Noninvasive blood glucose estimation method based on quaternary convolutional network
Through the method based on quaternary convolution network, discrete wavelet transformation and Mel spectrum coefficient extraction, combined with quaternary signal processing, the problems of noise and environmental interference in non-invasive blood glucose monitoring are solved, and high-precision and anti-interference blood glucose estimation are achieved. , suitable for wearable devices.
Patent Information
- Application Number
- CN202411029382.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-11-15
AI Technical Summary
The existing non-invasive blood glucose monitoring technology has noise and environmental interference when collecting photovoltaic pulse wave signals, and the sensor lacks sensitivity and accuracy when monitoring trace blood glucose changes, making it difficult to achieve individualized adaptation.
Using a method based on quaternary convolution network, the discrete wavelet transform denoising processing, Mel spectrum coefficient extraction and quaternary signal representation are used, combined with quaternary convolutional neural network, the number of parameters is reduced and the model efficiency and accuracy are improved.
Effectively denoising, improving the time and frequency resolution of the signal, enhancing the extraction of blood sugar-related information, improving the accuracy and anti-interference ability of non-invasive blood sugar estimation, and is suitable for wearable devices.
Smart Images

Figure CN118948270A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of blood sugar detection, and more specifically, to a non-invasive blood sugar estimation method based on a quaternion convolutional network. Background Art
[0002] Diabetes is a common chronic disease in modern life. Different types of diabetes may have different main causes, but no matter which type, it may lead to high blood sugar levels, which may cause serious health problems and complications. Therefore, the diagnosis and treatment of diabetes require close monitoring of blood sugar levels. Traditional blood sugar monitoring requires blood sampling by puncturing the finger, placing it on a pre-test strip, and monitoring the blood sugar value through an instrument. However, this invasive method of blood sugar testing is too uncomfortable and painful for the subject, and there is also a risk of infection, so it is not suitable for frequent use. And from a long-term perspective, blood sugar test strips are very expensive as consumables. Therefore, more and more diabetic patients are looking forward to non-invasive methods to achieve blood sugar monitoring. Near-infrared spectroscopy is a mainstream non-invasive blood sugar monitoring method. It uses the physical properties of light to monitor the glucose level in the blood. Because the skin has different absorption characteristics for light of different wavelengths, the transmission rate of near-infrared spectrum through the epidermis of the skin is 90-95%, which can better obtain glucose information. The photoelectric volume pulse wave signal (PPG) can reflect the change of blood volume in human blood vessels with the contraction of the heart by using photoelectric information. By placing a near-infrared light source on the fingertips, the principle of transmission is used to transmit the light changes passing through the skin tissue to the photodetector to obtain a light signal containing glucose information.
[0003] There are many wearable devices on the market that can achieve non-invasive blood sugar estimation. However, the following problems still exist: when collecting photoplethysmographic signals, there is noise, which makes it difficult to detect; when users exercise, environmental factors such as external light and temperature changes will interfere with sensor signals and affect measurement results; current sensor technology still lacks sensitivity and accuracy when monitoring trace blood sugar changes; algorithms need to be optimized for individual differences to improve data accuracy and individual adaptability. Summary of the invention
[0004] The present application provides a non-invasive blood glucose estimation method based on a quaternion convolutional network, which is used to solve the needs of both high precision and anti-interference in the prior art, and can be used on wearable devices for blood glucose monitoring.
[0005] In view of this, the present application provides a non-invasive blood glucose estimation method based on a quaternary convolutional network, comprising:
[0006] The device uses near-infrared light with a wavelength of 660nm, a signal sampling rate of 1000Hz, and a collection time of 60s to obtain the target object's photoplethysmographic signal (PPG);
[0007] Pre-process all collected PPG signals and use discrete wavelet transform to remove baseline drift and high-frequency noise of the signals, thereby improving the accuracy and reliability of subsequent signal processing;
[0008] Extract the logarithmic Mel filter group coefficient part of the preprocessed PPG signal. First, the PPG signal is windowed. The selected window is the Hamming window, which can effectively reduce the edge effect and provide a smooth edge. After windowing, the PPG signal can be guaranteed to be quasi-static in a short time. Where w[n] is the value of the window function at the nth sample. N is the length of the window (total number of samples), and n is the sample index of the window function.
