Electrocardio identity authentication method based on feature fusion

Abstract

The invention discloses an electrocardio identity authentication method based on feature fusion. An electrocardio identity recognition model constructed by the method comprises a local feature self-learning module, a global feature self-learning module, a feature fusion module and a classifier. A local feature self-learning module and a global feature self-learning module are adopted to extract local features and global features of electrocardiosignals, individual features are captured through the local sensing ability of the local feature self-learning module, and overall fluctuation of ECG is analyzed through the global association ability of the global feature self-learning module. The precision and robustness of identity authentication are improved; meanwhile, the electrocardiosignals are preprocessed through filtering, cardiac beat segmentation and data amplification, the processing capacity for unknown data is enhanced, the robustness and generalization capacity of the model are improved, the signal quality is improved, the diversification condition in reality is effectively simulated, and follow-up analysis is more accurate.

Description

Technical Field

[0001] The present invention belongs to the technical field of identity recognition, and in particular relates to an electrocardiogram identity authentication method based on feature fusion. Background Art

[0002] With the development of generative artificial intelligence, static biometric features such as fingerprints, faces, and irises are facing the threat of counterfeiting, and specific deception and attack cases against fingerprint and face recognition are emerging in an endless stream. In comparison, dynamic biometric features (such as electrocardiogram signals) are more expensive to counterfeit. Electrocardiogram (ECG) signals are one of the important bioelectric signals in the human body, reflecting the internal characteristic information and working mechanism of the heart from different levels. The waveform characteristics of ECG signals have the uniqueness required for identity authentication. Compared with other biometrics, ECG identity authentication (ECG Authentication) technology has the high anti-counterfeiting characteristics of "live" collection and authentication. It collects the user's ECG signals while performing specific physiological activities (such as deep breathing) to confirm their liveness, effectively resisting external attacks. In addition, ECG signals are affected by individual physiological states and heart activities, and emotions and other factors will also increase the difficulty of simulation and counterfeiting. ECG-based identity authentication is expected to become the next generation of identity authentication. However, there are currently technical challenges, which are mainly reflected in the following aspects:

[0003] The first challenge comes from the reliance of ECG recognition on high-quality data. Although the accuracy of some ECG authentication based on public databases is close to 100%, in actual applications, ECG signals are easily affected by various interferences such as movement, electromyography, breathing, emotions, and time, resulting in unstable accuracy of this authentication method. In addition, the quality of ECG signals collected by different acquisition methods is different, so a preprocessing method is needed to improve the signal-to-noise ratio of the signal and reduce the impact of practical factors.

[0004] The second challenge comes from the stability of the ECG recognition algorithm. First, traditional feature extraction methods perform poorly in the face of complex signal variability. ECG signals are affected by many factors, such as individual physiological differences and environmental noise, which leads to reduced stability and accuracy of the algorithm under different conditions. Second, many existing models lack effective global context modeling capabilities. The timing characteristics of ECG signals make the capture of local features insufficient to provide sufficient information for reliable identity authentication. This requires a network architecture that can integrate local and global information to improve the accuracy of identity recognition.

[0005] In view of this, a feature fusion-based ECG identity recognition method is provided, which realizes stable and accurate recognition of individual identity by fusing the local and global features of ECG signals, and has broad practical application value. Summary of the invention

[0006] The purpose of the present invention is to provide an electrocardiogram identity authentication method based on feature fusion.

