Wearable device-based electrocardiogram signal user authentication method
By collecting single-lead electrocardiogram signals on wearable devices and using convolutional autoencoders and fully connected neural networks for feature extraction and authentication, the problems of noise and counterfeiting in existing technologies are solved, achieving highly accurate and robust identity authentication.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI JIAOTONG UNIV
- Filing Date
- 2022-09-28
- Publication Date
- 2026-07-21
AI Technical Summary
Existing biometric identification technologies for wearable devices are susceptible to environmental noise, and traditional technologies such as fingerprint recognition are easily counterfeited, resulting in insufficient security.
Based on the acquisition of single-lead electrocardiogram signals by wearable devices, features are extracted using convolutional autoencoders, and user authentication is performed by combining fully connected neural networks. User tags are generated for identity authentication through preprocessing and feature point detection.
It improves the accuracy and robustness of identity authentication in high-noise environments, provides more intuitive biometric extraction, and enhances user data security and user experience.
Smart Images
Figure CN115587346B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a technology in the field of identity authentication, specifically a user authentication method based on electrocardiogram signals from wearable devices. Background Technology
[0002] In recent years, smart wearable devices on the market have integrated various sensors to collect data such as acceleration, orientation, and GPS. These sensors can collect data for biometric identification, using inherent physiological and behavioral characteristics of the human body, such as fingerprints, voiceprints, irises, gait, and even breathing, to authenticate an individual's identity. Among existing technologies, voiceprint recognition is easily affected by environmental noise, iris recognition technology requires sufficient light to capture the eye, and even widely used technologies like fingerprint recognition are often questioned regarding their security, as fingerprints are easily forged. Summary of the Invention
[0003] This invention addresses the shortcomings of existing technologies in processing high-noise signals acquired from wearable devices by proposing a user authentication method based on ECG signals from wearable devices. The method involves acquiring single-lead ECG signals from wearable devices, extracting ECG signal features using a convolutional autoencoder, and combining these features with the temporal domain characteristics of the ECG signals to authenticate users using a fully connected neural network.
[0004] This invention is achieved through the following technical solution:
[0005] This invention relates to a user authentication method based on electrocardiogram (ECG) signals from wearable devices. In the offline stage, the raw single-lead ECG signal data is preprocessed and a convolutional autoencoder is trained to extract features. The extracted feature vectors, namely the interval width and height of feature points and user labels, are used as samples to train a fully connected neural network. In the online stage, the trained convolutional autoencoder and fully connected neural network are used for real-time user authentication.
[0006] The method specifically includes:
[0007] The first step is to filter and reduce noise from the raw single-lead signal obtained from the electrodes of the wearable device, and to fit the ECG signal using the least squares method to obtain the trend term. The trend term is then subtracted from the raw signal to eliminate baseline drift.
[0008] The second step is to detect and identify the feature points of the electrocardiogram, extract the starting position of each P wave and the starting position of the next P wave as a complete heartbeat cycle, and perform data preprocessing on each heartbeat cycle.
[0009] The third step involves training a convolutional autoencoder using several uniformly sized electrocardiogram (ECG) signals generated in the second step. This autoencoder is used to extract latent variables from the intermediate layers in order to extract feature vectors from the ECG signals.
[0010] The fourth step involves using the feature vector extracted by the convolutional autoencoder, along with the ECG interval width and height features, as input, and setting the target output as the user's identity label to train the fully connected neural network.
[0011] The fifth step, in the online phase, involves inputting the feature vector extracted by the convolutional autoencoder within a measured heartbeat cycle and the ECG interval features into a fully connected neural network to obtain the user's identity label. Finally, the label generated by the neural network is compared with the preset label to determine the corresponding user's identity.
