Progressive ballistocardiogram signal-based identity database construction method and device, identity recognition method and device and storage medium

Abstract

The invention discloses a progressive ballistocardiogram signal-based identity database construction method, an identity recognition method and device and a storage medium, and belongs to the technical field of biological recognition. The method comprises the following steps: collecting ballistocardiogram signals of a plurality of users, and storing the ballistocardiogram signals in a two-tuple of time and amplitude to obtain original signals; the original signals are preprocessed, and invalid signal segments are removed; performing signal decomposition on the preprocessed signal by using a plurality of time sequence prediction modules, and extracting cardiac beat signal segments; performing random masking and signal reconstruction on the heart beat signal to obtain an intermediate feature vector, and performing comparative learning and metric learning based on the intermediate feature vector to obtain an identity feature vector of the user; and taking the feature vector set of each user as a cluster, determining the center of each cluster, and constructing a target identity database. And an efficient, accurate and stable sign state monitoring algorithm is realized, so that accurate and efficient identity recognition is realized.

Description

Technical Field

[0001] The present invention belongs to the field of biometric identification technology, and more specifically, relates to a progressive identity database construction method based on heartbeat signals, an identity identification method, a device and a storage medium. Background Art

[0002] In recent years, biometric identification has faced multiple challenges, including data security, privacy security, technical accuracy, legal ethics, and system scalability. Currently, commonly used biometric identification methods include fingerprint recognition, but it is susceptible to contamination, damage, and forgery. Although facial recognition is a non-contact method, it is greatly affected by changes in lighting and facial expressions, and privacy and ethical issues are prominent. In addition, most of the iris recognition, voiceprint recognition, and behavioral recognition methods have some inconveniences or are easily interfered with during testing.

[0003] Repeated personal identity verification is a tedious but critical daily task for people to access sensitive information and private assets. Unobtrusive personal identification methods are gaining attention due to their ability to protect individual rights while reducing the burden of repeated verification. BCG (Ballistocardiogram) is a technology that indirectly reflects cardiovascular activity by measuring the tiny movements caused by heart beats. Because the BCG signal contains information reflecting a person's cardiovascular activity, it has become a non-invasive biometric identification method. However, due to the complex components and diverse waveforms of the BCG signal, it poses a challenge to effectively extract identity information.

[0004] At present, the identification method based on BCG signal mainly extracts features through simple deep learning network and compares it with stable ECG (Electrocardiogram). However, these methods have the limitation of too high requirements on the quality of test signals. Traditional feature extraction methods usually rely on fixed feature sets, which makes them vulnerable to individual differences and unable to cope with small individual differences. When the amount of identification data is large, the specificity is not high. The identification of the identification by comparing with ECG signal depends on additional data overhead and cannot achieve true senseless authentication, and it cannot solve the open set problem. Therefore, there are problems in the prior art such as high cost of obtaining BCG signal and the ECG signal for comparison, and inaccurate identification results caused by low specificity during identification. Summary of the invention

[0005] In view of the defects of the related art, the purpose of the present invention is to provide a progressive identity database construction method, identity recognition method, device and storage medium based on cardiac shock signals, aiming to solve the problems in the prior art of high cost of obtaining BCG signals and ECG signals for comparison, inaccurate identity database caused by low specificity during identity recognition, and inaccurate recognition results.

[0006] To achieve the above object, the present invention provides a progressive method for constructing an identity database based on a heartbeat signal, comprising:

[0007] S1, collecting cardiac impulse signals of multiple users, storing them in binary pairs of time and amplitude, and obtaining original signals; preprocessing the original signals to remove invalid signal segments;

[0008] S2. Decompose the preprocessed original signal to extract the heartbeat signal segment q i,j ,q i,j Indicates the jth signal segment of the i-th user; i ranges from 0 to N1, N1 is the number of users, j ranges from 1 to N2, N2 is the number of signal segments;

[0009] S3, by processing the heartbeat signal segment q i,j After random mask processing, the signal is reconstructed to obtain the heartbeat signal segment q i,j Consistent reconstructed signal, obtain the intermediate feature vector f of signal encoding and decoding during signal reconstruction i,j , based on the intermediate eigenvector f i,j Perform contrastive learning and metric learning to obtain the user's feature vector;

[0010] S4. Treat each user's feature vector set as a cluster, determine the center of each cluster, and build a target identity database.

[0011] Optionally, the preprocessing of the original signal includes: performing first-order difference processing on the original signal, calculating the mean and standard deviation of the signal using a sliding window for the first-order difference result, and marking segments with window energy fluctuations exceeding 2 standard deviations as invalid signal segments.