[0007] The present invention provides an electrocardiogram identity authentication method based on feature fusion, which comprises the following steps: Step 1: Collect the ECG signals of the subjects at different time periods, and label the ECG signals of different subjects accordingly; Step 2: preprocessing the subject's ECG signal as a data set; Step 3, construct an ECG identity recognition model; the ECG identity recognition model includes a local feature self-learning module, a global feature self-learning module, a feature fusion module and a classifier; the local feature self-learning module includes an input layer, an output layer and five sequentially connected convolution blocks; the input layer includes a sequentially connected one-dimensional convolution layer, a batch normalization layer and an activation layer for preliminary feature extraction; the output layer includes a sequentially connected flattening layer, a fully connected layer, a batch normalization layer and an activation layer for outputting local features; the five convolution blocks have the same structure, including a sequentially connected one-dimensional convolution layer, a batch normalization layer, an activation layer and a maximum pooling layer; The global feature self-learning module includes a position encoding layer, a multi-head self-attention layer and a feedforward neural network layer connected in sequence, as well as a layer normalization, a flattening layer, a fully connected layer and an activation layer; the position encoding layer is used to introduce position information into the ECG signal; the multi-head self-attention layer is used to extract features; the feedforward neural network is used to further extract features; the layer normalization is used to receive and fuse the ECG signal with introduced position information and the features extracted by the feedforward neural network; the flattening layer, the fully connected layer and the activation layer process the layer normalization output features in sequence to output global features; The feature fusion module is used to fuse local features and global features; the classifier is used to receive the fused features and obtain classification results; Step 4: Use the data set to train the ECG identity recognition model; Step 5: Collect the subject's ECG signal, pre-process it, and input it into the trained ECG identity recognition model to identify the subject.

[0008] Preferably, in the step three, the sizes of the convolution kernels in the five convolution blocks are gradually increased to learn abstract features at different levels.

[0009] Preferably, in step 3, sine and cosine functions are used to generate the position code, and the position code is calculated as follows: in, is the encoding value of the even index position in the position encoding corresponding to sequence n; is the encoding value of the odd index position in the position encoding corresponding to sequence n; is the embedding dimension; is the dimension index; This is the ECG signal after the position information is introduced.

[0010] Preferably, in step 2, the preprocessing method is: filtering, heart beat segmentation and data amplification processing are performed on the ECG signals in the data set respectively.

[0011] Preferably, the filtering process is as follows: The subject's ECG signal is processed by a zero-phase bidirectional filter; the intermediate result is obtained by forward filtering y[n] The expression is: in, x[n] is the original ECG signal of the subject; a k and b k is the coefficient of the filter; k=0,1,...,M; 𝑀,N is the order of the filter, that is, the input ECG signal of the filter , The maximum number of lags; k is the time offset or lag steps of the ECG signal; n is the sequence of the ECG signal; The final result obtained by reverse filtering z[n] The expression is: The final result obtained by applying a finite impulse response filter to a zero-phase two-way filter z[n] Perform filtering and obtain the expression of the filtered output ECG signal r[n]: in, h[k] is the impulse response, and its expression is: in, w[k] is the Hanning window function; and are the high cut-off frequency and low cut-off frequency of the band-pass filter respectively; sinc is the sampling function.

[0012] Preferably, the process of heart beat segmentation is: The ECG signal obtained after filtering r[n] Perform power line interference suppression and myoelectric noise suppression; apply an adaptive threshold on the characteristic curve and check whether the threshold is exceeded at each sampling point to detect the R wave peak position; obtain the ECG signal after heart beat segmentation based on the R wave peak position and the length of the sampling points before and after the R wave peak .

[0013] Preferably, the process of data amplification is as follows: ECG signal Noise, shift and resample are performed respectively; ECG signal after noise The expression is: N [n]=E n +0.05×N 0,1 in, is a normal distribution with a mean of 0 and a standard deviation of 1; ECG signal after translation The expression is: Among them, fs is the sampling rate of the data set; is the floor symbol; Resampled ECG signal The expression is: The ECG signal obtained after noise addition, translation and resampling is compared with the ECG signal Compose an enhanced dataset and use the enhanced dataset to update the original dataset.

[0014] Preferably, in step 1, the ECG signals of the subject in different activity states and psychological states are collected respectively.

[0015] Preferably, in the step three, the feature fusion module fuses local features and global features by weighted superposition.

[0016] Preferably, in step 4, the training process of the ECG identity recognition model is guided by cross entropy loss.

