An accelerometer-based physiological information processing method and system
By acquiring data using accelerometers and utilizing pre-emphasis processing, adaptive filtering, and neural network technology, physiological signals are separated and enhanced, solving the problem of missing high-frequency information in bone conduction signals and achieving more accurate physiological information monitoring and diagnosis.
Patent Information
- Application Number
- CN202310938397.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-28
- Publication Date
- 2026-02-17
- Estimated Expiration
- 2043-07-28
AI Technical Summary
Existing technologies suffer from a lack of high-frequency information in bone conduction signals, which affects the quality of physiological information acquisition systems and monitoring and diagnostic results.
Physiological information is collected using an accelerometer. Through pre-emphasis processing, adaptive filtering, bandwidth expansion model and neural network processing, the signal is separated and enhanced to recover high-frequency information.
This improved the quality of the physiological information acquisition system, resulting in more accurate monitoring and diagnostic results and overcoming the problem of high-frequency data loss.
Smart Images
Figure CN116687379B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of physiological information processing, and more particularly to a physiological information processing method and system based on an accelerometer. BACKGROUND
[0002] Common audio physiological signals such as breathing sound, heartbeat sound and snoring sound have the characteristics of weak signal and being easily interfered by environmental noise and other noise. In an extremely quiet environment, a suitable method for collecting the above physiological signals is to use an air-conducted microphone. However, it is difficult to maintain a quiet enough environment in actual scenes, and the physiological signals collected by using the air-conducted microphone are often contaminated by environmental noise, especially the weak breathing sound and heartbeat sound, which are easily submerged by environmental noise. Therefore, in a noisy actual scene, it is difficult to collect clean physiological signals by using the air-conducted microphone, which brings great difficulty to subsequent physiological information processing and analysis.
[0003] The bone-conduction physiological information collection based on the accelerometer collects physiological signals by directly contacting the skin, and theoretically cannot collect signals transmitted through the air, so it can be free from the influence of environmental noise, and even in a noisy environment, the influence of environmental noise on weak signals can be avoided, so that relatively clean physiological signals can be collected by using the accelerometer and other devices. However, due to the inherent low-pass characteristic of the human body and the limitation of the sensor technology level, the collected signals lack high-frequency information, and the high-frequency part of the physiological signal also contains important physiological information as the low-frequency part, so it is necessary to perform spectrum bandwidth extension on the physiological signal with missing high-frequency information to restore the missing high-frequency information, which is helpful for more accurate monitoring and research of physiological information.
[0004] In the related art, a portable sleep respiratory state monitoring device is designed in a Chinese thesis (Sleep Respiratory State Monitoring and Health Analysis Based on Bone Conduction Microphone[D]. On Juncheng. Tianjin University, 2016), the sensor includes a bone conduction and air conduction microphone, a gyroscope, and the respiratory signal and snoring signal of the subject during sleep are synchronously collected, which is used to monitor the respiratory state and heart rate change of the subject during sleep. In this process, the bone conduction signal directly uses the high-frequency information missing, and no enhancement processing is performed on the signal. For example, an embedded system based on a bone conduction microphone is designed in a Chinese thesis (Research on Bone Conduction Snoring Detection System[D]. Liu Tao. Jilin University, 2020), which can collect bone conduction snoring signals, calculate AHI according to the snoring silent section, diagnose the severity of OSA, and achieve good diagnostic results. However, the above-mentioned physiological signal collection systems do not perform repair processing on the bone conduction signal missing high-frequency information, and the missing high-frequency information will affect the final monitoring results to some extent. The current bone conduction device for collecting physiological information does not have the defect of missing high-frequency information in the bone conduction signal, and the quality of the physiological information collection system is deeply affected.
[0005] As can be seen from the above, the related art does not provide an effective solution to the problem of missing high-frequency information in the bone conduction signal. SUMMARY
[0006] 1. Technical problem to be solved
[0007] In view of the problem of missing high-frequency information in the bone conduction signal in the prior art, the present application provides a physiological information processing method and system based on an accelerometer, which can enhance and repair the damaged physiological signal to make it closer to the real physiological signal, thereby obtaining more accurate and objective monitoring and diagnosis results.
[0008] 2. Technical scheme
[0009] The object of the present application is achieved by the following technical scheme.