[0012] Optionally, S2 specifically includes:

[0013] S21, performing sequence decomposition on the preprocessed signal X, the decomposition formula is as follows:

[0014] X br =AvgPool(Padding(X))

[0015] X s =XX br

[0016] Among them, AvgPool represents the average pooling layer, Padding represents padding, and X br represents the separated respiratory component, X s Contains heart beat component and noise component;

[0017] S22, use a sliding window with a window size of L1 from X s Extract the first sample point Use a sliding window with a window size of L2 to extract the s Second sampling point Among them, the value range of L1 is the window size corresponding to 0.5s to 0.7s, the value range of L2 is the window size corresponding to 0.04s to 0.08s, and g is the median coordinate of the signal in the sliding window;

[0018] S23, respectively calculating the energy of the first sample point and the energy of the second sample point The first sample point and the second sample point are input into the multi-layer perceptron MLP for prediction, and the heartbeat feature Y=MLP(X s [g-1],…,X s [g-L2]) and noise features Z = MLP(X s [g-1],…,X s [g-L1]), by The standardized heartbeat signal fragment is obtained; the loss function of the multi-layer perceptron MLP is: Among them, when the first sample point is input, V in the loss function adopts When the second sample point is input, V in the loss function adopts

[0019] Optionally, S3 specifically includes:

[0020] S31, by heartbeat signal segment q i,j Perform random mask processing and encode the processed signal to obtain the intermediate feature vector f i,j , based on the intermediate feature vector f i,j Decoding obtains a reconstructed signal, which is consistent with the heartbeat signal segment q i,j Consistency;

[0021] S32, the intermediate feature vector f i,j Input into the twin network model with self-attention as the core to obtain the contrast feature r i,j ; The loss function Loss2 of the twin network model is:

[0022]

[0023] Where k represents the kth user, l represents the lth signal segment;

[0024] S33, the comparison feature r i,j Input the multi-layer perceptron MLP to reduce the dimension and strengthen the contrast features r corresponding to different users i,j The difference between them is used to obtain the metric feature s i,j ; Metric features s corresponding to different users i,j The distance between them is greater than or equal to the preset value m, and the metric feature s i,j As the user's feature vector; the loss function Loss3 of the multi-layer perceptron MLP is expressed as:

[0025] Loss3=∑ i,j,k,l I i＝k (‖s i,j -s k,l ‖)+I i≠k (ReLU(m-‖s i,j -s k,l ‖)).

[0026] Optionally, determining the center of each cluster includes: transforming the feature vector s of each user i,j The average value of is taken as the center of the corresponding cluster.

[0027] In a second aspect, the present invention further provides a progressive identity recognition method based on cardiac impulse signals, which is applied to a target identity database constructed by any method described in the first aspect, comprising:

[0028] Acquire the cardiac signal of the user to be identified, store it in a binary group of time and amplitude, and obtain the original signal to be identified; pre-process the original signal to be identified and remove invalid signal segments;

[0029] Decompose the preprocessed original signal to be identified and extract the heartbeat signal segment q j ,q j represents the jth signal segment of the user; the value range of j is 1 to N2, where N2 is the number of signal segments;

[0030] By j After random mask processing, the signal is reconstructed to obtain the heartbeat signal segment q i,j Consistent reconstructed signal, obtain the intermediate feature vector f of signal encoding and decoding during signal reconstruction j , based on the intermediate eigenvector f j Perform contrastive learning and metric learning to obtain the metric features of the user to be identified;

[0031] Calculate the Euclidean distance between the metric features of the user to be identified and the centers of all clusters in the target identity database to obtain the identity features F of the user to be identified. i , the identity feature F i Input it into the multi-head self-attention module encoder, and input the result into the Softmax layer to obtain the identity feature F i The probability of being authenticated as the user corresponding to each cluster;

[0032] When the maximum probability exceeds the preset threshold, the identity feature F i Identify the identity information corresponding to the maximum probability.