[0017] The present invention has the following beneficial effects: 1. The present invention adopts a local feature self-learning module and a global feature self-learning module to extract local features and global features of electrocardiogram signals respectively. The local perception ability of the local feature self-learning module is used to capture individual features, and the global correlation ability of the global feature self-learning module is used to analyze the overall fluctuation of ECG, thereby improving the accuracy and robustness of identity authentication.

[0018] 2. The present invention pre-processes the ECG signal through filtering, heart beat segmentation and data amplification, thereby enhancing the processing capability of unknown data, improving the robustness and generalization ability of the model, improving the quality of the signal, and effectively simulating the diverse situations in reality, making subsequent analysis more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is the overall flow chart of the present invention.

[0020] Figure 2 It is a schematic diagram of the central electrical identity recognition model of the present invention. DETAILED DESCRIPTION

[0021] The present invention will be further described below in conjunction with the accompanying drawings.

[0022] like Figure 1 As shown, an ECG identity authentication method based on feature fusion includes the following steps: Step 1: The fingertip ECG acquisition system is used to collect ECG signals of different subjects in different activity states and psychological states. When measuring the fingertip ECG, the subjects only need to place the fingertips of both hands on the dry electrodes to ensure that the fingertips are in full contact with the dry electrodes for 10 seconds to complete the fingertip ECG acquisition. The ECG signals obtained by fingertip ECG acquisition of subjects at different time periods are used as data sets.

[0023] Step 2: Preprocess the subject’s ECG signal 2-1. Filtering The subject's ECG signal is processed through a zero-phase two-way filter to eliminate the noise in the subject's ECG signal without introducing phase delay. A zero-phase two-way filter is a filter used for signal processing. Its working principle is to first use a filter (such as an FIR filter) to forward filter the signal, and then reverse filter the filtered signal. Since the reverse filtering process is the reverse operation of the original signal, the phase delay introduced by filtering can be offset, thereby achieving zero-phase filtering. Intermediate results obtained by forward filtering y[n] The expression is:

[0024] in, x[n] is the original ECG signal of the subject; a k and b k is the coefficient of the filter; k=0,1,...,M; 𝑀,N is the order of the filter, that is, the input ECG signal of the filter , The maximum number of lags; k is the time offset or lag steps of the ECG signal; n is the sequence of the ECG signal.

[0025] The final result obtained by reverse filtering z[n] The expression is: The final result obtained by applying a finite impulse response (FIR) filter to a zero-phase two-way filter z[n] Perform filtering to obtain the output ECG signal r[n] The expression is: in, h[k] is the impulse response, and its expression is: in, w[k] is the Hanning window function; and are the high cut-off frequency and low cut-off frequency of the band-pass filter respectively; sinc is the sampling function.

[0026] Hanning Window Function w[k] The expression is: Through finite impulse response filter processing, it is possible to filter out signals in the range of 0.7 Hz to 90 Hz, while suppressing power supply noise near 50 Hz. The combination of zero-phase bidirectional filter and finite impulse response filter provides an effective filtering solution for signal processing, which can simultaneously eliminate phase distortion and smooth signals, thereby improving the overall quality of the signal and the accuracy of analysis. This method has a wide range of application value in medical signal processing, audio signal processing and other fields.

[0027] 2-2. Heart beat segmentation The ECG signal obtained after filtering r[n] Power line interference suppression and myoelectric noise suppression are performed, and then an adaptive threshold (a dynamic threshold formed by the historical maximum value, short-term maximum value, and decreasing slope) is applied to the characteristic curve, and the R wave peak position is detected by checking whether the threshold is exceeded at each sampling point. Considering that the average heartbeat length is 0.8s, the single segment length is fixed to 0.8fs, and the ratio of sampling points before and after the R peak of each heartbeat is 0.32:0.48. On this basis, the first and last heartbeats are discarded, and the remaining heartbeats are used as the ECG signal after heartbeat segmentation. This process effectively improves the analysis accuracy of the ECG and is suitable for real-time monitoring applications.