[0010] A physiological information processing method based on an accelerometer, comprising the following steps,
[0011] Collecting physiological information: collecting the neck sound signal of the subject and transmitting the collected signal to the computer for storage and processing;
[0012] Processing physiological information: including separating signals and enhancing signals.
[0013] Further, the neck sound signal includes snoring sound, breathing sound and carotid pulse sound.
[0014] Signal separation includes pre-emphasis processing, signal filtering, and separation of mixed signals;
[0015] Signal enhancement includes data preprocessing, establishing a bandwidth expansion model, extracting signal features, enhancing cross-channel features, and generating and judging signals.
[0016] Furthermore, in the pre-emphasis processing, the filter frequency response function for:
[0017]
[0018] This is the pre-emphasis coefficient, a value close to 1;
[0019] In the filtered signal, a low-pass filter is used to separate the neck pulse sound, and the remaining signal consists of snoring and breathing sounds.
[0020] Furthermore, separating the mixed signal involves separating the snoring and breathing sounds from the mixed signal of snoring and breathing sounds. The steps are as follows:
[0021] set up The input mixed signal sequence:
[0022] ,
[0023] It is the length of the filter. Represents the transpose of a vector. For the index of the signal sequence, It is the first of the signal sequence Numerical points;
[0024] The filter output signal is The expected signal is Then the error signal .
[0025] Furthermore, the filter output signal can be represented as:
[0026] ,
[0027] The error signal is represented as:
[0028] ,
[0029] Mean square error is expressed as:
[0030] ,
[0031] The autocorrelation matrix is expressed as:
[0032] ,
[0033] The cross-correlation matrix is expressed as:
[0034] ,
[0035] In the mean square error expression, for the filter coefficients Taking the derivative, we can obtain the gradient as:
[0036] ,
[0037] Setting the gradient to zero, the optimal filter coefficients are:
[0038] ,
[0039] yes The inverse matrix.
[0040] Furthermore, data preprocessing involves preprocessing the separated physiological signals. Preprocessing methods include normalization and framing. Framing needs to be set so that there is some overlap between adjacent frames.
[0041] The bandwidth extension model includes an encoding module, a cross-channel feature enhancement module, and a decoding module.
[0042] Furthermore, the feature extraction of the signal specifically involves: reducing the dimensionality of the original data and extracting its deep features through the encoding module; supplementing the relationships between feature channels through the cross-channel feature enhancement module; and recovering the signal from the deep features through the decoding module, thus restoring the encoded data to its original data dimension.
[0043] The encoding module consists of three identical encoding units, each of which consists of a residual unit, a one-dimensional convolutional layer, and a ReLU activation function. When data passes through the residual unit, the data dimension remains unchanged. When data passes through the one-dimensional convolutional layer, the number of feature channels is increased and the data dimension is compressed.
[0044] The decoding module consists of three identical decoding units, each consisting of a one-dimensional transposed convolution, a ReLU activation function, and a residual unit. When data passes through the residual unit, the data dimension is not changed, and the encoded signal is restored to the original data dimension after decoding.
[0045] The encoding and decoding modules are symmetrical structures, concatenating data of the same dimension along the feature channel dimension, and then decoding the corresponding data in the decoding module.
[0046] Furthermore, the cross-channel features are enhanced by redistributing the weights across channels, as follows:
[0047] The input data is mapped to the feature space through transformation operations;
[0048] Feature compression is performed by generating channel information through a global average pooling layer. Input converted The output generates C One feature channel;
[0049] The compressed feature result is subjected to a fully connected operation to fuse the feature information of each channel and output the weight coefficient of each feature channel.
[0050] The weight coefficients of each feature channel are weighted channel by channel through a channel-wise multiplication operation and applied to the feature map output by the encoding module, thereby completing the recalibration of the encoded features in the channel dimension.
[0051] Furthermore, the specific steps for generating and judging the signal are as follows: performing the inverse frame-segmentation operation on the output data of the decoding module, generating a time-domain audio signal using the overlap-addition method, which is a physiological signal after spectral broadening; and evaluating performance in both the time and frequency domains, as follows:
[0052] The signal-to-noise ratio is calculated as follows:
[0053] ,
[0054] For target signal, For the enhanced signal, For the signal's first frame, The first frame of a signal Data points, The number of numerical points in a frame of signal;
[0055] The calculation method for spectral distortion is as follows:
[0056] ,
[0057] The target signal's logarithmic spectrum. To enhance the logarithmic spectrum of the signal, For the signal's first frame, The total number of frames in the signal. The first frame of a signal One frequency point;
[0058] The resonance peak loss is calculated as follows:
[0059] ,
[0060] To enhance the signal The resonance peak of the frame, For the target signal The resonance peak of the frame.