[0033] In a third aspect, the present invention further provides a progressive identity recognition device based on a cardiac impulse signal, comprising:

[0034] The data acquisition and preprocessing module is used to collect the cardiac shock signals of multiple users, store them in binary groups of time and amplitude, and obtain the original signals; preprocess the original signals to remove invalid signal segments;

[0035] The signal decomposition module is used to decompose the preprocessed original signal and extract the heartbeat signal segment q i,j ,q i,j Indicates the jth signal segment of the i-th user; i ranges from 0 to N1, N1 is the number of users, j ranges from 1 to N2, N2 is the number of signal segments;

[0036] A training module is used to train the heartbeat signal segment q i,j After random mask processing, the signal is reconstructed to obtain the heartbeat signal segment q i,j Consistent reconstructed signal, obtain the intermediate feature vector f of signal encoding and decoding during signal reconstruction i,j , based on the intermediate eigenvector f i,j Perform contrastive learning and metric learning to obtain the user's feature vector;

[0037] The identity database construction module is used to treat the feature vector set of each user as a cluster, determine the center of each cluster, and construct the target identity database.

[0038] In a fourth aspect, the present invention further provides a computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a processor to implement the progressive method for constructing an identity database based on cardiac impulse signals as described in any one of the first aspects, or to implement the progressive method for identifying an identity based on cardiac impulse signals as described in the second aspect.

[0039] Compared with the prior art, the above technical solution conceived by the present invention can achieve the following beneficial effects:

[0040] 1. The present invention provides a progressive identity recognition method based on cardiac impact signal, which uses BCG signal as test signal, and performs dimension reduction, feature extraction and signal recovery on the signal through non-orthogonal projection, accurately extracts heartbeat components from composite signals, generates refined heartbeat waveforms, and uses progressive metric learning (contrast learning and metric learning in sequence) to project heartbeat segments into latent space to obtain identity feature information, and promotes the separation of each individual's unique features, creating accurate individual matching identifiers for identity recognition. This solution provides an efficient, accurate and stable vital sign status monitoring algorithm, thereby achieving accurate and efficient identity recognition, which is applicable to multiple fields, including professional medical diagnosis, smart home systems, personal health care, etc., and provides users with more reliable vital sign monitoring solutions.

[0041] 2. The present invention provides a progressive identity recognition method based on cardiac impact signals. By sampling the signal of a specific length at large scale and small scale and predicting the subsequent signal energy, the original data is corrected through a loss function to obtain a more accurate heartbeat signal component, ensuring that the BCG signal used for heartbeat segmentation reduces noise interference. The cardiac impact signal of the subject is decomposed, that is, the heartbeat feature vector is orthogonalized, which improves the availability and robustness of the algorithm. By clustering the clusters of user feature vectors, the vectors of different subjects are projected into the accumulated space, and the clusters of the subjects are distinguished by a certain distance threshold. This solution not only solves the problem of high data acquisition cost in the identity recognition method that requires ECG for comparison, but also solves the problem of inaccurate identity database caused by low specificity during identity recognition. Low-cost, continuous and efficient identity recognition is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flowchart of a progressive method for constructing an identity database based on a heartbeat signal provided by the present invention;

[0043] Figure 2 : is a signal decomposition diagram provided by the present invention, wherein (a) is an original signal diagram, (b) is a breathing trend diagram, (c) is an artifact noise diagram, and (d) is a heartbeat schematic diagram after filtering;

[0044] Figure 3 It is a heat map of the combination performance;

[0045] Figure 4 It is a schematic diagram of the heart beat segmentation at rest and after strenuous exercise;

[0046] Figure 5It is a visualization diagram of the results of metric learning, where (a) is the original image of identity information encoding, (b) is the feature after contrastive learning, and (c) is the final feature after metric learning. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0048] The contents involved in the above embodiment are described below in conjunction with a preferred embodiment.

[0049] Embodiment 1

[0050] like Figure 1 As shown, the embodiment of the present invention provides a progressive method for constructing an identity database based on a heartbeat signal, comprising:

[0051] S1, collecting cardiac impulse signals of multiple users, storing them in binary pairs of time and amplitude, and obtaining original signals; preprocessing the original signals to remove invalid signal segments;

[0052] S2. Decompose the preprocessed original signal to extract the heartbeat signal segment q i,j ,q i,j Indicates the jth signal segment of the i-th user; i ranges from 0 to N1, N1 is the number of users, j ranges from 1 to N2, N2 is the number of signal segments;

[0053] S3, by processing the heartbeat signal segment q i,j After random mask processing, the signal is reconstructed to obtain the heartbeat signal segment q i,j Consistent reconstructed signal, obtain the intermediate feature vector f of signal encoding and decoding during signal reconstruction i,j , based on the intermediate eigenvector f i,j Perform contrastive learning and metric learning to obtain the user's feature vector;

[0054] S4. Treat each user's feature vector set as a cluster, determine the center of each cluster, and build a target identity database.