[0028] 2-3. Data augmentation In order to adapt the model to the real environment, the ECG signal is amplified. The signal has the characteristics of noise, pitch domain and time domain. Data amplification will focus on these characteristics to simulate the changes in real-world conditions and amplify the ECG signal. Noise, shift and resample are performed respectively. ECG signal after noise addition The expression is:

[0029] N [n]=E n +0.05×N 0,1 # 6 in, is a normal distribution with a mean of 0 and a standard deviation of 1.

[0030] ECG signal after translation The expression is: Among them, fs is the sampling rate of the data set; The floor symbol.

[0031] Resampled ECG signal The expression is: The ECG signal obtained after noise addition, translation and resampling is compared with the ECG signal Compose an enhanced data set to enhance the ability to process unknown data and improve the robustness and generalization ability of the model.

[0032] Step 3: Build an ECG identity recognition model like Figure 2 As shown in the figure, the ECG identity recognition model includes a local feature self-learning module, a global feature self-learning module, a feature fusion module and a classifier; since some key individual physiological information of the one-dimensional ECG signal is hidden in the local band, local features are extracted through the local feature self-learning module; the local feature self-learning module includes an input layer, an output layer and five sequentially connected convolution blocks (Block); the input layer includes a sequentially connected one-dimensional convolution layer (Conv1d), a batch normalization layer (BN) and an activation layer (ReLU activation function) for preliminary feature extraction; the output layer includes a sequentially connected flattening layer (Flatten), a fully connected layer (Dense), a batch normalization layer and an activation layer; the structures of the five convolution blocks are the same, all of which include sequentially connected one-dimensional convolution layers, batch normalization layers, activation layers and maximum pooling layers (MaxPooling); the expression of the ECG signal Y output by each convolution block is: Among them, X is the ECG signal input to the convolution block; W is the convolution kernel matrix; b is the bias.

[0033] The size of the convolution kernel in the five convolution blocks gradually increases, allowing the network to learn abstract features at different levels. Preliminary feature extraction is performed through the input layer, and then higher-level abstract features are gradually learned by stacking five convolution blocks, which helps to stabilize training and improve computational efficiency. The specific structure of the local feature self-learning module is shown in Table 1.

[0034] Table 1 Network structure of local feature self-learning module based on multi-level enhancement The global feature self-learning module can automatically learn important time dependencies and feature patterns from the ECG signals in the enhanced data set, so that the signal data at each moment can focus on other parts of the entire signal, capture subtle differences in the ECG signal, and identify correlations, trends and periodic features in the signal, thereby more accurately distinguishing individual differences in identity authentication.

[0035] The global feature self-learning module includes a position encoding layer, a multi-head self-attention layer, and a feedforward neural network layer connected in sequence; the position encoding layer is used to introduce position information into the ECG signal after input data amplification, and uses sine and cosine functions to generate position encoding so that the feature of each position has unique position information. The calculation formula of the position encoding is as follows: in, is the encoding value of the even index position in the position encoding corresponding to sequence n; is the encoding value of the odd index position in the position encoding corresponding to sequence n; is the embedding dimension; is the dimension index; This is the ECG signal after the position information is introduced.

[0036] The multi-head self-attention layer is used for feature extraction. The multi-head self-attention mechanism can simultaneously focus on the features of different positions in the ECG signal, capture the correlations in different time domains and frequency domains, and better understand the structure of the ECG signal. In addition, it can also establish long-distance dependencies, which is crucial for the current registration and verification of a period of time. The calculation of the multi-head self-attention mechanism is as follows:

[0037] Among them, Q is the query vector, K is the key vector, and V is the value vector, which is calculated by the input signal X and the learnable weight matrix; is the dimension of the key; softmax is the normalized exponential function.

