposition layers and re-performing multidimensional VMD decomposition until a multidimensional high-noise IMF component and a multidimensional noise-free IMF component are marked; obtaining a multidimensional consistency optimization algorithm through synchronous update technology and Lagrange multiplier constraint adjustment technology; and using the multidimensional consistency optimization algorithm to further select and synchronize the multidimensional high-noise IMF components to obtain optimized multidimensional high-noise IMF components.

6. The method for adaptive denoising of PPG signals based on wavelet transform and autoencoder according to claim 5, characterized in that: In step T3, the multidimensional stationary wavelet denoising unit performs the following steps: selecting a wavelet basis for each dimension of the optimized multidimensional high-noise IMF component, applying stationary wavelet transform to remove high-frequency noise, and obtaining a denoised multidimensional IMF component.

7. The method for adaptive denoising of PPG signals based on wavelet transform and autoencoder according to claim 6, characterized in that: In step T3, the steps performed by the multidimensional signal reconstruction unit specifically include recombining the denoised multidimensional IMF components and the multidimensional noise-free IMF components, performing multidimensional reconstruction, and forming denoised multidimensional comprehensive health data.

8. The method for adaptive denoising of PPG signals based on wavelet transform and autoencoder according to claim 5, characterized in that: In the key value judgment unit, the multidimensional high-noise IMF component is further selected and synchronized using the multidimensional consistency optimization algorithm to obtain a process of optimizing the multidimensional high-noise IMF component, which specifically includes the following steps: Step S1: construct objective function: construct multi-dimensional joint optimization objective function; Step S2: IMF component update: based on the multi-dimensional joint optimization objective function, the multi-dimensional high-noise IMF components are synchronously updated to gradually reduce the noise in each dimension; Step S3: Lagrange multiplier update: After each synchronization update, Lagrange multipliers are introduced to perform constraint adjustment to ensure the frequency consistency of the multidimensional high-noise IMF component, and the parameters and Lagrange multipliers of the multidimensional high-noise IMF component are iteratively updated to finally output the optimized multidimensional high-noise IMF component.

9. The method for adaptive denoising of PPG signals based on wavelet transform and autoencoder according to claim 1, characterized in that: In step T4, the process of inputting the denoised multi-dimensional comprehensive health data into the enhanced sparse autoencoder to perform driver abnormality detection and determine the driver's health status specifically includes the following steps: Step T41: Data input: The enhanced sparse autoencoder includes an encoding layer, a hidden layer, and a decoding layer. The weights of the enhanced sparse autoencoder are initialized, and the denoised multi-dimensional comprehensive health data is input into the encoding layer; Step T42: forward propagation encoding: convert the denoised multi-dimensional comprehensive health data into a low-dimensional sparse representation through the encoding layer, and set the sparsity parameters to control the activity of the hidden layer neurons; Step T43: Calculate the sparsity loss: Calculate the sparsity regularization term through KL divergence to measure the deviation between the actual average activation value of the hidden layer neurons and the sparsity parameter; Step T44: Back propagation optimization: introduce contrast loss, weighted reconstruction error and weight decay terms, and construct a total loss function in combination with the sparsity regularization term. Minimize the total loss function through back propagation and adjust the weights of the enhanced sparse autoencoder. Step T45: Health status assessment: According to the weights of the enhanced sparse autoencoder, the low-dimensional sparse representation is restored through the decoding layer to obtain reconstructed data, the reconstruction error is calculated based on the reconstructed data, and the driver's health status is judged based on the reconstruction error.

Images