[0055] The present invention uses BCG signals as test signals and extracts feature vectors, and constructs an identity database through the divided heartbeats and continuous segments of heartbeats. First, signal acquisition and preprocessing are performed, and the heartbeat signal data of each user is stored separately, and invalid signal segments are removed. In this embodiment, the data is collected by PVDF flexible piezoelectric film sensors, which are usually built into mattresses and cushions, and are suitable for long-term non-sensing signal detection. The preprocessed data is subjected to signal separation work. In order to accurately characterize the user's identity through the BCG signal, a single heartbeat or continuous heartbeat is extracted as a basis for identification. The three-peak detection method is adopted, that is, three continuous energy value protrusions are found, which are defined as peaks in the BCG signal, and the middle peak of the three continuous energy protrusions is defined as the J peak. The heartbeat segmentation is performed through the identification of the J peak, and the length of each 2.5-second heartbeat signal cut out from the J peak is a heartbeat sample.

[0056] By the heartbeat signal q i,j Perform random masking and signal reconstruction, and use the loss function to make the reconstructed signal as consistent as possible with the original signal. In this process, the intermediate feature vector f of signal encoding and decoding is obtained. i,j The obtained signal intermediate feature f i,j The feature vectors f of two different users are transformed through the twin network i,j and f k,l Contrastive learning is performed, and the loss function is used to make the feature vectors obtained after encoding the heartbeat samples between two subjects orthogonal, that is, the angle between the two vectors is 90 degrees and the inner product of the two vectors is 0.

[0057] As the number of subjects increases, although the feature vectors of users are mutually orthogonal, the orthogonality of feature vectors alone is still defective. When the feature vectors obtained after encoding are all near the origin, it will be difficult to distinguish the identities of some users because their Euclidean distance is too close. Therefore, it is necessary to introduce progressive metric learning, and use contrastive learning and metric learning methods in stages. Triplet Loss is used to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs, so that the feature vectors of the same user are closer and the feature vectors of different users are farther apart, and the feature vectors are pulled apart to facilitate subsequent identity recognition.

[0058] After the user's feature vector is pulled away, the user identity database is constructed by referring to the fingerprint recognition method, and then the identity characteristics of the user to be tested are compared and calculated with the user identity database to perform the final identity recognition work. The feature vector of each user is concentrated into a cluster, each cluster has a center, and the feature vectors of all users constitute the target identity database. Then, the distance between the feature vector of the user to be identified and the target identity database is calculated to perform identity recognition.

[0059] Optionally, S1 specifically includes:

[0060] S11, collecting the cardiac shock signal of each user, and storing the cardiac shock signal data of each user in a CSV format file in the form of a binary storage of time and amplitude to obtain an original signal;

[0061] S12, performing first-order difference processing on the original signal, using a sliding window to calculate the mean and standard deviation of the signal for the first-order difference result, and marking the segments with window energy fluctuation exceeding 2 standard deviations as invalid signal segments;

[0062] S13: removing invalid signal segments from the original signal.

[0063] The BCG (Ballistocardiogram) signal is collected using a data collector, and a pressure sensor is used as an example; it is stored as a CSV (Comma-Separated Values) file in the form of a binary of time and energy value. Each user's ballistocardiogram signal data is stored in a separate CSV file, and each person is numbered starting from sequence number 1. The initial value of i is 1, and q i It is represented as the heart shock signal data of the i-th user. The preprocessing operation of the signal mainly removes invalid signal fragments (i.e., contains gross motion artifacts and excessive external noise), reads the storage file and cuts the file for subsequent work. Slice the data of each subject starting from sequence number 1, set the initial value of j to 1, and q i,j is the jth signal segment of the ith user. In order to identify such invalid segments with intense physical movements and excessive external noise interference, the signal is processed by first-order difference, and the average energy of the first-order difference result is calculated using a sliding window, and the mean and standard deviation are calculated. The segments with energy fluctuations exceeding 2 standard deviations will be marked as invalid segments.