[0009] Perform discrete Fourier transform (DFT) on each window to obtain the amplitude map; Where x[n] is the nth discrete time signal after windowing, and X[k] is the kth discrete frequency signal after transformation. is a complex exponential basis function.
[0010] The amplitude map is passed through Mel's triangular filter to obtain the Mel spectrum;
[0011] The Mel Spectral Coefficients (MFCC) are obtained by applying the discrete cosine transform (DCT) to the logarithmic Mel spectrum. Where x[n] is the nth discrete time signal after windowing, and X[k] is the kth discrete frequency signal after transformation.
[0012] The Mel spectrum coefficients are derivatized multiple times to obtain the first-order, second-order and third-order derivative coefficients, which are represented in parallel as quaternion signals. Where Q(f,t) represents the quaternion signal of different frequencies f and different times t. e(f,t) represents the Mel spectrum coefficient (MFCC), and and They represent the first, second and third order derivatives of MFCC respectively, and i, j and k represent the three imaginary components of the quaternion.
[0013] The quaternion convolutional network is constructed based on quaternion algebra and convolutional network, and includes: a quaternion convolutional layer, a quaternion pooling layer, a quaternion dense layer and a dense layer.
[0014] The quaternion convolutional network achieves weight sharing through Hamilton product, reduces the parameters required for the network, and is more suitable for use in wearable devices with limited resources.
[0015] The present invention has the following advantages: First, the discrete wavelet transform (DWT) is used to denoise the signal, providing both the time and frequency information of the signal at the same time. Through multi-scale analysis, the high-frequency part has high time resolution, while the low-frequency part has high frequency resolution. Through multi-resolution analysis, DWT can effectively capture the details of the signal and is suitable for processing various complex signals. Second, the Mel Spectral Coefficient (MFCC) is used as a feature, which can effectively capture the frequency domain characteristics in the signal, which is very important for analyzing blood glucose related information in PPG (photoplethysmography) signals. Through the first, second, and third order derivatives, the dynamic characteristics of the signal over time can be captured, which is particularly important for detecting changes in blood glucose levels. MFCC performs well in dealing with noise and interference, and can extract effective features from noisy signals, which helps to improve the accuracy of non-invasive blood glucose estimation. Third, a quaternion convolutional network is proposed based on quaternion and convolutional neural network, which achieves accurate non-invasive blood glucose estimation by using the quaternion network to learn the intrinsic connection between different order derivatives of MFCC. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 It is a flow chart of a non-invasive blood glucose estimation method based on a quaternary convolutional network of the present invention;
[0017] Figure 2 Schematic diagram of the quaternary convolutional network structure. DETAILED DESCRIPTION
[0018] In order to enable those skilled in the art to better understand the method of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application, and it is understandable to those skilled in the art that some well-known structures and their descriptions in the drawings may be omitted. The described embodiments are only part of the embodiments of the present application, not all of the embodiments.
[0019] The present invention provides a non-invasive blood glucose estimation method based on a quaternary convolutional network, and the overall process is as follows: Figure 1 As shown, the following steps are included:
[0020] The device uses near-infrared light with a wavelength of 660nm, a signal sampling rate of 1000Hz, and a collection time of 60s to obtain the target object's photoplethysmographic signal (PPG);
[0021] Preprocess all samples in the data set. First, perform discrete wavelet transform (DWT) on the collected PPG signal, and use the multi-resolution analysis characteristics of wavelet transform to decompose the signal into different frequencies and time scales, so as to effectively separate the signal and noise. First, select a suitable wavelet basis function to decompose the original signal into several layers of wavelet coefficients, including low-frequency approximate coefficients and high-frequency detail coefficients. Perform threshold processing on the decomposed coefficients to achieve the functions of signal denoising and baseline drift removal. And reconstruct it into a denoised signal through inverse discrete wavelet transform, which can improve the signal-to-noise ratio of the PPG signal;
[0022] The denoised signal is windowed to obtain a short segment signal with stable characteristics. The Hamming window can reduce edge effects and provide smooth edges. The windowing is achieved by multiplying the signal by the Hamming window, where the overlap rate of the window is 30%; Where w[n] is the value of the window function at the nth sample. N is the length of the window (total number of samples), and n is the sample index of the window function.
[0023] Use discrete Fourier transform (DFT) for each window frame after windowing to convert the signal into an amplitude map, and multiply the amplitude map by a triangular filter with the Mel frequency to obtain the Mel spectrum; Where x[n] is the nth discrete time signal after windowing, and X[k] is the kth discrete frequency signal after transformation. is a complex exponential basis function.
