ECG identity recognition method based on multi-scale feature selection fusion residual network

Abstract

The application provides an ECG identity recognition method based on a multi-scale feature selection and fusion residual network, which comprises the following steps: obtaining sample electrocardio data for preprocessing to obtain fixed-length electrocardio data; dividing the fixed-length electrocardio data to obtain a training set and a test set; inputting the training set and the test set into an ECG identity recognition network model to determine a final ECG identity recognition network model; obtaining to-be-recognized electrocardio data and inputting the to-be-recognized electrocardio data into the trained ECG identity recognition network model; inputting the to-be-recognized electrocardio data into a basic convolution block to extract a basic feature map, inputting the basic feature map into a first grouping convolution block to extract grouped features, inputting the grouped features into a multi-scale feature selection and fusion module to perform feature fusion and obtain fused features, inputting the fused features into a second grouping convolution block to perform dimension lifting and obtain a target identity label, and determining an identity recognition result according to the target identity label. Thus, the network model parameter quantity is reduced, and the model is lightened.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ECG identity recognition, and particularly relates to an ECG identity recognition method based on a multi-scale feature selection and fusion residual network. BACKGROUND

[0002] In recent years, with the rapid development of information technology and the popularity of the network, personal identity recognition has been widely applied in many fields. In order to improve the security of identity authentication, there is currently a technology that applies biological characteristics to the field of identity authentication. The current biological characteristic authentication method based on ECG generally includes four processes: data collection, preprocessing, feature extraction and recognition. The feature extraction stage usually adopts a deep learning model, however, the current ECG identity recognition technology requires more and more time and cost as the network volume becomes larger and the depth becomes deeper, which is not conducive to the application of the identity recognition algorithm in a larger actual scene. SUMMARY

[0003] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the present application provides an ECG identity recognition method based on a multi-scale feature selection and fusion residual network, which can reduce the network model parameter quantity and realize model lightweight.

[0004] In a first aspect, an ECG identity recognition method based on a multi-scale feature selection and fusion residual network is provided, comprising:

[0005] Obtaining sample electrocardio data, preprocessing the sample electrocardio data to obtain fixed-length electrocardio data, the preprocessing including wavelet denoising, R-point positioning, heart beat segmentation and normalization;

[0006] Dividing the fixed-length electrocardio data to obtain a training set and a test set;

[0007] Inputting the training set into a preset ECG identity recognition network model for training, inputting the test set into the trained ECG identity recognition network model for testing, and determining the final ECG identity recognition network model according to the test result, the ECG identity recognition network model comprising a basic convolution block, a residual network and a multi-scale feature selection and fusion module, the residual network comprising a first grouped convolution block and a second grouped convolution block;

[0008] Obtaining to-be-recognized electrocardio data, and inputting the to-be-recognized electrocardio data into the trained ECG identity recognition network model;

[0009] The electrocardio data to be identified is input to the basic convolution block to extract a basic feature map, the basic feature map is subjected to grouped feature extraction by the first grouped convolution block to obtain grouped features, the grouped features are input to the multi-scale feature selection and fusion module for feature fusion to obtain fused features, the fused features are input to the second grouped convolution block for dimension lifting to obtain a target identity label, and an identity recognition result is determined according to the target identity label.

[0010] In some embodiments of the present application, the sample electrocardio data is obtained, and the sample electrocardio data is preprocessed to obtain fixed-length electrocardio data, including:

[0011] The sample electrocardio data is subjected to wavelet decomposition by a Mallet algorithm to obtain first wavelet coefficients of each layer;

[0012] The first wavelet coefficients of each layer are input to a dynamic soft threshold function for noise coefficient discarding processing to obtain second wavelet coefficients;

[0013] The second wavelet coefficients are subjected to wavelet reconstruction to obtain denoised electrocardio data;

[0014] The denoised electrocardio data is subjected to sampling frequency normalization to obtain resampled electrocardio data;

[0015] The R-wave vertex in the resampled electrocardio data is located by a Pan-Tompkins algorithm;

[0016] The resampled electrocardio data is subjected to heartbeat segmentation with the R-wave vertex as a reference point;

[0017] The resampled electrocardio data subjected to heartbeat segmentation is normalized by a min-max normalization formula to obtain the fixed-length electrocardio data.

[0018] In some embodiments of the present application, the first wavelet coefficients of each layer are input to a dynamic soft threshold function for noise coefficient discarding processing to obtain second wavelet coefficients, including:

[0019] The first wavelet coefficients of each layer are compared with a preset critical threshold value of the corresponding layer respectively, when the first wavelet coefficient is greater than or equal to the critical threshold value, the first wavelet coefficient of the corresponding layer is subjected to dynamic soft threshold calculation to obtain the second wavelet coefficient;

[0020] The calculation formula for obtaining the second wavelet coefficient is:

[0021] At the same time, when the condition |w jk |>T j is met,

[0022] for the second wavelet coefficients, w jk for the first wavelet coefficients, T j for the critical threshold, Tc j is a wavelet decomposition scale, and λ is a universal threshold.

