An identity authentication method based on mixed physiological state electrocardiosignal

Abstract

The application discloses an identity authentication method based on a mixed physiological state electrocardiosignal, relates to the technical field of biological feature recognition, and aims to solve the problem that the identity recognition research on the electrocardiosignal in the prior art is mainly concentrated in the resting state and the research on the identity recognition before and after exercise is less. The method comprises the following steps: acquiring mixed electrocardiosignal data of a to-be-tested person; pre-processing the mixed electrocardiosignal data to obtain pre-processed mixed electrocardiosignal data; taking the pre-processed mixed electrocardiosignal data as input, outputting an embedding feature vector based on an electrocardiosignal extraction network model; calculating a cosine similarity value according to the embedding feature vector, and authenticating the identity of the to-be-tested person according to the cosine similarity value. The method can simultaneously adapt to and process the electrocardiosignals in the exercise state and the resting state, and improves the accuracy and robustness of identity authentication.

Description

TECHNICAL FIELD

[0001] The present application relates to an identity authentication method based on a mixed physiological state electrocardiosignal, belonging to the technical field of biometric recognition. BACKGROUND

[0002] Biometric recognition is an important means of identity authentication in today's society. The mainstream biometric recognition includes fingerprint recognition, palmprint recognition, face recognition, iris recognition, etc. However, these traditional recognition methods have certain limitations: fingerprint and palmprint recognition are easily affected by surface conditions and have the risk of being copied; face recognition has reduced recognition accuracy under factors such as light changes and age growth; iris recognition has high accuracy but requires professional acquisition equipment and is costly.

[0003] As one of the physiological characteristics of the human body, electrocardiosignal has its unique advantages. Since the location, size and structure of everyone's heart are different, electrocardiosignal has significant individual differences. At the same time, electrocardiosignal is a biological electric signal actively generated by the human body and has the characteristics of living body, which is difficult to be forged. However, the current electrocardiosignal-based identity authentication technology still has technical bottlenecks. Electrocardiosignal is easily affected by external factors such as emotions and movements, especially the electrocardiosignal under the movement state changes significantly compared with the resting state, which leads to a decrease in recognition accuracy.

[0004] In recent years, with the development of deep learning technology, electrocardiosignal identity recognition methods based on deep neural networks have made significant progress. Existing methods mainly include feature extraction based on convolutional neural network (CNN) and time series modeling based on recurrent neural network (RNN). However, these methods still have the following problems when dealing with electrocardiosignal under the movement state: first, the traditional single-scale convolutional structure is difficult to capture both local detailed features and long-range dependencies of electrocardiosignal; second, there is a lack of effective feature selection mechanism, which cannot highlight the contribution of important feature channels; third, the training strategy does not fully consider the distribution difference of samples under the movement and resting states, leading to inconsistent performance of the model under different states.

[0005] Currently, the research on electrocardiosignal identity recognition at home and abroad mainly focuses on the resting state, and the research on identity recognition before and after movement is relatively less. Movement will cause problems such as heart rate changes and electrocardiosignal waveform distortion, making the recognition algorithm developed based on the resting state perform poorly under the movement state. At the same time, there is also a lack of effective methods to deal with the differences between electrocardiosignal under the movement state and the resting state, affecting the practicality of the authentication system. Therefore, how to propose an electrocardiosignal identity authentication method that can adapt to both the movement state and the resting state, and improve the accuracy and robustness of authentication, is a problem that needs to be solved at present. SUMMARY

[0006] The present application aims at overcoming the deficiencies in the prior art, and provides an identity authentication method based on mixed physiological state electrocardiosignal, which can simultaneously adapt to and process electrocardiosignals in motion state and resting state, and improve the accuracy and robustness of identity authentication.