[0024] Take the logarithm of the obtained Mel spectrum and use discrete cosine transform (DCT) to obtain Mel spectrum coefficients (MFCC). DCT helps to concentrate most of the energy of the signal on a few low-frequency coefficients. This means that the first few DCT coefficients contain the main information of the signal, while the high-frequency coefficients can usually be ignored, thereby achieving data compression. DCT can sparsely represent the signal, which is very useful in compressed sensing and signal reconstruction. Sparse representation helps to reconstruct the original signal at a lower sampling rate. Where x[n] is the nth discrete time signal after windowing, and X[k] is the kth discrete frequency signal after transformation.
[0025] For the extracted MFCC features, the first, second and third order derivatives are calculated and expressed as quaternion signals; Where Q(f,t) represents the quaternion signal of different frequencies f and different times t. e(f,t) represents the Mel spectrum coefficient (MFCC), and and They represent the first, second and third order derivatives of MFCC respectively, and i, j and k represent the three imaginary components of the quaternion.
[0026] Because traditional convolutional neural networks need to learn parameters for each channel independently when processing multi-channel inputs. This leads to a substantial increase in the number of parameters, thereby increasing the storage requirements and computational complexity of the model. Convolution operations and pooling operations will cause the spatial position information of the input image to be lost. Although convolution kernels can capture local features, as the number of layers increases, position information may gradually be lost, affecting the network's ability to model spatial relationships. Therefore, the present invention proposes the use of a quaternion-based convolutional neural network.
[0027] The constructed quaternion convolution network consists of four parts: quaternion convolution layer, quaternion pooling layer, quaternion dense layer and dense layer. The convolution kernel of QCNN is composed of quaternions, so the number of parameters is significantly reduced compared to traditional CNN. This not only reduces the computational complexity of the model, but also reduces memory usage, making model training and inference more efficient. Quaternions not only contain amplitude information, but also represent the phase relationship between signals. For some applications, such as image processing and signal processing, phase information is very important. QCNN can maintain and utilize these phase relationships, improving the accuracy of feature extraction;
[0028] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A non-invasive blood glucose estimation method based on a quaternion convolutional network, characterized in that: include: Step 1: Use near-infrared light technology to collect photoplethysmography (PPG) signals from the fingertips; Step 2: Preprocess all collected PPG signals and use discrete wavelet transform to remove baseline drift and high-frequency noise of the signals, thereby improving the accuracy and reliability of subsequent signal processing; Step 3: Extract the logarithmic Mel filter bank coefficients of the preprocessed PPG signal, calculate the first, second, and third order derivatives, convert them into quaternion signals and perform normalization; Step 4: Then construct a quaternion convolutional neural network, and use the training set and loss function to train the constructed quaternion neural network; Step 5: Finally, the PPG signal data to be detected is processed in the same way and input into the trained neural network for regression prediction, and finally the result of non-invasive blood glucose estimation is obtained.
2. The non-invasive blood glucose estimation method based on a quaternary convolutional network according to claim 1, characterized in that: The device for collecting PPG signals uses near-infrared light with a wavelength of 660nm, the signal sampling rate is 1000Hz, and the collection time is 60s.
3. The non-invasive blood glucose estimation method based on a quaternary convolutional network according to claim 1, characterized in that: The discrete wavelet transform decomposes the signal into sub-band coefficients of different scales, removes low-frequency components and high-frequency components, and then performs inverse wavelet transform on the processed coefficients to obtain a signal with high-frequency noise and baseline drift removed.
4. The non-invasive blood glucose estimation method based on a quaternary convolutional network according to claim 1, characterized in that: Extract the logarithmic Mel filter group coefficient part of the preprocessed PPG signal. First, use the Hamming window to reduce the edge effect, perform discrete Fourier transform (DFT) on each window, obtain the amplitude map, pass the amplitude map through a set of Mel filters, take the logarithm to obtain the Mel spectrum coefficient (MFCC), and solve the multi-order derivatives to express it as a unit quaternion signal.
5. The non-invasive blood glucose estimation method based on a quaternary convolutional network according to claim 1, characterized in that: The quaternion convolutional neural network is an extension of the convolutional neural network based on quaternion algebra, which can explore more hidden information.