[0038] The feedforward neural network is used to further extract features and independently transform the features of each sequence position of the ECG signal to improve the nonlinear expression ability of the model. The feedforward neural network independently applies a fully connected layer to each position to enrich the representation of the features at each position. The calculation method is as follows:

[0039] Where x is the input ECG signal; and is the weight of the feedforward neural network; and is the bias of the feedforward neural network.

[0040] The ECG signal is normalized by layer The features extracted by the multi-head self-attention layer and the feedforward neural network are added, and the global features are output in sequence through the flattening layer, the fully connected layer and the activation layer. This enables the global feature self-learning module to not only learn new abstract features but also retain the original signal information. It can solve the problems of gradient vanishing and gradient exploding in deep neural network training and improve training stability.

[0041] The feature fusion module includes a batch normalization layer, a ReLU activation function, and a random dropout layer (Dropout). The batch normalization layer fuses local features and global features in a weighted superposition manner, and adds the weight coefficient to the training process to achieve self-learning of the weight coefficient. The expression of the fusion output C is: in, The local features extracted by the local feature self-learning module; The global features extracted by the global feature self-learning module; is the weight coefficient.

[0042] The classifier (SoftMax) obtains the classification result based on the output of the feature fusion module.

[0043] Step 4: Use the enhanced dataset to train the ECG identity recognition model, and use the cross entropy loss to guide the training process of the ECG identity recognition model.

[0044] Step 5: Collect the subject's ECG signal, pre-process it, and input it into the trained ECG identity recognition model to identify the subject. If the subject's ECG signal segment recognition success rate is higher than 96%, the tester is successfully authenticated, otherwise the authentication fails.

[0045] Step 6: Model Evaluation 6-1. The ECG authentication mainstream databases ECGID, MIT-ARR and MIT-NSR were used to conduct ablation experiments on the ECG authentication method of the present invention; the accuracy, F1 score and equal error rate were used to evaluate the model. The accuracy is the average accuracy of the test set. Each individual adopts a voting mechanism. When more than 96% of the heartbeats are successfully identified, the individual is verified. The accuracy represents the number of individuals / total number of people verified by the database. The equal error rate is the value when the false recognition rate and the rejection rate are equal, indicating the performance balance of the system between identifying legitimate users and rejecting illegal users. The ablation experiment results are shown in Table 2.

[0046] Table 2 Ablation experiment results It can be seen from Table 2 that the effect of ECG identity authentication is poor when data augmentation is not performed, and the effect of using local feature self-learning module and global feature self-learning module to extract features for ECG identity authentication is also far behind the present invention.

[0047] 6-2. The present invention and other mainstream models are used to perform ECG identity authentication on ECG-ID, MIT-ARR, MIT-NSR and CYBHi databases. The experimental results are shown in Table 3.

[0048] Table 3 Comparison results of ECG identity authentication using different models It can be seen from Table 3 that the accuracy of the present invention in ECG identity authentication is better than that of the other three models, and it can achieve a high level on different databases.

Claims

1. An ECG identity authentication method based on feature fusion, characterized in that: The following steps are involved: Step 1: Collect the ECG signals of the subjects at different time periods, and label the ECG signals of different subjects accordingly; Step 2: preprocessing the subject's ECG signal as a data set; Step 3, construct an ECG identity recognition model; the ECG identity recognition model includes a local feature self-learning module, a global feature self-learning module, a feature fusion module and a classifier; the local feature self-learning module includes an input layer, an output layer and five sequentially connected convolution blocks; the input layer includes a sequentially connected one-dimensional convolution layer, a batch normalization layer and an activation layer for preliminary feature extraction; the output layer includes a sequentially connected flattening layer, a fully connected layer, a batch normalization layer and an activation layer for outputting local features; The five convolutional blocks have the same structure, which includes a one-dimensional convolutional layer, a batch normalization layer, an activation layer, and a maximum pooling layer connected in sequence; The global feature self-learning module includes a position encoding layer, a multi-head self-attention layer and a feedforward neural network layer connected in sequence, as well as a layer normalization, a flattening layer, a fully connected layer and an activation layer; the position encoding layer is used to introduce position information into the electrocardiogram signal; The multi-head self-attention layer is used for feature extraction; the feedforward neural network is used to further extract features; the layer normalization is used to receive and fuse the ECG signal with the introduced position information and the features extracted by the feedforward neural network; the flattening layer, the fully connected layer and the activation layer process the layer normalization output features in turn and output the global features; The feature fusion module includes a batch normalization layer, an activation layer and a random inactivation layer, which are used to fuse local features and global features; The classifier is used to receive the fused features and obtain the classification results; Step 4: Use the data set to train the ECG identity recognition model; Step 5: Collect the subject's ECG signal, pre-process it, and input it into the trained ECG identity recognition model to identify the subject.