[0007] To achieve the above-mentioned purpose, the present application is implemented by using the following technical scheme:

[0008] The present application provides an identity authentication method based on mixed physiological state electrocardiosignal, comprising:

[0009] Obtaining mixed electrocardiosignal data of a to-be-tested person, wherein the mixed electrocardiosignal data is obtained by mixing electrocardiosignals of the to-be-tested person in motion state and resting state;

[0010] Preprocessing the mixed electrocardiosignal data to obtain preprocessed mixed electrocardiosignal data;

[0011] Taking the preprocessed mixed electrocardiosignal data as input, outputting an embedding feature vector based on an electrocardiosignal extraction network model;

[0012] Calculating a cosine similarity value according to the embedding feature vector, and authenticating the identity of the to-be-tested person according to the cosine similarity value.

[0013] Further, the preprocessing of the motion-resting state mixed electrocardiosignal data to obtain preprocessed motion-resting state mixed electrocardiosignal data comprises:

[0014] Filtering the motion-resting state mixed electrocardiosignal by using a band-pass filter, wherein the band-pass filter is a fourth-order Butterworth band-pass filter, and the frequency band range is set to 0.5-40Hz;

[0015] Eliminating the residual baseline drift in the motion-resting state mixed electrocardiosignal by using a high-pass filter to obtain filtered mixed electrocardiosignal, wherein the high-pass filter is a first-order filter, and the cutoff frequency is set to 0.5Hz;

[0016] Standardizing the filtered mixed electrocardiosignal to obtain processed mixed electrocardiosignal, and the expression of the standardization is:

[0017]

[0018] wherein, represents the filtered mixed electrocardiosignal, represents the signal mean of the filtered mixed electrocardiosignal, represents the signal standard deviation of the filtered mixed electrocardiosignal, represents a stability coefficient, a processed mixed electrocardiosignal;

[0019] segmenting and quality controlling the processed mixed electrocardiosignal to obtain preprocessed mixed exercise-rest electrocardiosignal data.

[0020] Further, the segmenting and quality controlling the processed mixed electrocardiosignal to obtain preprocessed mixed exercise-rest electrocardiosignal data comprises:

[0021] detecting the position of R wave in the preprocessed mixed electrocardiosignal by using Hamilton algorithm, and calculating the time interval of the interval of adjacent R waves;

[0022] segmenting the window at the position of R wave to obtain segmented signal, the window is centered on R wave, 25% of the window length is cut forward, the remaining part is cut backward, and the window does not exceed the effective range of the signal, wherein the window length is set according to the signal state;

[0023] detecting the quality control of the segmented signal, and screening to obtain the segmented signal meeting the quality control condition, i.e. as preprocessed mixed exercise-rest electrocardiosignal data;

[0024] The quality control condition includes that there is no non-numeric value and infinite value, the difference range of the maximum value and the minimum value of the segmented signal is 1.0-15.0, and at least one valid R wave is contained.

[0025] Further, the electrocardiosignal extraction network model comprises a multi-scale convolution module, a feature enhancement module, a deep feature extraction module, a global context module and a self-attention module connected in sequence, and the output end of the self-attention module is connected with a global average pooling layer, a full connection layer and a normalization layer in sequence.

[0026] The multi-scale convolution module is used for feature extraction of the preprocessed mixed electrocardiosignal data to obtain a feature tensor, and the feature tensor is input to the feature enhancement module after feature fusion by channel dimension splicing.

[0027] The feature enhancement module is used for feature channel adaptive adjustment of the input feature tensor.

[0028] The deep feature extraction module is used for deep feature extraction of the input feature tensor.

[0029] The global context module is used for global information aggregation of the input feature to obtain channel attention weight, and the channel attention weight is multiplied with the input feature tensor and input to the self-attention module.

[0030] The self-attention module is used for establishing long-range dependency relationship between the input feature tensors.

[0031] Further, the feature enhancement module comprises an adaptive average pooling layer, a first fully connected layer, a second fully connected layer and an output layer connected in sequence;

[0032] The adaptive average pooling layer is configured to compress the input feature tensor into a one-dimensional feature vector;

[0033] The first fully connected layer is configured to compress the channel dimension of the one-dimensional feature vector to 1 / 16 of the original dimension and pass it through a ReLU activation function;

[0034] The second fully connected layer is configured to restore the channel dimension of the one-dimensional feature vector to the original dimension and pass it through a Sigmoid activation function to obtain the attention weight coefficients of each channel;

[0035] The output layer is configured to multiply the weight coefficients of each channel with the input feature tensor channel by channel and then output.