[0023] In some embodiments of the present application, the locating the R-wave peaks in the resampled electrocardiogram data by the Pan-Tompkins algorithm comprises:

[0024] inputting the resampled electrocardiogram data into a band-pass filter for signal filtering processing;

[0025] enhancing components of the QRS complex in the resampled electrocardiogram data by a difference equation to obtain a difference signal;

[0026] performing nonlinear operation on the difference signal to obtain an integral signal;

[0027] when the difference signal satisfies a preset first signal threshold and the integral signal satisfies a preset second signal threshold, determining a maximum value in the QRS complex as an R-wave peak.

[0028] In some embodiments of the present application, the first grouped convolution block includes a plurality of convolution kernels, and the performing grouped feature extraction on the base feature map by the first grouped convolution block to obtain grouped features comprises:

[0029] dividing the base feature map into a plurality of base sub-feature maps evenly, wherein the number of the base sub-feature maps is the same as the number of the convolution kernels;

[0030] each of the convolution kernels respectively performs convolution operation on a corresponding base sub-feature map to obtain a plurality of grouped feature maps;

[0031] splicing the plurality of grouped feature maps according to the channel dimension;

[0032] performing information fusion on the spliced grouped feature maps by a channel rearrangement function to obtain grouped features.

[0033] In some embodiments of the present application, inputting the grouped features into the multi-scale feature selection and fusion module to perform feature fusion to obtain fused features comprises:

[0034] performing feature extraction on the grouped features by three convolution kernels of different sizes to obtain corresponding three feature vectors, and performing element summation on the three feature vectors to obtain a first feature;

[0035] embedding global information into the first feature by a global average pooling layer to obtain channel statistics.

[0036] linearly operating the channel statistics through a full connection layer to obtain a second feature;

[0037] performing multi-classification processing on the second feature through a softmax activation function to obtain three feature attention vectors;

[0038] respectively multiplying elements in the three feature attention vectors to obtain corresponding three fusion sub-features;

[0039] performing element summation on the three fusion sub-features to obtain the fusion feature;

[0040] wherein a calculation formula of element summation on the three feature vectors to obtain the first feature is: U = U1 + U2 + U3, U is the first feature, U1 is a feature vector obtained by performing feature extraction on the grouped feature using a 1x3 convolution kernel, U2 is a feature vector obtained by performing feature extraction on the grouped feature using a 1x5 convolution kernel, and U3 is a feature vector obtained by performing feature extraction on the grouped feature using a 1x7 convolution kernel;

[0041] a calculation formula of the channel statistics is:

[0042] S is the channel statistics, F gp (U) represents a global average pooling function with a parameter U, H is a convolution height, W is a convolution width, c is a total number of channels, U is the first feature, i is a selected convolution height serial number, j is a selected convolution width serial number, is a domain;

[0043] a calculation formula of the second feature is: Z = F fc (S) = δ(B(WS)),

[0044] Z is the second feature, F fc (S) represents a full connection function with a parameter S, δ represents a ReLU function, B is a batch normalization function, S is the channel statistics, W is a convolution width, d is a channel number after linear compression, and c is a total number of original channels, is a domain;

[0045] a calculation formula of the three feature attention vectors is:

[0046]

[0047] wherein a represents a feature attention vector of U1, b represents a feature attention vector of U2, and c represents a feature attention vector of U3.

[0048] In some embodiments of the present application, the training set is input into a preset ECG identity recognition network model for training, the test set is input into the trained ECG identity recognition network model for testing, and the final ECG identity recognition network model is determined according to the test result, comprising:

[0049] The fixed-length electrocardio data is divided into a plurality of data sets;

[0050] One of the plurality of data sets is selected in turn to determine the test set, and the remaining data sets are determined as the training set;

[0051] The ECG identity recognition network model is trained by the training set each time, and the trained ECG identity recognition network model is tested by the test set to obtain a test sub-result, until all test sub-results obtained by taking each data set as the test set are obtained;

[0052] The average value of all test sub-results is obtained, and the average value is determined as the target test result of the ECG identity recognition network model;

[0053] The target test result is compared with a preset model standard value, and when the target test result is greater than the model standard value, it is determined that the ECG identity recognition network model is trained.

[0054] In some embodiments of the present application, the ECG identity recognition device based on the multi-scale feature selection fusion residual network is provided with a basic convolution block, a residual network and a multi-scale feature selection fusion module, the residual network includes a first grouped convolution block and a second grouped convolution block, and the device is used for:

[0055] Obtaining sample electrocardio data, pre-processing the sample electrocardio data to obtain fixed-length electrocardio data, the pre-processing including wavelet denoising, R-point positioning, heart beat segmentation and normalization;

[0056] The fixed-length electrocardio data is divided to obtain a training set and a test set;

[0057] The training set is input into a preset ECG identity recognition network model for training, the test set is input into the trained ECG identity recognition network model for testing, and the final ECG identity recognition network model is determined according to the test result, the ECG identity recognition network model including a basic convolution block, a residual network and a multi-scale feature selection fusion module, the residual network including a first grouped convolution block and a second grouped convolution block;

[0058] obtain to-be-identified electrocardio data, input the to-be-identified electrocardio data into the ECG identity recognition network model which is trained;

[0059] input the to-be-identified electrocardio data into the basic convolution block to extract a basic feature map, perform grouped feature extraction on the basic feature map through the first grouped convolution block to obtain grouped features, input the grouped features into the multi-scale feature selection fusion module to perform feature fusion to obtain fused features, input the fused features into the second grouped convolution block to perform dimension lifting to obtain a target identity label, and determine an identity recognition result according to the target identity label.