[0012] This invention relates to a system for implementing the above-mentioned method, comprising: an electrocardiogram (ECG) signal acquisition and preprocessing unit, a waveform detection and feature point localization unit, a time-domain feature extraction unit, a convolutional autoencoder feature extraction unit, a user tag generation unit, and a user authentication unit. Specifically: the waveform detection and feature point localization unit performs noise reduction, ECG signal segmentation, and feature point localization on the raw single-lead ECG signal data acquired by the wearable device, obtaining the denoised signal and signal feature points; the time-domain feature extraction unit calculates the relationship between feature points based on the signal feature points to obtain time-domain features; the convolutional autoencoder feature extraction unit compresses the segmented single-beat ECG signal using a one-dimensional convolutional autoencoder to obtain the frequency-domain features of the ECG signal; the user tag generation unit trains user tag data using medically significant time-domain features and the frequency-domain features extracted by the autoencoder; and the user authentication unit compares the user's tag data, calculates the similarity, and compares it with a threshold to obtain the user authentication result. Technical effect
[0013] This invention performs multimodal analysis on electrocardiogram (ECG) signals to extract feature points, calculates intuitive medical features, and combines these with frequency domain features extracted by a convolutional autoencoder to generate user tags for authentication. The extracted features are more intuitive and interpretable, and exhibit better robustness to high-noise signals collected by wearable devices. Attached Figure Description
[0014] Figure 1 A schematic diagram illustrating the intuitive medical features of existing electrocardiogram readings;
[0015] Figure 2 This is a schematic diagram of the convolutional autoencoder network structure designed for this invention;
[0016] Figure 3 A schematic diagram of an electrocardiogram signal compressed and restored using the convolutional autoencoder of the present invention;
[0017] Figure 4 Flowchart for invention;
[0018] Figure 5 This is a schematic diagram illustrating the effect of an example. Detailed Implementation
[0019] like Figure 1 As shown, a standard electrocardiogram (ECG) signal includes the P wave, QRS complex, and T wave. The P wave represents the potential changes that occur during atrial depolarization; the QRS complex reflects the rapid depolarization of the left and right ventricles and is a group of waves composed of Q, R, and S waves, with the Q and S waves exhibiting downward waveforms and the R wave being a high peak; the T wave represents the potential changes during ventricular repolarization. To increase the accuracy of ECG authentication, this embodiment combines the time-domain characteristics of the ECG signal, namely the width and height of the ECG signal intervals, and the selected features all have certain medical significance. This embodiment uses the differences in the horizontal and vertical coordinates of 10 identified feature points as input to the ECG user authentication system, including the PR interval, QT interval, time difference between the R wave and Q and S waves, time difference between the P wave peak and Q wave, time difference between the T wave peak and Q wave, potential difference between the R wave and Q and S waves, potential difference between the P wave peak and Q wave, and potential difference between the T wave peak and S wave.
[0020] like Figure 4 As shown, this embodiment relates to a user authentication method based on electrocardiogram signals from a wearable device, including:
[0021] Step one involves data preprocessing to transform the raw data collected from the wearable device electrodes into data usable by the ECG signal user authentication system. This specifically includes:
[0022] 1.1) Use a Butterworth filter to filter out noise below 0.5Hz, specifically: Where: n is the order of the filter, ω c The cutoff frequency;
[0023] 1.2) To eliminate 50Hz power frequency interference, a moving average method with a width of 50 is used to remove noise. The same weight is set for the 50 samples before and after to generate a new value at the current position.
[0024] 1.3) The least squares method is used to perform polynomial fitting on the acquired electrocardiogram signal to obtain the trend term. The trend term is then subtracted from the original signal to eliminate baseline drift.
[0025] Step two involves using the Pan-Tompkins algorithm to detect the R-wave peaks of the preprocessed ECG signal obtained in step one, and then locating the PQST wave and its start and end positions using continuous wavelet transform. Specifically, this includes:
[0026] 2.1 The first derivative of the preprocessed signal is calculated and then squared, followed by sliding window integration. The position of the R-wave peak is detected using a threshold.
[0027] 2.2 After performing wavelet multi-scale decomposition on the signal, positive and negative threshold detection is then performed to determine the characteristic points of the PQST wave.
[0028] Step 3: Extract the starting position of each P wave obtained in Step 2 and the starting position of the next P wave as a complete heartbeat cycle, and preprocess the data within each heartbeat cycle.
[0029] The preprocessing includes: performing a 16-length moving average on the extracted ECG signal, then performing linear interpolation on the processed signal to standardize all heartbeat cycles to a length of 500 sampling points, and then normalizing the processed signal to limit the ECG signal value range to 0 to 1.