[0036] Further, the multi-scale convolution module comprises four parallel convolution branches, the convolution kernel sizes of the convolution branches are 3, 5, 7 and 11 respectively, and the input channel number of each convolution branch is 1 and the output channel number is 64;

[0037] The deep feature extraction module comprises two convolution units connected in sequence, each of the convolution units comprises a convolution layer, a batch normalization layer and a ReLU activation function, wherein the first convolution unit maintains the channel number unchanged, and the second convolution unit increases the channel number to 512;

[0038] The global context module comprises an adaptive average pooling layer, two 1x1 convolution layers and an activation function connected in sequence, wherein the first 1x1 convolution layer reduces the channel number to 256 and passes it through a ReLU activation function, and the second 1x1 convolution layer restores the channel number to 512 and passes it through a Sigmoid activation function to obtain the channel attention weight;

[0039] The self-attention module comprises a Query branch, a Key branch and a Value branch, the channel number of the Query branch and the Key branch is reduced to 1 / 8 of the input feature tensor through a 1x1 convolution layer, the channel number of the Value branch is maintained through a 1x1 convolution layer, the product of the Query branch and the Key branch is processed through a Softmax activation function to obtain an attention score, and the attention feature is calculated with the Value branch to obtain an enhanced feature representation through a learnable scaling parameter and a residual connection.

[0040] Further, an output end of the normalization layer is further connected with a classifier, and the classifier is used for mapping features to a classification space for calculation of a classification loss in a pre-training process of the electrocardiosignal extraction network model.

[0041] Further, the electrocardiosignal extraction network model is pre-trained, and the pre-training method comprises:

[0042] An exercise-rest state mixed electrocardiosignal dataset is acquired.

[0043] The exercise-rest state mixed electrocardiosignal dataset is divided into a training set, a verification set and a test set.

[0044] The training set data is taken as input to train the electrocardiosignal extraction network model, in the training process, an Adam optimizer is used for parameter optimization, a learning rate is adjusted, and a loss function is calculated, then the test set data is taken as input to test the electrocardiosignal extraction network model, and the training is iterated until the loss function in the test is not reduced for 10 rounds in succession to stop the training, and a trained electrocardiosignal extraction network model is obtained.

[0045] A preset decision threshold is set, the verification set data is input into the trained electrocardiosignal extraction network model, an error acceptance rate and an error rejection rate are calculated, the error acceptance rate is a probability that an unauthorized person is wrongly identified as an authorized person, the error rejection rate is a probability that an authorized person is wrongly identified as an unauthorized person, and the decision threshold corresponding to the case that the error acceptance rate is equal to the error rejection rate is the optimal decision threshold;

[0046] The decision threshold is set as the optimal decision threshold to obtain a pre-trained electrocardiosignal extraction network model.

[0047] Further, the loss function comprises a sum of a focal loss function and a multiple similarity loss function, an expression of the focal loss function is:

[0048]

[0049] wherein, is a class weight, is a probability of a correct class predicted, is a focal factor, is a cross-entropy loss;

[0050] The multiple similarity loss function comprises a similarity loss of a positive sample pair and a similarity loss of a negative sample pair, and an expression of the multiple similarity loss function is:

[0051]

[0052]

[0053]

[0054] wherein, is a multiple similarity loss function, is a similarity loss of a positive sample pair, is a similarity loss of a negative sample pair, is a positive sample pair similarity, is a negative sample pair similarity, is a positive sample scaling factor, is a negative sample scaling factor, thresh is a similarity threshold, and clamp represents limiting the input value in a preset range.

[0055] Further, the method comprises:

[0056] calculating a cosine similarity value between the embedding feature vectors, and

[0057]

[0058] wherein, and both represent embedding feature vectors, and n represents the dimension of the embedding feature vectors, represents a cosine similarity value;

[0059] authenticating the identity of the subject according to the cosine similarity value between the embedding feature vectors.