[0061] The system based on the above-mentioned accelerometer-based physiological information processing method includes,
[0062] The information acquisition module is used to collect physiological information. It uses an accelerometer to collect the neck sound signal of the subject. The accelerometer is specifically worn at the carotid artery pulse point. At the same time, the collected signals are transmitted to a computer for storage and processing.
[0063] The signal processing module is used to separate and enhance signals, and specifically includes a signal separation unit and a signal enhancement unit.
[0064] 3. Beneficial effects
[0065] Compared with the prior art, the advantages of this invention are:
[0066] This invention discloses a physiological information processing method and system based on an accelerometer. By enhancing and repairing damaged physiological signals at high frequencies, the collected physiological information becomes closer to the real physiological signals, resulting in more accurate and objective monitoring and diagnostic results. This invention uses a neural network-based method to repair damaged signals at high frequencies, solving the problem of high-frequency missing information in the collected physiological information, improving the quality of the physiological information acquisition system, and thus increasing the accuracy of people using physiological information for health monitoring or disease diagnosis. Attached Figure Description
[0067] Figure 1 This is a schematic flowchart of a physiological information processing method based on an accelerometer according to an embodiment of the present invention;
[0068] Figure 2 This is a schematic diagram of the process for processing physiological information in one embodiment of the present invention;
[0069] Figure 3 This is a schematic flowchart of the signal separation process in one embodiment of the present invention;
[0070] Figure 4 This is a schematic diagram of the signal enhancement process in one embodiment of the present invention;
[0071] Figure 5 This is a flowchart illustrating a physiological information processing system based on an accelerometer according to an embodiment of the present invention.
[0072] Figure 6 This is a schematic diagram illustrating the separation effect according to an embodiment of the present invention;
[0073] Figure 7 This is a time-frequency diagram of a signal acquired by an accelerometer according to an embodiment of the present invention;
[0074] Figure 8 This is a time-frequency schematic diagram of the target signal according to an embodiment of the present invention;
[0075] Figure 9 This is a time-frequency diagram of the high-frequency recovered signal according to an embodiment of the present invention. Detailed Implementation
[0076] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.
[0077] Combination Figures 1 to 9 The present invention provides a physiological information processing method based on an accelerometer, comprising the following steps:
[0078] Collect physiological information
[0079] Accelerometers are used to collect sound signals from the subject's neck, including snoring, breathing sounds, and carotid pulse sounds. The accelerometer is worn at the carotid artery pulse point, specifically between the anterior border of the sternocleidomastoid muscle and the thyroid cartilage. The collected signals are then transmitted to a computer for storage and processing.
[0080] Processing physiological information
[0081] Processing physiological information includes signal separation and signal enhancement. The specific steps for signal separation are as follows:
[0082] Pre-weighting treatment:
[0083] Pre-emphasis processing can flatten the signal spectrum and reduce loss in the high-frequency components. This process is achieved through filters, and the filter's frequency response function... for:
[0084]
[0085] In the response function, This is the pre-emphasis coefficient, a value close to 1, here we take 0.98; It is a complex variable; It represents the unit delay of a sampling point and is used to describe the frequency response of a discrete filter.
[0086] Filtered signal:
[0087] Since the frequency range of the neck pulse sound is low, generally below 100Hz, which is much smaller than the frequency range of snoring and breathing sounds, a Butterworth low-pass filter with a cutoff frequency of 100Hz is used to filter the signal and separate the neck pulse sound. The remaining signal is snoring and breathing sounds.
[0088] Separate mixed signals:
[0089] Separating the mixed signal involves applying adaptive filtering to separate snoring and breathing sounds. This can be viewed as an unknown system where the input is a mixed signal of snoring and breathing, and the output is snoring and breathing sounds. For this unknown system, the self-adjusting capability of the adaptive filter allows the output to gradually approximate clean snoring and breathing sounds. The specific steps are as follows:
[0090] set up The input mixed signal sequence:
[0091]
[0092] in, It is the length of the filter. Represents the transpose of a vector. For the index of the signal sequence, It is the first of the signal sequence The filter output signal is given by [number] numerical points. The expected signal is Then the error signal .