[0060] In a second aspect, an ECG identity recognition device based on a multi-scale feature selection fusion residual network is provided, including at least one control processor and a memory in communication connection with the at least one control processor; the memory stores instructions executable by the at least one control processor, and the instructions are executed by the at least one control processor to enable the at least one control processor to execute the ECG identity recognition method based on the multi-scale feature selection fusion residual network as described in the first aspect. The ECG identity recognition device based on the multi-scale feature selection fusion residual network is preconfigured with a basic convolution block, a residual network and a multi-scale feature selection fusion module, the residual network includes a first grouped convolution block and a second grouped convolution block, and the device is used for:

[0061] obtain sample electrocardio data, pre-process the sample electrocardio data to obtain fixed-length electrocardio data, and the pre-processing includes wavelet denoising, R-point positioning, heart beat segmentation and normalization;

[0062] divide the fixed-length electrocardio data to obtain a training set and a test set;

[0063] input the training set into a pre-set ECG identity recognition network model for training, input the test set into the ECG identity recognition network model which is trained for testing, determine a final ECG identity recognition network model according to a test result, and the ECG identity recognition network model includes a basic convolution block, a residual network and a multi-scale feature selection fusion module, and the residual network includes a first grouped convolution block and a second grouped convolution block;

[0064] obtain to-be-identified electrocardio data, input the to-be-identified electrocardio data into the ECG identity recognition network model which is trained;

[0065] The ECG data to be identified is input to the basic convolution block to extract a basic feature map, the basic feature map is subjected to grouped feature extraction by the first grouped convolution block to obtain grouped features, the grouped features are input to the multi-scale feature selection and fusion module for feature fusion to obtain fused features, the fused features are input to the second grouped convolution block for dimension increase to obtain a target identity label, and an identity recognition result is determined according to the target identity label.

[0066] In a third aspect, an electronic device is provided, which comprises the ECG identity recognition device based on the multi-scale feature selection and fusion residual network according to the second aspect.

[0067] In a fourth aspect, a computer readable storage medium is provided, which stores computer executable instructions for executing the ECG identity recognition method based on the multi-scale feature selection and fusion residual network according to the first aspect.

[0068] The ECG identity recognition method based on the multi-scale feature selection and fusion residual network according to the embodiments of the present application has at least the following beneficial effects: fixed-length ECG data with noise reduction and most distinctive feature information retained is obtained by preprocessing sample ECG data through wavelet denoising, R-point positioning, heartbeat segmentation and normalization. The final ECG identity recognition network model is obtained by training the ECG identity recognition network model according to the test set and the training set corresponding to the fixed-length ECG data. The ECG data to be identified is input to the ECG identity recognition network model, the basic feature map output by the basic convolution block is subjected to grouped feature extraction by the first grouped convolution block in the ECG identity recognition network model, thereby reducing the network model parameter quantity and improving the model lightweight, and the feature selection and fusion module is used for feature fusion, thereby selectively fusing information of different scales together, enhancing the network model feature extraction capability, and improving the ECG identity recognition accuracy and generalization. The fused features are input to the second grouped convolution block for dimension increase to obtain a target identity label, and an identity recognition result is determined according to the target identity label. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1 is a flowchart of an ECG identity recognition method based on a multi-scale feature selection and fusion residual network provided by the embodiments of the present application;

[0070] Figure 2 is a flowchart of preprocessing sample ECG data provided by another embodiment of the present application;

[0071] Figure 3 is a flowchart of determining an R-wave vertex provided by another embodiment of the present application;

[0072] Figure 4 is a flowchart of processing a base feature map by a first grouped convolutional block according to another embodiment of the present application;

[0073] Figure 5 is a flowchart of processing grouped features by a multi-scale feature selection fusion module according to another embodiment of the present application;

[0074] Figure 6 is a flowchart of obtaining a target test result of an ECG identity recognition network model according to another embodiment of the present application;

[0075] Figure 7 is a general flowchart of an ECG identity recognition network model according to another embodiment of the present application;

[0076] Figure 8 is a schematic diagram of a block in an ECG identity recognition network model according to another embodiment of the present application;

[0077] Figure 9 is a schematic diagram of a multi-scale feature selection fusion module MSFConv according to another embodiment of the present application;

[0078] Figure 10 is a schematic diagram of a grouped convolutional block Gconv in an ECG identity recognition network model according to another embodiment of the present application;

[0079] Figure 11 is a schematic diagram of a short connection block in an ECG identity recognition network model according to another embodiment of the present application;

[0080] Figure 12 is a structural diagram of an ECG identity recognition device based on a multi-scale feature selection fusion residual network according to another embodiment of the present application. DETAILED DESCRIPTION

[0081] Embodiments of the present application are described in detail below with reference to the accompanying drawings, in which the same or similar elements or elements having the same or similar functions are denoted by the same or similar reference numerals throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present application, and should not be understood as limiting the present application.

[0082] In the description of the present application, it should be understood that the orientation description, such as the orientation or position relationship indicated by up, down, front, back, left, right, etc., is based on the orientation or position relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore should not be understood as limiting the present application.