[0060] Compared with the prior art, the present application has the following beneficial effects:

[0061] The present application trains the electrocardiogram data of the mixed motion state and the resting state, realizes the unified processing of the electrocardiogram signals in different states, realizes good recognition performance of the electrocardiogram signals in the motion state and the resting state, and has strong practical value.

[0062] The combined loss function training strategy provided by the application solves the sample class imbalance problem through the focal loss, optimizes the feature space distribution by using the multi-similarity loss, and effectively improves the training effect and generalization ability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 A flowchart of an embodiment of the identity authentication method based on the mixed physiological state electrocardiosignal of the application is shown.

[0064] Figure 2 A flowchart of pre-training and implementation of an embodiment of the identity authentication method based on the mixed physiological state electrocardiosignal of the application is shown.

[0065] Figure 3 A structural diagram of an electrocardiosignal extraction network model of an embodiment of the identity authentication method based on the mixed physiological state electrocardiosignal of the application is shown.

[0066] Figure 4 A structural diagram of a feature enhancement module of the electrocardiosignal extraction network model of an embodiment of the identity authentication method based on the mixed physiological state electrocardiosignal of the application is shown.

[0067] Figure 5 A structural diagram of a deep feature extraction module of the electrocardiosignal extraction network model of an embodiment of the identity authentication method based on the mixed physiological state electrocardiosignal of the application is shown.

[0068] Figure 6 A structural diagram of a global context module of the electrocardiosignal extraction network model of an embodiment of the identity authentication method based on the mixed physiological state electrocardiosignal of the application is shown.

[0069] Figure 7 A structural diagram of a self-attention module of the electrocardiosignal extraction network model of an embodiment of the identity authentication method based on the mixed physiological state electrocardiosignal of the application is shown. DETAILED DESCRIPTION

[0070] The application will be further described below in conjunction with the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the application, and cannot be used to limit the protection scope of the application.

[0071] As shown in Figure 1 and Figure 2 , the embodiment of the application provides an identity authentication method based on a mixed physiological state electrocardiosignal, which includes two stages, i.e., an electrocardio registration stage and an electrocardio authentication stage, wherein the electrocardio registration stage is used to train an electrocardiosignal extraction network model, and the electrocardio authentication stage is used to perform identity authentication on a to-be-tested person by using the pre-trained electrocardiosignal extraction network model.

[0072] First, the electrocardio registration stage specifically includes:

[0073] The electrocardiosignal extraction network model is constructed, as shown in Figure 3As shown, the electrocardiosignal extraction network model comprises a multi-scale convolution module, a feature enhancement module, a deep feature extraction module, a global context module and a self-attention module connected in sequence, and the output end of the self-attention module is connected with a global average pooling layer, a full connection layer and a normalization layer in sequence.

[0074] The multi-scale convolution module comprises four parallel convolution branches, the convolution kernel sizes of the convolution branches are 3, 5, 7 and 11 respectively, and the input channel number of each convolution branch is 1 and the output channel number is 64. In the electrocardiosignal, the QRS complex is a characteristic waveform representing ventricular depolarization, which is composed of Q wave, R wave and S wave, and is the most important feature mark in electrocardiosignal processing. The shape, time interval and amplitude of the QRS complex can be used as an important basis for individual physiological feature recognition. Smaller convolution kernels are used to capture local detail features such as R waves, and larger convolution kernels are used to capture morphological features such as QRS waveforms and relationship features between wave groups.

[0075] The multi-scale convolution module is used for feature extraction of the preprocessed mixed electrocardiosignal data through four convolution branches to obtain a feature tensor, and then the 256-dimensional feature tensor is obtained after fusion by splicing and input to the feature enhancement module.

[0076] In combination Figure 4 The feature enhancement module comprises an adaptive average pooling layer, a first full connection layer, a second full connection layer and an output layer connected in sequence.

[0077] The processing process of the feature enhancement module includes two processing stages of feature compression and adaptive re-labeling. In the feature compression stage, the input feature tensor is compressed into a one-dimensional feature vector by the adaptive average pooling layer. In the adaptive re-labeling stage, the first full connection layer and the second full connection layer are used to learn the dependence between channels, wherein the first full connection layer compresses the channel dimension to 1 / 16 of the original dimension and is processed by the ReLU activation function, the second full connection layer restores the channel dimension to the original dimension and is processed by the Sigmoid activation function to obtain the weight coefficients of each channel, and finally the weight coefficients of each channel are multiplied with the input feature tensor by the output layer to output to the deep feature extraction module.