[0064] Optionally, S2 specifically includes:

[0065] S21, performing sequence decomposition on the preprocessed signal X, the decomposition formula is as follows:

[0066] X br =AvgPool(Padding(X))

[0067] X s =XX br

[0068] Among them, AvgPool represents the average pooling layer, Padding represents padding, and X br represents the separated respiratory component, X sContains heart beat component and noise component;

[0069] S22, use a sliding window with a window size of L1 from X s Extract the first sample point Use a sliding window with a window size of L2 to extract the s Second sampling point Among them, the value range of L1 is the window size corresponding to 0.5s to 0.7s, the value range of L2 is the window size corresponding to 0.04s to 0.08s, and g is the median coordinate of the signal in the sliding window;

[0070] S23, respectively calculating the energy of the first sample point and the energy of the second sample point The first sample point and the second sample point are input into the multi-layer perceptron MLP for prediction, and the heartbeat feature Y=MLP(X s [g-1],…,X s [g-L2]) and noise features Z = MLP(X s [g-1],…,X s [g-L1]), by The standardized heartbeat signal fragment is obtained; the loss function of the multi-layer perceptron MLP is: Among them, when the first sample point is input, V in the loss function adopts When the second sample point is input, V in the loss function adopts

[0071] The original one-dimensional BCG signal is X = {X_1, X_2, ..., X_n}, where n represents the number of discrete points, i.e. the length of the one-dimensional signal. The original BCG signal X can be decomposed into three parts: heart beat component X hb , breathing weight X br and noise component ε,X=X hb +X br +ε.

[0072] In order to extract the pure heartbeat component, the BCG signal is first decomposed into a sequence. The algorithm is as follows

[0073] X br =AvgPool(Padding(X))

[0074] X s =XX br

[0075] Among them, X br represents the separated respiratory component, X s Contains heart beat component and noise component.

[0076] In this embodiment, a sliding window with a window size of L2=7 and L1=101 is used to sequentially select from X s The first sample point is obtained by sampling and the second sampling point Calculate the energy of the first sample point separately and the energy of the second sample point Input the multi-layer perceptron MLP for prediction, and then obtain the heartbeat feature Y and noise feature Z; then use the loss function Loss1 to verify whether the heartbeat feature Y and noise feature Z together are signal X s , so as to correct the data. Figure 2 , we can see the schematic diagram of each component, Figure 3 When the sliding window value is L2=7 and L1=101, the generated heat map shows that the effect is best under this value.

[0077] The feedforward neural network used in this embodiment is composed of multiple layers, including an input layer, a hidden layer, and an output layer. Each layer is composed of trainable neurons, which are connected by weights. MLP is one of the most commonly used supervised learning algorithms and is widely used in classification and regression problems. The training process of MLP involves using the back propagation algorithm and the gradient descent method to adjust the weights to minimize the loss function. This process usually requires a large amount of data and computing resources, and may require adjustment of hyperparameters such as the learning rate, regularization term, number of hidden layers, and number of neurons. When the amount of data is limited, the loss function Loss1 is used to solve the problem of easy overfitting that may exist in MLP.

[0078] Optionally, S3 specifically includes:

[0079] S31, by heartbeat signal segment q i,j Perform random mask processing and encode the processed signal to obtain the intermediate feature vector f i,j , based on the intermediate feature vector f i,j Decoding obtains a reconstructed signal, which is consistent with the heartbeat signal segment q i,j Consistency;

[0080] S32, the intermediate feature vector f i,j Input into the twin network model with self-attention as the core to obtain the contrast feature r i,j ; The loss function Loss2 of the twin network model is:

[0081]

[0082] Where k represents the kth user, l represents the lth signal segment;

[0083] S33, the comparison feature r i,j Input the multi-layer perceptron MLP to reduce the dimension and strengthen the contrast features r corresponding to different users i,j The difference between them is used to obtain the metric feature s i,j ; Metric features s corresponding to different users i,j The distance between them is greater than or equal to the preset value m, and the metric feature s i,j As the user's feature vector; the loss function Loss3 of the multi-layer perceptron MLP is expressed as:

[0084] Loss3=∑ i,j,k,l I i＝k (‖s i,j -s k,l ‖)+I i≠k (ReLU(m-‖s i,j -s k,l ‖)).

[0085] Heartbeat signal segment q i,j Random masking is performed, that is, the signal is divided into 6 patches, the contents of 3 patches are randomly set to 0, and position encoding (Position Embedding) information is added to each patch. The patch is input into the encoding layer and decoding layer of the Transformer. The results are the intermediate feature vectors f i,j , reconstruct the signal p i,j , the loss function is Loss = ∑ i,j ‖p i,j -q i,j ‖. This stage is mainly used to extract the intermediate feature vector f of the signal segment i,j .

[0086] The intermediate eigenvector f i,j Input into the Siamese Network (Twin Network) network architecture with self-attention as the core, that is, f i,j Get Q, K, and V through linear layer encoding respectively, and calculate the contrast features Where d represents the vector dimension.