[0083] In the description of the present application, the meaning of one or more is one or more, the meaning of multiple is more than two, greater than, less than, more than, etc. are understood as not including the number, above, below, within, etc. are understood as including the number. If the first, second is described, it is only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features or the sequence of indicated technical features.

[0084] In the description of the present application, unless otherwise explicitly limited, the words such as setting, installing, connecting, etc. should be broadly understood, and those skilled in the art can reasonably determine the specific meaning of the above words in the present application in combination with the specific content of the technical solution.

[0085] The embodiment of the present application provides an ECG identity recognition method based on a multi-scale feature selection and fusion residual network: through wavelet denoising, R point positioning, heartbeat segmentation and normalization preprocessing of sample electrocardio data, fixed length electrocardio data which is denoised and retains the most distinguishing feature information is obtained. According to the test set and the training set corresponding to the fixed length electrocardio data, the ECG identity recognition network model is trained to obtain the final ECG identity recognition network model, and the credit data to be identified is input into the ECG identity recognition network model. The first grouping convolution block in the ECG identity recognition network model is used for grouping feature extraction of the basic feature map output by the basic convolution block, so as to reduce the network model parameter quantity and improve the realization of model lightweight. At the same time, the feature selection and fusion module is used for feature fusion, so that information of different scales can be selectively fused together, the ability of network model feature extraction is enhanced, and the accuracy and generalization of ECG identity recognition are improved. The fusion feature is input into the second grouping convolution block to obtain a target identity label, and the identity recognition result is determined according to the target identity label.

[0086] The control method of the embodiment of the present application will be further described below based on the drawings.

[0087] Reference Figure 1 , Figure 1 A flowchart of an ECG identity recognition method based on a multi-scale feature selection and fusion residual network is provided for the embodiment of the present application. The ECG identity recognition method based on the multi-scale feature selection and fusion residual network includes but is not limited to the following steps:

[0088] In step S110, sample electrocardio data is obtained, and the sample electrocardio data is preprocessed to obtain fixed length electrocardio data. The preprocessing includes wavelet denoising, R point positioning, heartbeat segmentation and normalization.

[0089] In step S120, the fixed length electrocardio data is divided to obtain a training set and a test set.

[0090] Step S130, input the training set to the preset ECG identity recognition network model for training, input the test set to the trained ECG identity recognition network model for testing, determine the final ECG identity recognition network model according to the test result, the ECG identity recognition network model comprises a basic convolution block, a residual network and a multi-scale feature selection fusion module, the residual network comprises a first grouped convolution block and a second grouped convolution block;

[0091] Step S140, obtaining the to-be-identified electrocardio data, inputting the to-be-identified electrocardio data into the ECG identity recognition network model completed the training;

[0092] Step S150, inputting the to-be-identified electrocardio data into the basic convolution block to extract the basic feature map, extracting the grouped feature by the first grouped convolution block on the basic feature map, inputting the grouped feature into the multi-scale feature selection fusion module to perform feature fusion to obtain the fusion feature, inputting the fusion feature into the second grouped convolution block to perform dimensionality lifting to obtain the target identity label, and determining the identity recognition result according to the target identity label.

[0093] It should be noted that there are many noises and interference data in the original sample electrocardio data, which affect the accuracy, therefore, the preprocessing is needed, and the wavelet denoising is used to remove the electrocardio signal baseline drift, power frequency interference and electromyographic noise. The R-point positioning is used to find the position for heart beat segmentation, the sample electrocardio data is segmented by heart beat, the data with the most distinguishing feature information in the electrocardio data is reserved, and then the segmented electrocardio data is normalized to obtain the fixed length electrocardio data.

[0094] It should be noted that the fixed length electrocardio data is used to train the ECG identity recognition network model, the test set is used to test the ECG identity recognition network model, and if the test result is greater than the preset model precision threshold, it is considered that the training of the ECG identity recognition network model is completed, and the ECG identity recognition network model completed the training is used as the final model for use.

[0095] It should be noted that the ECG identity recognition network model includes a basic convolution block, a residual network and a multi-scale feature selection and fusion module. The residual network includes a first group convolution block and a second group convolution block. The current parameter quantity in the traditional residual network is large, and the operation efficiency is also low. Therefore, the first group convolution block is used to replace the corresponding group convolution operation of the traditional convolution block, so that the efficiency is higher and the parameter quantity is smaller. Then, the group features output by the first group convolution block are input into the multi-scale feature selection and fusion module for fusion to obtain fused features. The multi-scale feature selection and fusion module can fuse information of multiple scales, thereby improving the feature extraction capability of the network model. Then, the features are output through the second group convolution block, and the final target identity label is obtained through the average pooling layer and the full connection layer. Further, referring to Figure 7 and Figure 8 , the ECG identity recognition network model includes a basic convolution block and a plurality of block blocks. Each block block is connected with a short connection block. For example, 9 block blocks are used, and the short connection block also has 9 corresponding short connection blocks. Each block block includes two lightweight group convolution blocks Gconv and a multi-scale feature selection and fusion module MSFConv. The two group convolution blocks are the first group convolution block and the second group convolution block. Therefore, the to-be-identified electrocardio data is sequentially subjected to the basic convolution block, the plurality of block blocks including the first group convolution block, the multi-scale feature fusion module and the second group convolution block, the short connection block connected with the block, and is output after entering the average pooling layer and the full connection layer. Further, referring to Figure 11 , Figure 11 is a schematic diagram of the short connection block in the ECG identity recognition network model according to another embodiment of the present application. The information in the block block is input into the short connection block, sequentially subjected to a 1x1 convolution layer, a batch norm (BN) batch normalization process and a nonlinear activation ReLU function, and then output and input into the next block block. The multi-scale feature selection and fusion module can selectively fuse information of different scales together, enhance the feature extraction capability of the network model, improve the accuracy and generalization of the ECG identity recognition method, and reduce the parameter quantity of the network model by using the group convolution block, which is conducive to realizing the lightweight of the model, so that the model can be applied to a larger actual scene.