[0078] In combination Figure 5 The deep feature extraction module is used for further extracting deep features, which comprises two convolution units connected in sequence, each convolution unit is composed of a convolution layer, a batch normalization layer and a ReLU activation function, wherein the first convolution unit keeps the channel number unchanged, and the second convolution unit increases the channel number to 512 dimensions.

[0079] In combination Figure 6The global context module comprises an adaptive average pooling layer, two 1x1 convolution layers and an activation function connected in sequence, wherein the first 1x1 convolution layer reduces the channels to 256 and is processed by a ReLU activation function, and the second 1x1 convolution layer restores the channels to 512 and is processed by a Sigmoid activation function to obtain channel attention weights. The channel attention weights are multiplied with the input feature tensor and output to the self-attention module.

[0080] In combination Figure 7 The self-attention module establishes long-range dependencies between features through a Query-Key-Value mechanism. The Query and Key branches reduce the channel number to 1 / 8 of the original through 1x1 convolution, and the Value branch maintains the original channel number. The product of Query and Key is activated by Softmax to obtain attention scores, and the attention features are calculated with the Value branch. Finally, the enhanced feature representation is obtained through a learnable scaling parameter and a residual connection to adjust the attention intensity by introducing a learnable scaling parameter γ.

[0081] The enhanced feature representation is globally averaged by a global average pooling layer to obtain a 512-dimensional feature tensor, which is reduced to 64-dimensional features by a fully connected layer and normalized by L2 to obtain an embedding feature vector.

[0082] In this embodiment, in order to pre-train the electrocardiosignal extraction network model, the output end of the normalization layer is further connected with a classifier, which is used to map the features to a classification space for calculation of a classification loss in the pre-training process of the electrocardiosignal extraction network model.

[0083] The electrocardiosignal extraction network model constructed is pre-trained, and the pre-training method specifically comprises:

[0084] An exercise-rest mixed electrocardiosignal dataset is obtained.

[0085] The exercise-rest mixed electrocardiosignal dataset is divided into a training set, a validation set and a test set.

[0086] The training set data is taken as input to train the electrocardiosignal extraction network model. In the training process, an Adam optimizer is used for parameter optimization, the learning rate is adjusted, and a loss function is calculated. Then, the test set data is taken as input to test the electrocardiosignal extraction network model. The training is iterated until the loss function does not decrease for 10 consecutive rounds during testing, and the training of the electrocardiosignal extraction network model is stopped. A trained electrocardiosignal extraction network model is obtained.

[0087] The loss function is the sum of a focal loss function and a multiple similarity loss function, and the expression of the focal loss function is:

[0088]

[0089] where, is the class weight, is the probability of the predicted correct class, is the focus factor, controlling the attention to different samples, is the cross-entropy loss.

[0090] The multiple similarity loss function includes the similarity loss of positive sample pairs and the similarity loss of negative sample pairs, and the expression is:

[0091]

[0092]

[0093]

[0094] where, is the positive sample pair similarity, is the negative sample pair similarity, is the positive sample scaling factor, is the negative sample scaling factor, thresh is the similarity threshold, and clamp means limiting the input value within a certain range to ensure numerical stability.

[0095] A preset decision threshold is input into the trained electrocardiosignal extraction network model, the false acceptance rate and the false rejection rate are calculated, the false acceptance rate is the probability of identifying an unauthorized person as an authorized person, the false rejection rate is the probability of identifying an authorized person as an unauthorized person, and when the false acceptance rate and the false rejection rate are equal, the corresponding decision threshold is the optimal decision threshold;

[0096] The decision threshold is set to the optimal decision threshold to obtain the pre-trained electrocardiosignal extraction network model.

[0097] Next, enter the electrocardio authentication phase, which specifically includes:

[0098] The mixed electrocardiosignal data of the to-be-tested person is obtained, and the mixed electrocardiosignal data is obtained by mixing the electrocardiosignal of the to-be-tested person in the motion state and the resting state.