[0093] The filter output can be expressed as:
[0094]
[0095] For the filter in the first The weight coefficients at each delay unit constitute a length of... The weight vector; This represents a column vector composed of all value vectors arranged in order.
[0096] The error signal can be represented as:
[0097]
[0098] The mean square error can be expressed as:
[0099]
[0100] in The autocorrelation matrix can be represented as:
[0101]
[0102] The cross-correlation matrix can be represented as:
[0103]
[0104] In the mean square error expression, for the filter coefficients Taking the derivative, we can obtain the gradient as:
[0105]
[0106] Setting the gradient to zero, the optimal filter coefficients are:
[0107]
[0108] in yes The inverse matrix.
[0109] Through continuous iteration of the adaptive filter, when the error signal gradually stabilizes and approaches 0, the filter coefficients can be regarded as the system's impulse response, thus achieving signal separation. For the separation of snoring and breathing sounds, this process requires the use of an adaptive filter to simulate this unknown separation system.
[0110] The unknown system is input using a mixed signal of snoring and breathing sounds, and output using pure snoring sounds. An adaptive filter is then iterated using this method. When the error signal stabilizes, the weighting coefficients can be used as the coefficients of the adaptive filter.
[0111] After obtaining the weight coefficients of the adaptive filter, the remaining signal is used for testing. The mixed signal of snoring and breathing is taken as input and passed through the obtained adaptive filter. The output is the separated snoring signal, while the breathing signal is obtained by subtracting the mixed signal from the separated snoring signal.
[0112] Signal enhancement involves expanding the spectral bandwidth of the separated signal to achieve high-frequency recovery of physiological signals. The specific implementation steps are as follows:
[0113] Data preprocessing:
[0114] The separated physiological signals are preprocessed, including normalization and framing. Framing is required to allow for some overlap between adjacent frames.
[0115] Establish a bandwidth expansion model:
[0116] The bandwidth extension model includes an encoding module, a decoding module, and a cross-channel feature enhancement module. The encoding module is used for feature extraction, which reduces the dimensionality of the original data and extracts its deep features. The cross-channel feature enhancement module is used to supplement the relationships between feature channels. The decoding module is used for signal recovery, which restores the deep feature information.
[0117] Extracting signal features:
[0118] The encoding module consists of three identical encoding units used to extract signal features layer by layer. Each encoding unit consists of a residual unit, a one-dimensional convolutional layer, and a ReLU activation function. Adding a residual unit can alleviate gradient vanishing and prevent network degradation while extracting features.
[0119] When data passes through residual units, the data dimension remains unchanged. When passing through one-dimensional convolutional layers, the data dimension is compressed while increasing the number of feature channels. Specifically, the kernel lengths are set to 8, 6, and 4, corresponding to 64, 128, and 256 output channels, respectively, with a convolution stride of 2 for all layers. A ReLU activation function is added after each convolutional layer. After feature extraction by the encoding module, the original signal generates 256-dimensional features. The main function of this module is to extract shallow features from the original input data, gradually increasing the number of feature channels and decreasing the kernel length to minimize the loss of important information in the original data.
[0120] Enhanced cross-channel characteristics:
[0121] Enhancing cross-channel features involves strengthening the connections between feature channels through a cross-channel feature enhancement module. This involves redistributing weights among channels, clarifying the interdependencies between them, helping the network learn important feature information, and using global information to selectively emphasize useful features while suppressing useless ones. The specific implementation steps are as follows:
[0122] The input data is mapped to the feature space through transformation operations;
[0123] Feature compression is performed by generating channel information through a global average pooling layer. Input converted The output generates C One feature channel;
[0124] The compressed feature result is subjected to a fully connected operation to fuse the feature information of each channel, and the output result is the weight coefficient of each feature channel.
[0125] The weight coefficients of each feature channel represent the importance of each feature channel. By multiplying each channel, the weights are applied to the feature map output by the encoding module, thus completing the recalibration of the encoded features in the channel dimension.
[0126] Generate a signal and determine its nature:
[0127] The decoding module consists of three identical decoding units, which are used to recover the signal from the features layer by layer. Each decoding unit consists of a one-dimensional transposed convolution, a ReLU activation function, and a residual unit. The addition of the residual unit can alleviate gradient vanishing and prevent network degradation while recovering the signal.