[0096] In another embodiment, referring to Figure 2 , Figure 2 is a flowchart of a pre-processing process for sample electrocardio data according to another embodiment of the present application. Step S110 includes:

[0097] Step S111, wavelet decomposition is performed on the sample electrocardio data by using the Mallet algorithm to obtain first wavelet coefficients of each layer.

[0098] Step S112, input the first wavelet coefficients of each layer to a dynamic soft threshold function to perform noise coefficient discarding processing to obtain second wavelet coefficients;

[0099] Step S113, perform wavelet reconstruction according to the second wavelet coefficients to obtain denoised electrocardio data;

[0100] Step S114, perform sampling frequency normalization on the denoised electrocardio data to obtain resampled electrocardio data;

[0101] Step S115, locate the R-wave peak in the resampled electrocardio data by using a Pan-Tompkins algorithm;

[0102] Step S116, perform heartbeat segmentation on the resampled electrocardio data with the R-wave peak as a reference point;

[0103] Step S117, perform normalization on the resampled electrocardio data after heartbeat segmentation by using a min-max normalization formula to obtain fixed-length electrocardio data.

[0104] It should be noted that the sample electrocardio data includes a lot of noise, and in order to improve the accuracy of subsequent training of the ECG identity recognition network model, noise removal processing is required. The Mallat algorithm is used to perform wavelet decomposition on the sample electrocardio data to obtain first wavelet coefficients of each layer; a dynamic soft threshold function is proposed to process the wavelet coefficients of each layer in combination with the characteristics of the electrocardio data. The main function of the dynamic soft threshold function is to provide the critical threshold value required for wavelet coefficient processing. The specific operation process is to compare the first wavelet coefficients of each layer with the corresponding critical threshold value to determine whether they contain noise components. Generally, if the first wavelet coefficient is less than the critical threshold value, it is considered to reflect a noise component and should be discarded; if it is greater than the critical threshold value, it is considered to be caused by a signal and should be retained. The mathematical form of the above processing process is shown in the formula:

[0105]

[0106] In the formula, the critical threshold value The specific value is affected by the wavelet decomposition scale j, and λ is a general threshold value proposed by Donoho et al. The dynamic soft threshold function can ensure that the deviation between the second wavelet coefficients and the first wavelet coefficients w jk is at ± λ, and the deviation will become smaller as the decomposition scale increases.

[0107] At the same time, when the condition |w jk |> T j is met, there is

[0108]

[0109] It can be seen that the second wavelet coefficient value obtained according to the dynamic soft threshold function processing between the soft threshold and the hard threshold. When |w jk |>T j , is a nonlinear function, at this time, if the value of w jk increases, the value of gradually approaches the soft threshold; if the wavelet decomposition layer number j increases, the value of decreases inversely, the value of gradually approaches the hard threshold. The baseline drift of the sample electrocardio data, the power frequency interference and the electromyographic noise are removed. The denoised electrocardio data can be obtained by wavelet reconstruction on the basis of the second wavelet coefficient obtained after processing. Then, the Pan-Tompkins algorithm is used for R-point positioning. The Pan-Tompkins algorithm is used for R-point positioning, heart beat segmentation and normalization of the denoised electrocardio data, to obtain fixed-length electrocardio data and individual labels. The sampling frequency of the denoised electrocardio data in each database is normalized, and the resampled electrocardio data is obtained by uniform resampling. For example, the resampling is uniformly performed at 500 Hz. The Pan-Tompkins algorithm is used for R-point positioning of the resampled electrocardio data, and the resampled electrocardio data is segmented based on the R-wave peak point. Further, in order to retain the QRS complex which is the most distinctive feature information in the electrocardio signal, the dimension of the heart beat unit is uniformly set to 256 based on the sampling frequency of 500 Hz, with 64 sampling points on the left and 192 sampling points on the right of the R-wave peak point as the center. The segmented resampled electrocardio data is normalized by using the min-max standardization, to obtain the fixed-length electrocardio data.

[0110] In another embodiment, with reference to Figure 3 , Figure 3 is a flowchart for determining the R-wave peak point provided by another embodiment of the present application, and step S115 includes:

[0111] Step S1151, inputting the resampled electrocardio data into a band-pass filter for signal filtering processing;

[0112] Step S1152, enhancing the component of the QRS complex in the resampled electrocardio data by using a difference equation, to obtain a difference signal;

[0113] Step S1153, performing nonlinear operation on the difference signal to obtain an integral signal;

[0114] Step S1154, when the difference signal satisfies a preset first signal threshold and the integral signal satisfies a preset second signal threshold, determining the maximum value in the QRS complex as the R-wave peak point.