[0099] The mixed electrocardiosignal data is preprocessed, specifically:

[0100] A fourth-order Butterworth band-pass filter is used to filter the signal. The frequency band range of the filter is set to 0.5Hz-40Hz. The reason for selecting this frequency band range is that the main effective components of the electrocardiosignal are concentrated in this range, which can effectively filter out power frequency interference (50Hz / 60Hz) and baseline drift (<0.5Hz).

[0101] A high-pass filter with a cutoff frequency of 0.5 Hz is used to further eliminate the remaining baseline drift in the signal. The high-pass filter adopts a first-order filter to remove the baseline drift while trying to maintain the original morphology of the signal as much as possible.

[0102] The filtered mixed electrocardiosignal is standardized to obtain a processed mixed electrocardiosignal, and the expression of the standardization is:

[0103]

[0104] Among them, represents the filtered mixed electrocardiosignal, represents the signal mean of the filtered mixed electrocardiosignal, represents the signal standard deviation of the filtered mixed electrocardiosignal, represents the stability coefficient, represents the processed mixed electrocardiosignal.

[0105] The processed mixed electrocardiosignal is segmented: the Hamilton algorithm is used to detect the R-wave position in the preprocessed mixed electrocardiosignal, and the time interval of the interval between adjacent R-waves is calculated.

[0106] The window segmentation is performed at the R-wave position to obtain a segmented signal, the window is divided around the R-wave, the window length of 25% is cut forward, the remaining part is cut backward, and the window does not exceed the effective range of the signal, wherein the window length is set according to the signal state.

[0107] The segmented signal is detected for quality control, and the segmented signal meeting the quality control condition is screened, the quality control condition includes that there is no non-numeric value and infinite value, the difference range of the maximum value and the minimum value of the segmented signal is 1.0-15.0, and at least one valid R-wave is included, and the obtained segmented signal is used as the preprocessed exercise-rest state mixed electrocardiosignal data.

[0108] The preprocessed exercise-rest state mixed electrocardiosignal data is used as input, and a plurality of embedding feature vectors are obtained through the pre-trained electrocardiosignal extraction network model, each embedding feature vector corresponds to a segmented signal data.

[0109] The cosine similarity values of the embedding feature vectors are calculated, and the calculation expression is:

[0110]

[0111] Among them, and The cosine similarity value is in the range of [-1, 1], and the value is closer to 1, indicating that the two feature vectors are more similar, and closer to -1, indicating that they are less similar.

[0112] According to the cosine similarity value, the identity of the to-be-tested person is authenticated, if the cosine similarity value is greater than the optimal judgment threshold of the electrocardiosignal extraction network model, the to-be-tested person is determined as an authorized person, and if the cosine similarity value is less than the optimal judgment threshold of the electrocardiosignal extraction network model, the to-be-tested person is determined as an unauthorized person.

[0113] The present application has been fully experimentally verified on the mixed-state electrocardio data set. The optimal threshold value determined on the verification set is 0.6600, and using the threshold value for authentication evaluation on the test set, excellent performance is obtained, the authentication accuracy rate reaches 93.57%, the false acceptance rate is 1.97%, and the false rejection rate is 10.9%, indicating that the present application method has strong robustness and generalization ability. These experimental results also fully prove the effectiveness and practical value of the present application in processing mixed electrocardio data of motion-rest state.

[0114] The above only describes the preferred embodiments of the present application, and it should be pointed out that for ordinary skilled persons in the art, without departing from the technical principles of the present application, a number of improvements and modifications can be made, and these improvements and modifications should also be considered as the protection scope of the present application.