[0087] In order to ensure that the samples of different users are more different in the feature space, the metric learning method is introduced to further strengthen the differences between the features of different users on the basis of contrastive learning. Specifically, by setting a preset value m, the metric feature s is obtained by dimensionality reduction through the multi-layer perceptron MLP. i,j =MLP(r i,j ).

[0088] Measuring feature s of dimension reduction resulti,j Construct a distance function to strengthen the appropriate features so that the distance between samples of different users is greater than or equal to the preset value m, and use the triplet to calculate the loss function Loss3.

[0089] Optionally, determining the center of each cluster includes: transforming the feature vector s of each user i,j The average value of is taken as the center of the corresponding cluster.

[0090] After the above alignment operation, feature clusters in the latent space are obtained, and a corresponding target identity database is established based on these clusters.

[0091] On the basis of the above embodiment, a method for identifying the identity of a user to be tested can be implemented based on the target identity database, specifically including:

[0092] Acquire the cardiac signal of the user to be identified, store it in a binary group of time and amplitude, and obtain the original signal to be identified; pre-process the original signal to be identified and remove invalid signal segments;

[0093] Decompose the preprocessed original signal to be identified and extract the heartbeat signal segment q j ,q j represents the jth signal segment of the user; the value range of j is 1 to N2, where N2 is the number of signal segments;

[0094] By j After random mask processing, the signal is reconstructed to obtain the heartbeat signal segment q i,j Consistent reconstructed signal, obtain the intermediate feature vector f of signal encoding and decoding during signal reconstruction j , based on the intermediate eigenvector f j Perform contrastive learning and metric learning to obtain the metric features of the user to be identified;

[0095] Calculate the Euclidean distance between the metric features of the user to be identified and the centers of all clusters in the target identity database to obtain the identity features F of the user to be identified. i , the identity feature F i Input it into the multi-head self-attention module encoder, and input the result into the Softmax layer to obtain the identity feature F i The probability of being authenticated as the user corresponding to each cluster;

[0096] When the maximum probability exceeds the preset threshold, the identity feature F i Identify the identity information corresponding to the maximum probability.

[0097] By comparing the distances between the metric features of the user to be identified and each cluster point, the identity feature F of the user to be tested is obtained. i, identified through the encoding layer of TRM. Specifically, the method of constructing each person's ECG identity is to calculate the core features: The static ECG database is obtained, namely, the target identity database R = [R1, R2, ...]. The method for identifying the signal segment of the user to be identified is to calculate the Euclidean distance between the identity feature corresponding to the signal segment and each feature in the target identity database (the center of all clusters), namely: F i =[∑(s x -R i ) 2 ,for i in range(len(R))]. i The information is input into the encoder of the multi-head self-attention module, and the result is input into the Softmax layer to finally obtain the probability of the identity feature being authenticated as each user. When the maximum probability among the probabilities of being authenticated as each user exceeds the preset threshold, the user information corresponding to the cluster corresponding to the probability is determined as the identity information of the signal segment, thereby obtaining the identity information of the user to be identified; wherein the preset threshold is 0.6.

[0098] This solution provides an efficient, accurate and stable method for vital sign status, thereby realizing accurate user database construction for accurate and rapid user identification. Specifically, it adopts a progressive method for identification based on cardiac impact signals, which is applicable to multiple fields, including professional medical diagnosis, smart home systems, personal health care, etc., to provide users with more reliable vital sign monitoring solutions. The application prospects of this method include, but are not limited to, in medical environments, combining cardiac impact signals with patient identification to ensure the security and accuracy of medical information; in mobile devices such as smartphones, using heart rate sensors for identity authentication to provide a contactless security authentication method; in some sensitive and confidential positions, continuous identification can be performed through the collection of cardiac impact signals built into the seat.

[0099] In a specific embodiment, a pressure sensor is set on the mattress to obtain the cardiac signal of each user. Continuous identity authentication can prevent unauthorized access and potential security threats, reduce the number of times users need to log in frequently, and provide a seamless experience. After passing the identity authentication, the user's corresponding personalized settings can be called to change the mattress, or the user's status can be monitored according to the personalized settings. The Matplotlib toolkit in the Python programming tool is used to generate the image. The horizontal axis in the figure represents time, and the vertical axis represents body characteristics. From Figure 4It can be seen that at the beginning and end, the user is in the out-of-bed state (state1), there are two body movement states (state3), and the rest of the time is in a calm state (state2). It can be clearly observed that the algorithm accurately calculates the different physical states of the user, and is consistent with the actual physical state of the user. It proves that the algorithm of the present invention can effectively and accurately calculate the physical state of the user. Figure 4 This is a 300-point heartbeat pattern diagram of a person before and after exercise. The different symbols indicate that our method extracts the specificity of each person's heartbeat without being affected by heart rate. Figure 5 (a) is the original cluster graph of each user’s identity feature. Figure 5 (b) in the figure is the result obtained by distinguishing different users through contrastive learning. Figure 5 (c) in the figure is the result obtained by increasing the distance between users through metric learning. Furthermore, on mobile devices, users can access applications securely without entering a password.