[0115] It should be noted that the resampled electrocardio data is first processed by a band-pass filter, and then the QRS complex component is enhanced by a difference equation to obtain a difference signal; then the difference signal is subjected to a series of nonlinear operations such as point-by-point squaring and window sliding integration to obtain an integral signal; the points on the difference signal and the integral signal are searched, and when the value of the point on the difference signal meets the preset first signal threshold and the value of the point on the integral signal meets the preset second signal threshold, it is determined that the QRS complex is detected; finally, the maximum value in the complex range is found, and the R wave vertex positioning is completed. In order to improve the positioning accuracy, the "refractory period" and the backtracking mechanism can be introduced, which can effectively reduce the occurrence of missed detection and false detection.

[0116] In another embodiment, referring to Figure 4 , Figure 4 is a flowchart of a first grouped convolution block processing a base feature map according to another embodiment of the present application, the first grouped convolution block comprising a plurality of convolution kernels, comprising:

[0117] Step S1511, the base feature map is evenly divided into a plurality of base sub-feature maps, wherein the number of base sub-feature maps is the same as the number of convolution kernels;

[0118] Step S1512, each convolution kernel respectively performs convolution operation on the corresponding base sub-feature map to obtain a plurality of grouped feature maps;

[0119] Step S1513, the plurality of grouped feature maps are spliced according to the channel dimension;

[0120] Step S1514, the spliced grouped feature map is informationally mixed by a channel rearrangement function to obtain a grouped feature.

[0121] It should be noted that referring to Figure 10 is a schematic diagram of a grouped convolution block Gconv in an ECG identity recognition network model according to another embodiment of the present application. An improved residual network model is used as the backbone network of the ECG identity recognition network model, and the convolution operation in the traditional residual network is replaced by a grouped convolution operation with higher efficiency and fewer parameters. The grouped convolution is to divide the base feature map into N groups of base sub-feature maps, and the convolution kernel is also divided into N groups. Each group of convolution kernels is convolved with the corresponding base sub-feature map, and the N groups of base sub-feature maps obtained are spliced according to the channel dimension and output. The parameter quantity of the grouped convolution is 1 / N of the normal convolution, which effectively reduces the parameter quantity of the network model without reducing the feature extraction capability of the network. For example, refer to Figure 10 The number of grouped convolutions is set to 8 groups. Refer to Figure 8The grouped convolution makes each group independent of each other without feature fusion. Therefore, a channel shuffle function is added after the first grouped convolution block to realize information fusion between groups and improve information exchange between grouped channels.

[0122] In another embodiment, with reference to Figure 5 , Figure 5 is a flowchart of a multi-scale feature selection fusion module provided by another embodiment of the application for processing grouped features, comprising the steps of:

[0123] In step S1521, three feature vectors corresponding to the grouped features are obtained by using three convolution kernels of different sizes to extract features, and a first feature is obtained by element summation of the three feature vectors.

[0124] In step S1522, a global information is embedded into the first feature by a global average pooling layer to obtain a channel statistic quantity.

[0125] In step S1523, a second feature is obtained by linear operation of the channel statistic quantity by a full connection layer.

[0126] In step S1524, three feature attention vectors are obtained by multi-classification processing of the second feature by a softmax activation function.

[0127] In step S1525, three fusion sub-features corresponding to the three feature attention vectors are obtained by element multiplication of the elements in the three feature attention vectors.

[0128] In step S1526, a fusion feature is obtained by element summation of the three fusion sub-features.

[0129] The calculation formula of the first feature obtained by element summation of the three feature vectors is U=U1+U2+U3, where U is the first feature, U1 is a feature vector obtained by using a 1x3 convolution kernel to extract features of the grouped features, U2 is a feature vector obtained by using a 1x5 convolution kernel to extract features of the grouped features, and U3 is a feature vector obtained by using a 1x7 convolution kernel to extract features of the grouped features.

[0130] The calculation formula of the channel statistic quantity is:

[0131] S is the channel statistic quantity, F gp (U) represents a global average pooling function with parameter U, H is the convolution height, W is the convolution width, C is the channel, U is the first feature, i is the selected convolution height serial number, j is the selected convolution width serial number, is the domain.

[0132] The calculation formula of the second feature is: Z=F fc (S) = delta(B(WS)),

[0133] Z is the second feature, F fc (S) represents a full connection function with parameter S, delta represents a ReLU function, B is a batch normalization function, S is a channel statistic, W is a convolution width, d is a channel number after linear compression, c is a total channel number without linear compression, is a domain;

[0134] The calculation formula of the three feature attention vectors is:

[0135]

[0136] Wherein a represents a feature attention vector of U1, b represents a feature attention vector of U2, and c represents a feature attention vector of U3.