[0128] When data passes through residual units, the data dimension remains unchanged. When passing through one-dimensional transposed convolutions, the data dimension is restored while reducing the number of feature channels. Specifically, the kernel lengths are set to 4, 6, and 8, corresponding to 256, 128, and 64 output channels, respectively, with a stride of 2 for each layer. A ReLU activation function is added after each convolutional layer. After decoding, the encoded signal is restored to its original data dimension. The main function of the decoding module is to recover the signal from deep features by gradually reducing the number of feature channels and increasing the kernel length to restore the encoded data to its original data dimension.
[0129] The encoding and decoding modules are symmetrical structures. A jump connection unit is added between data of the same dimension, that is, data with the same dimension are spliced together along the feature channel dimension, and the spliced data is the corresponding data in the decoding module.
[0130] The output data of the decoding module is subjected to the inverse operation of frame segmentation, that is, the time-domain audio signal is generated by the overlapping and addition method. This signal is the physiological signal after spectral broadening.
[0131] To measure the effectiveness of signal enhancement, performance metrics are proposed from both the time and frequency domains: signal-to-noise ratio, logarithmic spectral distortion, and formant loss. Details are as follows:
[0132] Signal-to-noise ratio (SNR) () measures signal recovery from a time-domain perspective, and is calculated as follows:
[0133]
[0134] in For target signal, For the enhanced signal, For the signal's first frame, The first frame of a signal Numerical points, This represents the number of numerical points in a single frame of signal.
[0135] Log-spectral distortion ( () measures signal recovery from the frequency domain perspective, and is calculated as follows:
[0136]
[0137] in The target signal's logarithmic spectrum. To enhance the logarithmic spectrum of the signal, For the signal's first frame, The total number of frames in the signal. The first frame of a signal One frequency point, This represents the number of numerical points in a single frame of signal.
[0138] The resonance parameters of snoring signals are important parameters reflecting the physiological structure of the upper airway; therefore, resonance loss ( This can measure the recovery of physiological signals, and the calculation method is as follows:
[0139]
[0140] in To enhance the signal The resonance peak of the frame, For the target signal The resonance peak of the frame.
[0141] The accelerometer-based physiological information processing method of this invention enhances damaged physiological signals and performs high-frequency repair on the damaged signals, making them closer to real physiological signals, thus obtaining more accurate and objective monitoring and diagnostic results. This invention uses a neural network-based method to perform high-frequency repair on damaged signals, overcoming the deficiency of high-frequency signals in bone conduction signals and improving the quality of the physiological information acquisition system.
[0142] Combination Figures 1 to 9 The present invention provides a physiological information processing system based on an accelerometer, comprising:
[0143] The information acquisition module is used to collect physiological information. An accelerometer is used to collect the subject's neck sound signals, including snoring, breathing sounds, and carotid pulse sounds. The accelerometer is specifically worn between the anterior edge of the sternocleidomastoid muscle and the thyroid cartilage in the subject's neck, i.e., at the carotid artery pulsation site. At the same time, the collected signals are transmitted to a computer for storage and processing.
[0144] The signal processing module is used to separate and enhance signals. Specifically, it includes a signal separation unit and a signal enhancement unit. The specific implementation steps of the signal separation unit are as follows:
[0145] Pre-emphasis processing makes the signal spectrum flatter and reduces loss in the high-frequency part, which is specifically achieved through filters;
[0146] Low-pass filtering uses a Butterworth low-pass filter with a cutoff frequency of 100Hz to filter the signal, separating the neck pulse sound, which is much smaller than the frequency range of snoring and breathing sounds. The remaining signal is snoring and breathing sounds.
[0147] Adaptive filtering is applied to separate snoring and breathing sounds, thus separating them from the mixed signal of snoring and breathing.
[0148] The signal enhancement unit then expands the spectral bandwidth of the separated signal to achieve high-frequency recovery of the physiological signal. The specific implementation steps are as follows:
[0149] Data preprocessing involves preprocessing the separated physiological signals. The preprocessing methods include normalization and framing, where the framing process is set to allow for partial overlap between adjacent frames.