[0100] The embodiment of the present invention uses BCG signals as test signals, does not need to rely on ECG signals for comparison, accurately extracts heartbeat components from composite signals, generates refined heartbeat waveforms, uses progressive metric learning (contrast learning and metric learning) to project heartbeat segments into latent space to obtain identity feature information, and promotes the separation of each individual's unique features, combines fingerprint recognition strategies, creates accurate individual matching identifiers, builds an identity database, and performs identity recognition. The technical problems of the prior art, such as the high cost of obtaining BCG signals and ECG signals for comparison, inaccurate identity databases due to low specificity during identity recognition, and inaccurate recognition results, are solved, and efficient, accurate, and stable monitoring of vital signs is achieved, and an accurate user identity database is built, thereby achieving accurate and efficient identity recognition.

[0101] Embodiment 2

[0102] The present invention also provides a progressive identity recognition device based on cardiac impulse signals, comprising:

[0103] The signal decomposition module is used to decompose the preprocessed original signal and extract the heartbeat signal segment q i,j ,q i,j Indicates the jth signal segment of the i-th user; i ranges from 0 to N1, N1 is the number of users, j ranges from 1 to N2, N2 is the number of signal segments;

[0104] A training module is used to train the heartbeat signal segment q i,j After random mask processing, the signal is reconstructed to obtain the heartbeat signal segment q i,jConsistent reconstructed signal, obtain the intermediate feature vector f of signal encoding and decoding during signal reconstruction i,j , based on the intermediate eigenvector f i,j Perform contrastive learning and metric learning to obtain the user's feature vector;

[0105] The identity database construction module is used to treat the feature vector set of each user as a cluster, determine the center of each cluster, and construct the target identity database.

[0106] A progressive identity recognition device based on cardiac impulse signals provided by an embodiment of the present invention is used to execute a progressive identity recognition method based on cardiac impulse signals provided by any embodiment of the present invention, and has corresponding beneficial effects.

[0107] Embodiment 3

[0108] The present invention also provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable a processor to implement the progressive heartbeat signal-based identity recognition method as described in any one of the first embodiments when executing the computer instructions.

[0109] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A progressive method for constructing an identity database based on a heartbeat signal, characterized in that: include: S1, collecting cardiac impulse signals of multiple users, storing them in binary pairs of time and amplitude, and obtaining original signals; Preprocess the original signal to remove invalid signal fragments; S2. Decompose the preprocessed original signal to extract the heartbeat signal segment q i,j ,q i,j Indicates the jth signal segment of the i-th user; i ranges from 0 to N1, N1 is the number of users, j ranges from 1 to N2, N2 is the number of signal segments; S3, by processing the heartbeat signal segment q i,j After random mask processing, the signal is reconstructed to obtain the heartbeat signal segment q i,j Consistent reconstructed signal, obtain the intermediate feature vector f of signal encoding and decoding during signal reconstruction i,j , based on the intermediate eigenvector f i,j Perform contrastive learning and metric learning to obtain the user's feature vector; S4. Treat each user's feature vector set as a cluster, determine the center of each cluster, and build a target identity database.

2. The method according to claim 1, characterized in that The preprocessing of the original signal includes: performing first-order difference processing on the original signal, using a sliding window to calculate the mean and standard deviation of the signal for the first-order difference result, and marking the fragments whose window energy fluctuation exceeds 2 standard deviations as invalid signal fragments.