2. The ECG identity authentication method based on feature fusion according to claim 1 is characterized in that: In the step 3, the sizes of the convolution kernels in the five convolution blocks are gradually increased to learn abstract features at different levels.

3. The ECG identity authentication method based on feature fusion according to claim 1 is characterized in that: In step 3, sine and cosine functions are used to generate position codes, and the calculation formula of the position codes is as follows: in, is the encoding value of the even index position in the position encoding corresponding to sequence n; is the encoding value of the odd index position in the position encoding corresponding to sequence n; is the embedding dimension; is the dimension index; This is the ECG signal after the position information is introduced.

4. The ECG identity authentication method based on feature fusion according to claim 1, characterized in that: In the step 2, the preprocessing method is: filtering, heart beat segmentation and data amplification processing are performed on the ECG signals in the data set respectively.

5. The ECG identity authentication method based on feature fusion according to claim 4 is characterized in that: The filtering process is as follows: The subject's ECG signal is processed by a zero-phase bidirectional filter; the intermediate result is obtained by forward filtering The expression is: in, is the original ECG signal of the subject; a k and b k is the coefficient of the filter; k=0,1,...,M; 𝑀,N is the order of the filter, that is, the input ECG signal of the filter , The maximum number of lags; k is the time offset or lag steps of the ECG signal; n is the sequence of the ECG signal; The final result obtained by reverse filtering The expression is: The final result obtained by applying a finite impulse response filter to a zero-phase two-way filter Perform filtering and obtain the expression of the filtered output ECG signal r[n]: in, is the impulse response, and its expression is: in, is the Hanning window function; and are the high cut-off frequency and low cut-off frequency of the band-pass filter respectively; sinc is the sampling function.

6. The ECG identity authentication method based on feature fusion according to claim 5, characterized in that: The process of heart beat segmentation is as follows: The ECG signal obtained after filtering Perform power line interference suppression and myoelectric noise suppression; An adaptive threshold is applied to the characteristic curve, and at each sampling point, a check is made to see whether the threshold is exceeded to detect the R wave peak position; the ECG signal after heart beat segmentation is obtained based on the R wave peak position and the length of the sampling points before and after the R wave peak .

7. The ECG identity authentication method based on feature fusion according to claim 6, characterized in that: The process of data augmentation is as follows: ECG signal Noise, shift and resample are performed respectively; ECG signal after noise The expression is: in, is a normal distribution with a mean of 0 and a standard deviation of 1; ECG signal after translation The expression is: Where, fs is the sampling rate of the data set; is the floor symbol; Resampled ECG signal The expression is: The ECG signal obtained after noise addition, translation and resampling is compared with the ECG signal Compose an enhanced dataset and use the enhanced dataset to update the original dataset.

8. The ECG identity authentication method based on feature fusion according to claim 1, characterized in that: In the step 1, the electrocardiogram signals of the subjects in different activity states and psychological states are collected respectively.

9. The ECG identity authentication method based on feature fusion according to claim 1, characterized in that: In the step three, the feature fusion module fuses local features and global features by weighted superposition.

10. The ECG identity authentication method based on feature fusion according to claim 1, characterized in that: In the step 4, the cross entropy loss is used to guide the training process of the ECG identity recognition model.

Images