[0030] The moving average processing described herein may employ, but is not limited to: Where: n is the length of the moving average.
[0031] The interpolation methods used include, but are not limited to: Where: (x0, y0)(x1, y1) represents the two points on the signal that are closest to x.
[0032] The normalization mentioned herein may employ, but is not limited to: Where MAX and MIN are the maximum and minimum values of the current heartbeat cycle, respectively, and the resulting ECG signals of length 500 are used as inputs to the convolutional autoencoder.
[0033] Step four: Using several ECG signals of length 500 that have been preprocessed in step three as a training set, train the convolutional autoencoder and extract the latent variables of the intermediate layers to extract feature vectors from the ECG signals.
[0034] like Figure 2 As shown, the convolutional autoencoder includes an encoder and a decoder, wherein the encoder includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a flattening layer, and a first fully connected layer; the decoder includes a second fully connected layer, an anti-flattening layer, an inverse first convolutional layer, an inverse second convolutional layer, a third fully connected layer, and an output layer.
[0035] This embodiment of the convolutional autoencoder uses ReLU as the activation function and Binary Cross Entropy as the loss function between layers. After training, it can compress and restore the original image, and extract feature vectors from the intermediate layers, such as... Figure 3As shown, the raw signal represents the original electrocardiogram (ECG) signal, and the output signal represents the signal after compression and restoration. The convolutional autoencoder can restore the waveform of the original signal, which means that the autoencoder can compress the ECG signal as a feature after training.
[0036] Step 5: In the online phase, the feature vector extracted by the convolutional autoencoder within a measured heartbeat cycle and the ECG interval features are input into a fully connected neural network to obtain the user's identity label. Finally, the label generated by the neural network is compared with the preset label to determine the corresponding user's identity.
[0037] The fully connected neural network takes 20 features as input, including temporal features and 12 features extracted by a convolutional autoencoder. The intermediate layers are, in order, a 64-node fully connected layer, a 128-node fully connected layer, and a 64-node fully connected layer. Finally, it outputs 5 parameters as user labels and is trained using the Adam optimizer.
[0038] Through specific practical experiments, data sampling and user tag generation were conducted under different body postures, heart rates, and time periods. After repeated testing and verification, the average accuracy was 91.31%, the false acceptance rate was 2.08%, and the rejection rate was 3.32%. The false acceptance and rejection rates under different conditions are shown below. Figure 5 As shown.
[0039] Compared with existing technologies, the convolutional autoencoder used in this invention can obtain the trend characteristics of ECG signals through training. Combined with the medical significance of ECG signal interval width and height, unique biometric features can be extracted for user authentication. The features extracted by this invention are more intuitive and interpretable than those of existing technologies. It also has better robustness to high-noise signals collected by wearable devices and can be combined with wearable devices and other interactive terminals for authentication, thereby improving user data security and enhancing the user experience.
[0040] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.
Claims
1. A user authentication method based on electrocardiogram signals from wearable devices, characterized in that, In the offline phase, the raw single-lead ECG signal data collected is preprocessed and then a convolutional autoencoder is trained to extract features. The extracted feature vectors, namely the interval width and height of the feature points and the user labels, are used as samples to train a fully connected neural network. In the online phase, a trained convolutional autoencoder and a fully connected neural network are used for real-time user authentication, specifically including: The first step is to filter and reduce noise from the raw single-lead signal obtained from the electrodes of the wearable device, and to fit the ECG signal using the least squares method to obtain the trend term. The trend term is then subtracted from the raw signal to eliminate baseline drift. The second step is to detect and identify the feature points of the electrocardiogram, extract the starting position of each P wave and the starting position of the next P wave as a complete heartbeat cycle, and perform data preprocessing on each heartbeat cycle. The third step involves training a convolutional autoencoder using several uniformly sized electrocardiogram (ECG) signals generated in the second step. This autoencoder is used to extract latent variables from the intermediate layers in order to extract feature vectors from the ECG signals. The fourth step involves using the feature vector extracted by the convolutional autoencoder, along with the ECG interval width and height features, as input, and setting the target output as the user's identity label to train the fully connected neural network. The fifth step, in the online phase, involves inputting the feature vector extracted by the convolutional autoencoder within a heartbeat cycle and the ECG interval features into a fully connected neural network to obtain the user's identity label. Finally, the label generated by the neural network is compared with the preset label to determine the corresponding user's identity. The convolutional autoencoder includes an encoder and a decoder, wherein the encoder includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a flattening layer, and a first fully connected layer; the decoder includes a second fully connected layer, an anti-flattening layer, an inverse first convolutional layer, an inverse second convolutional layer, a third fully connected layer, and an output layer. The convolutional autoencoder uses ReLU as the activation function between layers and Binary CrossEntropy as the loss function. After training, it can compress and restore the original image and extract feature vectors from the intermediate layers to restore the waveform of the approximate original signal. The input of the fully connected neural network includes: interphase width and height features and 12 feature vectors extracted by the convolutional autoencoder. The intermediate layers of the fully connected neural network are, in order, a 64-node fully connected layer, a 128-node fully connected layer and a 64-node fully connected layer. Finally, the fully connected neural network outputs 5 parameters as user labels. The fully connected neural network is trained using the Adam optimizer.