3. The method according to claim 1, characterized in that S2 specifically includes: S21, performing sequence decomposition on the preprocessed signal X, the decomposition formula is as follows: X br =AvgPool(Padding(X)) X s =X-X br Among them, AvgPool represents the average pooling layer, Padding represents padding, and X br represents the separated respiratory component, X s Contains heart beat component and noise component; S22, use a sliding window with a window size of L1 to obtain s Extract the first sample point Use a sliding window with a window size of L2 to extract the s Second sampling point Among them, the value range of L1 is the window size corresponding to 0.5s to 0.7s, the value range of L2 is the window size corresponding to 0.04s to 0.08s, and g is the median coordinate of the signal in the sliding window; S23, respectively calculating the energy of the first sample point and the energy of the second sample point The first sample point and the second sample point are input into the multi-layer perceptron MLP for prediction, and the heartbeat feature Y=MLP(X s [g-1],…,X s [g-L2]) and noise features Z = MLP(X s [g-1],…,X s [g-L1]), by The standardized heartbeat signal fragment is obtained; the loss function of the multi-layer perceptron MLP is: Among them, when the first sample point is input, V in the loss function adopts When the second sample point is input, V in the loss function adopts 4. The method according to claim 1, characterized in that S3 specifically includes: S31, by heartbeat signal segment q i,j Perform random mask processing and encode the processed signal to obtain the intermediate feature vector f i,j , based on the intermediate feature vector f i,j Decoding obtains a reconstructed signal, which is consistent with the heartbeat signal segment q i,j Consistency; S32, the intermediate feature vector f i,j Input into the twin network model with self-attention as the core to obtain the contrast feature r i,j ; The loss function Loss2 of the twin network model is: Where k represents the kth user, l represents the lth signal segment; S33, the comparison feature r i,j Input the multi-layer perceptron MLP to reduce the dimension and strengthen the contrast features r corresponding to different users i,j The difference between them is used to obtain the metric feature s i,j ; Metric features s corresponding to different users i,j The distance between them is greater than or equal to the preset value m, and the metric feature s i,j As the user's feature vector; the loss function Loss3 of the multi-layer perceptron MLP is expressed as: Loss3=∑ i,j,k,l I i＝k (‖s i,j -s k,l ‖)+I i≠k (ReLU(m-‖s i,j -s k,l ‖))。 5. The method according to claim 1, characterized in that Determining the center of each cluster includes: i,j The average value of is taken as the center of the corresponding cluster.

6. A progressive identity recognition method based on cardiac impulse signals, using a target identity database constructed by the method according to any one of claims 1 to 5, characterized in that: include: Acquire the cardiac signal of the user to be identified, store it in a binary group of time and amplitude, and obtain the original signal to be identified; pre-process the original signal to be identified and remove invalid signal segments; Decompose the preprocessed original signal to be identified and extract the heartbeat signal segment q j ,q j represents the jth signal segment of the user; the value range of j is 1 to N2, where N2 is the number of signal segments; By j After random mask processing, the signal is reconstructed to obtain the heartbeat signal segment q i,j Consistent reconstructed signal, obtain the intermediate feature vector f of signal encoding and decoding during signal reconstruction j , based on the intermediate eigenvector f j Perform contrastive learning and metric learning to obtain the metric features of the user to be identified; Calculate the Euclidean distance between the metric features of the user to be identified and the centers of all clusters in the target identity database to obtain the identity features F of the user to be identified. i , the identity feature F i Input it into the multi-head self-attention module encoder, and input the result into the Softmax layer to obtain the identity feature F i The probability of being authenticated as the user corresponding to each cluster; When the maximum probability exceeds the preset threshold, the identity feature F i Identify the identity information corresponding to the maximum probability.

7. A progressive identity database construction device based on heartbeat signals, characterized in that: include: A data acquisition and preprocessing module is used to collect the heartbeat signals of multiple users and store them in binary groups of time and amplitude to obtain the original signals; Preprocess the original signal to remove invalid signal fragments; The signal decomposition module is used to decompose the preprocessed original signal and extract the heartbeat signal segment q i,j ,q i,j Indicates the jth signal segment of the i-th user; i ranges from 0 to N1, N1 is the number of users, j ranges from 1 to N2, N2 is the number of signal segments; A training module is used to train the heartbeat signal segment q i,j After random mask processing, the signal is reconstructed to obtain the heartbeat signal segment q i,j Consistent reconstructed signal, obtain the intermediate feature vector f of signal encoding and decoding during signal reconstruction i,j , based on the intermediate eigenvector f i,j Perform contrastive learning and metric learning to obtain the user's feature vector; The identity database construction module is used to treat the feature vector set of each user as a cluster, determine the center of each cluster, and construct the target identity database.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, which are used to enable a processor to implement the progressive identity database construction method based on cardiac impulse signals as described in any one of claims 1 to 5, or to implement the progressive identity recognition method based on cardiac impulse signals as described in claim 6.

Images