[0150] A bandwidth expansion model is established, including an encoding module, a decoding module, and a cross-channel feature enhancement module. The encoding module is used for feature extraction, which reduces the dimensionality of the original data and extracts its deep features. The cross-channel feature enhancement module is used to supplement the relationships between feature channels. The decoding module is used for signal recovery, which restores the deep feature information.
[0151] The encoding module consists of three identical encoding units, which are used to extract signal features layer by layer. Each encoding unit consists of a residual unit, a one-dimensional convolutional layer and a ReLU activation function in sequence. Adding a residual unit can alleviate gradient vanishing and prevent network degradation while extracting features.
[0152] The cross-channel feature enhancement module is used to enhance the connection between feature channels. The main function of this module is to redistribute the weights of each channel, clarify the interdependencies between feature channels, help the network learn important feature information, use global information to selectively emphasize useful information features, and suppress useless feature information.
[0153] The decoding module consists of three identical decoding units, which are used to recover the signal from the features layer by layer. Each decoding unit consists of a one-dimensional transposed convolution, a ReLU activation function and a residual unit in sequence. The addition of the residual unit can alleviate gradient vanishing and prevent network degradation while recovering the signal.
[0154] The encoding and decoding modules are symmetrical structures. A skip connection unit is added between data of the same dimension, that is, data with the same dimension are spliced together in the feature channel dimension, and the spliced data replaces the corresponding data in the decoding module.
[0155] The output data of the decoding module is subjected to the inverse operation of frame segmentation, that is, the time-domain audio signal is generated by the overlap and addition method. This signal is the physiological signal after spectral broadening.
[0156] We propose performance metrics from both the time and frequency domains: signal-to-noise ratio, logarithmic spectral distortion, and formant loss, to measure the effect of signal enhancement.
[0157] This invention discloses a physiological information processing method and system based on an accelerometer. Utilizing a physiological information acquisition device with bone conduction characteristics, it offers advantages such as low cost, wearability, portability, and immunity to environmental noise interference. This type of device can be used to collect physiological information in a home environment for health monitoring or self-assessment during rehabilitation training. This invention also solves the problem of high-frequency missing data in the collected physiological information, making the collected physiological information closer to real-world data and improving the accuracy of using physiological information for health monitoring or disease diagnosis.
[0158] The invention and its embodiments have been described above illustratively. This description is not restrictive, and the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. The accompanying drawings are only one embodiment of the invention, and the actual structure is not limited thereto. No reference numerals in the claims should limit the scope of the claims. Therefore, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the invention, such design should fall within the scope of protection of this patent. Furthermore, the word "comprising" does not exclude other elements or steps, and the word "a" preceding an element does not exclude the inclusion of "a plurality" of that element. Multiple elements stated in the product claims may also be implemented by a single element through software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any specific order.
Claims
1. A physiological information processing method based on an accelerometer, comprising the following steps: Physiological information collection: Accelerometers were used to collect neck sound signals from the subjects, including snoring, breathing sounds and neck pulse sounds, and the collected signals were transmitted to a computer for storage and processing; Processing physiological information: including signal separation and signal enhancement; in, The separated signals include: Pre-emphasis processing: Implemented through a filter, the filter's frequency response function. , This is the pre-emphasis coefficient; Let be the frequency response function of the filter. It is a complex variable; It represents the unit delay at one sampling point and is used to describe the frequency response of a discrete filter; Signal filtering: The signal was filtered using a Butterworth low-pass filter with a cutoff frequency of 100Hz to separate the neck pulse sound; the remaining signal consisted of snoring and breathing sounds. Separating the mixed signal: Applying adaptive filtering to separate snoring and breathing sounds from the mixed signal; The enhanced signal includes: Data preprocessing: The separated physiological signals are normalized and framed. The framed processing is set to allow for partial overlap between adjacent frames. A bandwidth expansion model is established, comprising an encoding module, a decoding module, and a cross-channel feature enhancement module. The encoding module is used for feature extraction, reducing the dimensionality of the original data and extracting its deep features. The cross-channel feature enhancement module is used to supplement the relationships between feature channels. The decoding module is used to recover deep-level feature information. Extracting signal features: The coding module consists of three identical coding units, which are used to extract signal features layer by layer. Each coding unit consists of a residual unit, a one-dimensional convolutional layer and a ReLU activation function in sequence. Enhance cross-channel features: The cross-channel feature enhancement module strengthens the connection between feature channels and redistributes the weights of each channel. Signal generation and evaluation: Perform the inverse operation of frame segmentation on the output data of the decoding module, and generate a time-domain audio signal using the overlapping addition method; measure the performance in both the time and frequency domains.