2. The user authentication method based on ECG signals from wearable devices according to claim 1, characterized in that, The first step, in particular, includes: 1.1) Use a Butterworth filter to filter out noise below 0.5Hz, specifically: Where: n is the order of the filter, The cutoff frequency; 1.2) To eliminate 50Hz power frequency interference, a moving average method with a width of 50 is used to remove noise. The same weight is set for the 50 samples before and after to generate a new value at the current position. 1.3) The least squares method is used to perform polynomial fitting on the acquired electrocardiogram signal to obtain the trend term. The trend term is then subtracted from the original signal to eliminate baseline drift.
3. The user authentication method based on ECG signals from wearable devices according to claim 1, characterized in that, The method for detecting and identifying feature points in an electrocardiogram (ECG) uses the Pan-Tompkins algorithm to detect the R-wave peaks of the preprocessed ECG signal, and then uses continuous wavelet transform to locate the PQST wave and its start and end positions.
4. The user authentication method based on ECG signals from wearable devices according to claim 1 or 3, characterized in that, The detection and identification of feature points in an electrocardiogram specifically includes: 2.1 The first derivative of the preprocessed signal is calculated, then squared, followed by sliding window integration, and the position of the R-wave peak is detected using a threshold. 2.2 After performing wavelet multi-scale decomposition on the signal, positive and negative threshold detection is then performed to determine the characteristic points of the PQST wave.
5. The user authentication method based on ECG signals from wearable devices according to claim 1, characterized in that, The preprocessing includes: performing a 16-length moving average on the extracted ECG signal, then performing linear interpolation on the processed signal to standardize all heartbeat cycles to a length of 500 sampling points, and then normalizing the processed signal to limit the ECG signal value range to 0 to 1.
6. The user authentication method based on ECG signals from wearable devices according to claim 5, characterized in that, The moving average processing mentioned above refers to: , where: n is the length of the moving average; The interpolation mentioned refers to: Where (x0, y0) and (x1, y1) are the two points on the signal that are closest to x; The normalization referred to is: , where MAX and MIN are the maximum and minimum values of the current heartbeat cycle, respectively, and thus generate several ECG signals of length 500.
7. A system for implementing the ECG signal user authentication method based on wearable devices according to any one of claims 1-6, characterized in that, include: The system comprises an ECG signal acquisition and preprocessing unit, a waveform detection and feature point localization unit, a time-domain feature extraction unit, a convolutional autoencoder feature extraction unit, a user tag generation unit, and a user authentication unit. Specifically: the waveform detection and feature point localization unit performs noise reduction, ECG signal segmentation, and feature point localization on the raw single-lead ECG signal data acquired by the wearable device, obtaining the denoised signal and signal feature points; the time-domain feature extraction unit calculates the relationship between feature points based on the signal feature points, obtaining time-domain features; and the convolutional autoencoder feature extraction unit compresses the segmented single-beat ECG signal using a one-dimensional convolutional autoencoder, obtaining the frequency domain features of the ECG signal. The user tag generation unit uses medically significant time-domain features and frequency-domain features extracted by an autoencoder to train user tag data; the user authentication unit compares the user tag data, calculates the similarity and compares it with a threshold to obtain the user authentication result.