[0137] It should be noted that the reference Figure 9 , Figure 9 is a schematic diagram of a multi-scale feature selection and fusion module MSFConv provided by another embodiment of the application, after realizing cross-group information intermingling through a channel rearrangement function, a multi-scale feature selection and fusion module MSFConv that can replace an ordinary convolution is adopted, the advantage of the module lies in that multiple scales of information are fused, features of signals under multiple scales are mined at a finer granularity level, and the size of a receptive field of a neuron can be adaptively changed during inference through the addition of an attention mechanism, automatic selection operations between multiple kernels with different convolution kernel sizes are performed, different scales of information are selectively fused together, and the feature extraction capability of the network model is improved. For any given feature map Firstly, three convolution kernels with sizes of 1x3, 1x5 and 1x7 are used for feature extraction, and batch normalization (BN) and a nonlinear activation ReLU function are sequentially performed respectively, to obtain three feature vectors U1, U2 and U3, in order to enable a neuron to adaptively adjust the size of its receptive field according to stimulation content, the feature vectors of the three branches are input into neurons of the next layer carrying different scales of information, and the feature vectors of the three branches are summed element by element to obtain a first feature:

[0138] U = U1 + U2 + U3

[0139] Then, global average pooling operation is performed on the first feature to embed global information in the first feature, to obtain a channel statistic Wherein:

[0140]

[0141] Then use the full connection layer to Transformed into That is, through the linear operation Compressed into B in the following formula represents batch normalization, and δ represents the ReLU function, wherein:

[0142]

[0143] In order to realize accurate and adaptive selection, a compact feature is created by a simple FC layer Reduce the dimension, improve the efficiency.

[0144] At the same time, in order to study the influence of d on the efficiency of the model, the compression ratio r is used to control its value, wherein L represents the minimum value of d:

[0145] d=max(C / r,L)

[0146] Then Z is divided into three branches through the softmax activation function, and is expanded back to c*1*1;

[0147]

[0148] Wherein a, b, c represent the feature attention vectors of U1, U2 and U3 respectively. We obtain A1, A2 and A3 through the element product of the three vectors.

[0149] Finally, the multi-scale selection fusion feature A is obtained by summing the elements, and the module A is refined compared with the initial X, and the information of multiple receptive fields is fused.

[0150] It should be noted that reference Figure 6 , Figure 6 The flowchart for obtaining the target test result of the ECG identity recognition network model provided by another embodiment of the present application is shown in Figure 130, comprising:

[0151] Step S131, the fixed length electrocardio data is divided into multiple data sets;

[0152] Step S132, alternately select one data set from the multiple data sets to determine as a test set, and determine the remaining data sets as training sets;

[0153] Step S133, the ECG identity recognition network model is trained through the training set each time, and the trained ECG identity recognition network model is tested through the test set to obtain a test sub result, until all test sub results obtained by taking each data set as a test set are obtained;

[0154] Step S134, obtaining the average value of all test sub-results, and determining the average value as the target test result of the ECG identity recognition network model;

[0155] Step S135, comparing the target test result with a preset model standard value, and when the target test result is greater than the model standard value, it is determined that the ECG identity recognition network model is trained.

[0156] It should be noted that the test set is input into the ECG identity recognition network model, and the final ECG identity recognition classification result is obtained through ten-fold cross-validation. For example, the fixed-length electrocardiogram data is divided into 10 data sets, and each data set is tested once, and the rest is used as a training set. Cross-validation is repeated 10 times, each time selecting a data set as a test set, and the average cross-validation recognition accuracy of 10 times is taken as the result. In this way, all samples are used as training sets and test sets, and each sample is verified once, which improves the generalization performance of the model and makes the evaluation index more convincing.

[0157] Further, in actual application, standard public electrocardiogram databases can be used as experimental data sources, and ECG signals of different data blocks have different characteristics. Four databases, ECG-ID, MIT-BIH Arrhythmia, MIT-BIH STChange and PTB Diagnostic ECG, can be selected for research. The four databases cover normal electrocardiogram signals, cardiovascular disease electrocardiogram signals, and exercise state electrocardiogram signals, and can verify the generalization performance of the model.

[0158] Further, the target test result obtained is evaluated, including:

[0159] Heartbeat recognition accuracy: the proportion of the number of single heartbeat recognition correct to the total number of heartbeats;

[0160] Precision (Precision), which refers to the proportion of true positive examples in all positive examples: Pre=TP / (TP+FP);

[0161] Recall (Recall), which refers to the proportion of correctly identified positive examples to all positive samples: Recall=TP / (TP+FN);

[0162] F1 score (F1 Score), which is the harmonic mean of precision and recall, and takes into account the precision and recall of the classification model:

[0163] In the formula, TP (True Positive) is the number of positive class samples predicted as positive class labels; FP (False Positive) is the number of positive class samples predicted as negative class labels; FN (False Negative) is the number of negative class samples predicted as positive class labels; and N (True Negative) is the number of negative class samples predicted as negative class labels.

[0164] In the ECG-ID, MIT-BIH ST Change, MIT-AHA, and PTB databases, the identity recognition accuracy and other evaluation index results of the ECG identity recognition method based on the multi-scale feature selection and residual network model of the embodiment are shown in Table 1. It can be seen that in different databases, the identity recognition accuracy and other evaluation index results of the ECG identity recognition method based on the multi-scale feature selection and residual network model are relatively high.

[0165] Table 1

[0166] Database Heartbeat recognition accuracy Precision Pre Recall F1 score ECG-ID 94.16％ 93.12％ 92.42％ 92.67％ MITBIH Arrhythmia 99.11％ 99.13％ 99.13％ 99.13％ PTB 99.65％ 99.61％ 99.62％ 99.61％ MITBIH ST Change 99.54％ 99.50％ 99.54％ 99.52％

[0167] As shown in Table 2, the identity recognition accuracy results of different backbone networks in the four data sets of the embodiment are shown in the table. In the table, MSFNet is a multi-scale feature selection and residual network model. It can be seen that the recognition accuracy of the multi-scale feature selection and residual network model is higher than that of other models.