2. The physiological information processing method based on an accelerometer according to claim 1, characterized in that, Separating the mixed signal involves separating snoring and breathing sounds from a mixed signal of snoring and breathing. The steps are as follows: set up The input mixed signal sequence: ， It is the length of the filter. Represents the transpose of a vector. For the index of the signal sequence, It is the first of the signal sequence Numerical points; The filter output signal is The expected signal is Then the error signal .
3. The physiological information processing method based on an accelerometer according to claim 2, characterized in that, The filter output signal is represented as: ， For the filter in the first The weight coefficients at each delay unit constitute a length of... The weight vector; The error signal is represented as: ， Mean square error is expressed as: ， The autocorrelation matrix is expressed as: ， The cross-correlation matrix is expressed as: ， In the mean square error expression, for the filter coefficients Taking the derivative, we can obtain the gradient as: ， Setting the gradient to zero, the optimal filter coefficients are: ， yes The inverse matrix.
4. The physiological information processing method based on an accelerometer according to claim 1, characterized in that, The specific steps for extracting signal features are as follows: the encoding module reduces the dimensionality of the original data and extracts its deep features; the cross-channel feature enhancement module supplements the relationships between feature channels; and the decoding module recovers the signal from the deep features, restoring the encoded data to its original dimensionality. The encoding module consists of three identical encoding units, each of which consists of a residual unit, a one-dimensional convolutional layer, and a ReLU activation function. When data passes through the residual unit, the data dimension remains unchanged. When data passes through the one-dimensional convolutional layer, the number of feature channels is increased and the data dimension is compressed. The decoding module consists of three identical decoding units, each consisting of a one-dimensional transposed convolution, a ReLU activation function, and a residual unit. When data passes through the residual unit, the data dimension is not changed, and the encoded signal is restored to the original data dimension after decoding. The encoding and decoding modules are symmetrical structures, concatenating data of the same dimension along the feature channel dimension, and then decoding the corresponding data in the decoding module.
5. The physiological information processing method based on an accelerometer according to claim 1, characterized in that, To enhance cross-channel features, the weights of each channel are redistributed as follows: The input data is mapped to the feature space through transformation operations; Feature compression is performed by generating channel information through a global average pooling layer. Input converted The output generates C There are 1 feature channels, among which H The height of the feature map, W The width of the feature map. C Number of feature channels; The compressed feature result is subjected to a fully connected operation to fuse the feature information of each channel and output the weight coefficient of each feature channel. The weight coefficients of each feature channel are weighted channel by channel through a channel-wise multiplication operation and applied to the feature map output by the encoding module, thereby completing the recalibration of the encoded features in the channel dimension.
6. The physiological information processing method based on an accelerometer according to claim 1, characterized in that, The specific steps for generating and judging the signal are as follows: performing the inverse frame segmentation operation on the output data of the decoding module, generating a time-domain audio signal using the overlap-addition method, which is a physiological signal after spectral broadening; and evaluating performance in both the time and frequency domains, as follows: The signal-to-noise ratio is calculated as follows: ， For target signal, For the enhanced signal, For the signal's first frame, The first frame of a signal Data points, The number of numerical points in a frame of signal; The calculation method for spectral distortion is as follows: ， The target signal's logarithmic spectrum. To enhance the logarithmic spectrum of the signal, For the signal's first frame, The total number of frames in the signal. The first frame of a signal One frequency point; The resonance peak loss is calculated as follows: ， To enhance the signal The resonance peak of the frame, For the target signal The resonance peak of the frame, This represents the total number of frames in the signal.
7. A system based on the physiological information processing method using an accelerometer according to any one of claims 1-6, characterized in that, include, The information acquisition module is used to collect physiological information. It uses an accelerometer to collect the neck sound signal of the subject. The accelerometer is specifically worn at the carotid artery pulse point. At the same time, the collected signals are transmitted to a computer for storage and processing. The signal processing module is used to separate and enhance signals, and specifically includes a signal separation unit and a signal enhancement unit.
Citation Information
Patent Citations
Depth and breadth neural network combined speech enhancement algorithm of digital hearing aid
CN105611477A
Voice bandwidth expansion method based on UNET architecture
CN115376543A