Claims

1. A method for identity authentication based on mixed physiological state electrocardiosignal, characterized in that, The method comprises the following steps: obtaining mixed electrocardio signal data of a to-be-tested person, wherein the mixed electrocardio signal data is obtained by mixing electrocardio signals of the to-be-tested person in a motion state and a resting state; preprocessing the mixed electrocardio signal data to obtain preprocessed mixed electrocardio signal data; inputting the preprocessed mixed electrocardio signal data as an input, and outputting an embedding feature vector based on an electrocardio signal extraction network model; calculating a cosine similarity value based on the embedding feature vector, and authenticating the identity of the to-be-tested person based on the cosine similarity value; the electrocardio signal extraction network model comprises a multi-scale convolution module, a feature enhancement module, a deep feature extraction module, a global context module and a self-attention module connected in sequence, and the output end of the self-attention module is connected with a global average pooling layer, a full connection layer and a normalization layer in sequence; the multi-scale convolution module is used for extracting features from the preprocessed mixed electrocardio signal data to obtain a feature tensor, and the feature tensor is input into the feature enhancement module after feature fusion through channel dimension splicing; the feature enhancement module is used for adaptively adjusting the feature channel of the input feature tensor; the deep feature extraction module is used for deep feature extraction of the input feature tensor; the global context module is used for global information aggregation of the input feature to obtain a channel attention weight, and the channel attention weight is multiplied with the input feature tensor and then input into the self-attention module; the self-attention module is used for establishing a long-range dependency relationship between the input feature tensors; wherein the multi-scale convolution module comprises four parallel convolution branches, the convolution kernel sizes of the convolution branches are 3, 5, 7 and 11 respectively, and the input channel number of each convolution branch is 1 and the output channel number is 64; the deep feature extraction module comprises two convolution units connected in sequence, each of the convolution units comprises a convolution layer, a batch normalization layer and a ReLU activation function, wherein the channel number of the first convolution unit remains unchanged, and the channel number of the second convolution unit is increased to 512; the global context module comprises an adaptive average pooling layer, two 1×1 convolution layers and an activation function connected in sequence, wherein the first 1×1 convolution layer reduces the channel number to 256 and processes the channel number through a ReLU activation function, and the second 1×1 convolution layer restores the channel number to 512 and processes the channel number through a Sigmoid activation function to obtain a channel attention weight; the self-attention module comprises a Query branch, a Key branch and a Value branch, the channel number of the Query branch and the Key branch is reduced to 1 / 8 of the input feature tensor through a 1×1 convolution layer, the channel number of the Value branch is maintained through a 1×1 convolution layer, the product of the Query branch and the Key branch is processed through a Softmax activation function to obtain an attention score, and then the attention feature is calculated with the Value branch, and the attention feature is processed through a learnable scaling parameter and a residual connection to obtain an enhanced feature representation.

2. The method for identity authentication based on mixed physiological state electrocardiosignal according to claim 1, characterized in that, the preprocessing of the motion-resting state mixed electrocardio signal data to obtain preprocessed motion-resting state mixed electrocardio signal data comprises: The motion-rest state mixed electrocardiogram signal is filtered by a band-pass filter, the band-pass filter is a fourth-order Butterworth band-pass filter, and the frequency band range is set to 0.5-40 Hz; The residual baseline drift in the motion-rest state mixed electrocardiogram signal is removed by a high-pass filter to obtain a filtered mixed electrocardiogram signal, the high-pass filter is a first-order filter, and the cutoff frequency is set to 0.5 Hz; The filtered mixed electrocardiogram signal is standardized to obtain a processed mixed electrocardiogram signal, and the expression of the standardization is: ； wherein, denotes the filtered mixed electrocardiosignal, denotes the signal mean of the filtered mixed electrocardiosignal, denotes the signal standard deviation of the filtered mixed electrocardiosignal, denotes the stability coefficient, denotes the processed mixed electrocardiosignal; The processed mixed electrocardiogram signal is segmented and quality controlled to obtain preprocessed motion-rest state mixed electrocardiogram signal data. 3.The method of claim 2, wherein, The processed mixed electrocardiogram signal is segmented and quality controlled to obtain preprocessed motion-rest state mixed electrocardiogram signal data, including: The R-wave position in the preprocessed mixed electrocardiogram signal is detected by using the Hamilton algorithm, and the time interval of the interval of adjacent R-waves is calculated; The segmented signal is obtained by window segmentation at the R-wave position, the window is centered on the R-wave, 25% of the window length is cut forward, the remaining part is cut backward, and the window does not exceed the effective range of the signal, wherein the window length is set according to the signal state; The segmented signal is quality controlled and detected to screen the segmented signal meeting the quality control condition, that is, as the preprocessed motion-rest state mixed electrocardiogram signal data; The quality control condition includes that there is no non-numeric value and infinite value, the difference range of the maximum value and the minimum value of the segmented signal is 1.0-15.0, and at least one valid R-wave is included. 4.The method of claim 1, wherein, The feature enhancement module includes an adaptive average pooling layer, a first full connection layer, a second full connection layer and an output layer connected in sequence; The adaptive average pooling layer is used to compress the input feature tensor into a one-dimensional feature vector; The first full connection layer is used to compress the channel dimension of the one-dimensional feature vector to 1 / 16 of the original dimension and pass through a ReLU activation function for processing; The second full connection layer is used to restore the channel dimension of the one-dimensional feature vector to the original dimension, and obtain the attention weight coefficients of each channel through a Sigmoid activation function; The output layer is used to multiply the weight coefficients of each channel with the input feature tensor channel by channel and then output.

5. The method for identity authentication based on mixed physiological state electrocardiosignal according to claim 1, characterized in that, The output end of the normalization layer is further connected with a classifier, and the classifier is used to map the features to a classification space for calculation of a classification loss in a pre-training process of the electrocardiogram signal extraction network model. 6.The method of claim 1, wherein, The electrocardiogram signal extraction network model is further pre-trained, and the pre-training method includes: An motion-rest state mixed electrocardiogram signal dataset is obtained; The motion-rest state mixed electrocardiogram signal dataset is divided into a training set, a validation set and a test set; The training set data is taken as input to train the electrocardiosignal extraction network model, in the training process, the Adam optimizer is used for parameter optimization, the learning rate is adjusted, and the loss function is calculated, then the test set data is taken as input to test the electrocardiosignal extraction network model, and the training is iterated until the loss function does not decrease for 10 consecutive rounds when testing, and the trained electrocardiosignal extraction network model is obtained; A preset decision threshold is set, the verification set data is input into the trained electrocardiosignal extraction network model, the false acceptance rate and the false rejection rate are calculated, the false acceptance rate is the probability of identifying the unauthorized person as the authorized person, and the false rejection rate is the probability of identifying the authorized person as the unauthorized person, and the decision threshold corresponding to the equal false acceptance rate and false rejection rate is the optimal decision threshold; The decision threshold is set as the optimal decision threshold to obtain the pre-trained electrocardiosignal extraction network model.

7. The method for identity authentication based on mixed physiological state electrocardiosignal according to claim 6, characterized in that, The loss function comprises a sum of a focal loss function and a multiple similarity loss function, and an expression of the focal loss function is: ； wherein, is a class weight, is a probability of the predicted correct class, is a focus factor, is a cross-entropy loss; The multiple similarity loss function comprises a similarity loss of a positive sample pair and a similarity loss of a negative sample pair, and an expression thereof is: ； ； ； wherein, is a multi-similarity loss function, is a similarity loss for positive sample pairs, is a similarity loss for negative sample pairs, is a positive sample pair similarity, is a negative sample pair similarity, is a positive sample scaling factor, is a negative sample scaling factor, thresh is a similarity threshold, and clamp indicates limiting the input value within a preset range. 8.The method of claim 1, wherein, The cosine similarity value between the data is calculated according to the embedding feature vector, and the identity of the testee is authenticated according to the cosine similarity value between the data, comprising: The cosine similarity value between the data is calculated according to the embedding feature vector, and an expression of the cosine similarity value is: ； wherein, and both represent an embedding feature vector, n represents a dimension of the embedding feature vector, represents a cosine similarity value; The identity of the testee is authenticated according to the cosine similarity value between the data, if the cosine similarity value is greater than the optimal judgment threshold of the electrocardiosignal extraction network model, the testee is determined as the authorized person, and if the cosine similarity value is less than the optimal judgment threshold of the electrocardiosignal extraction network model, the testee is determined as the unauthorized person.

Images