[0168] Table 2

[0169] ResNet18 ResNet34 MSFNet MITBIH ST Change 99.16％ 99.20％ 99.54％ MITBIH Arrhythmia 98.62％ 98.77％ 99.11％ ECG-ID 90.54％ 91.65％ 94.16％ PTB 99.32％ 99.40％ 99.65％

[0170] As shown in Table 2, the identity recognition accuracy results of different backbone networks in the four data sets of the embodiment are shown in the table. In the table, MSFNet is a multi-scale feature selection and residual network model. It can be seen that the recognition accuracy of the multi-scale feature selection and residual network model is higher than that of other models. Figure 12 Figure 8 As shown in Table 2, the identity recognition accuracy results of different backbone networks in the four data sets of the embodiment are shown in the table. In the table, MSFNet is a multi-scale feature selection and residual network model. It can be seen that the recognition accuracy of the multi-scale feature selection and residual network model is higher than that of other models.

[0171] The processor 1210 can be implemented in a general-purpose central processing unit (CPU), a microprocessor, an application specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute related programs to implement the technical solutions provided by the embodiments of the present application.

[0172] ​The memory 1220 can be implemented in the form of a Read Only Memory (ROM), a static storage device, a dynamic storage device, or a Random Access Memory (RAM), etc. The memory 1220 can store an operating system and other application programs, and when the technical solutions provided by the embodiments of the present specification are implemented by software or firmware, the related program codes are stored in the memory 1220 and are called and executed by the processor 1210 to implement the ECG identity recognition method based on the multi-scale feature selection fusion residual network according to the embodiments of the present application;

[0173] The input / output interface 1230 is configured to realize information input and output.

[0174] The communication interface 1240 is configured to realize the communication interaction between the device and other devices, and can realize the communication through a wired manner (for example, a USB, a network cable, etc.) or a wireless manner (for example, a mobile network, WIFI, Bluetooth, etc.).

[0175] The bus 1250 is configured to transmit information between various components (for example, the processor 1210, the memory 1220, the input / output interface 1230, and the communication interface 1240) of the device.

[0176] The processor 1210, the memory 1220, the input / output interface 1230, and the communication interface 1240 are connected to each other through the bus 1250 to realize the communication connection between them in the device.

[0177] The embodiments of the present application also provide an electronic device including the ECG identity recognition device based on the multi-scale feature selection fusion residual network as described above.

[0178] The embodiments of the present application also provide a storage medium, which is a computer readable storage medium, and the storage medium stores a computer program. The computer program is executed by a processor to implement the ECG identity recognition method based on the multi-scale feature selection fusion residual network.

[0179] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. In addition, the memory can include a high-speed random access memory, and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory can optionally include a memory disposed remotely with respect to the processor, which can be connected to the processor through a network. Examples of the above network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof. The above-described device embodiments are only illustrative, and units described as separate components can or can not be physically separated, implemented in one place, or distributed to multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment.

[0180] Those of ordinary skill in the art can understand that all or some steps in the above disclosed method and system can be implemented as software, firmware, hardware, and appropriate combinations thereof. Some or all physical components can be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transitory media). As known to those of ordinary skill in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, magnetic disk storage, or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. In addition, as known to those of ordinary skill in the art, communication media typically includes computer readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transport mechanisms, and can include any information delivery medium.

[0181] The above is a specific description of the preferred embodiment of the present application, but the present application is not limited to the above-described embodiments, and those skilled in the art can make various equivalent modifications or replacements without departing from the spirit of the present application. These equivalent modifications or replacements are all included in the scope defined by the claims of the present application.

Claims

1. An ECG identity recognition method based on a multi-scale feature selection fusion residual network, characterized in that, The method comprises the following steps: obtaining sample electrocardio data, pre-processing the sample electrocardio data to obtain fixed-length electrocardio data, the pre-processing comprising wavelet denoising, R-point positioning, heart beat segmentation and normalization; dividing the fixed-length electrocardio data to obtain a training set and a test set; inputting the training set into a preset ECG identity recognition network model for training, inputting the test set into the trained ECG identity recognition network model for testing, and determining a final ECG identity recognition network model according to the test result, the ECG identity recognition network model comprising a basic convolution block, a residual network and a multi-scale feature selection and fusion module, the residual network comprising a first grouped convolution block and a second grouped convolution block; obtaining to-be-recognized electrocardio data, and inputting the to-be-recognized electrocardio data into the trained ECG identity recognition network model; inputting the to-be-recognized electrocardio data into the basic convolution block to extract a basic feature map, performing grouped feature extraction on the basic feature map through the first grouped convolution block to obtain grouped features, inputting the grouped features into the multi-scale feature selection and fusion module for feature fusion to obtain fused features, inputting the fused features into the second grouped convolution block for dimension lifting to obtain a target identity label, and determining an identity recognition result according to the target identity label; the first grouped convolution block comprises a plurality of convolution kernels, and the grouped feature extraction on the basic feature map through the first grouped convolution block comprises: dividing the basic feature map into a plurality of basic sub-feature maps, wherein the number of the basic sub-feature maps is the same as the number of the convolution kernels; each convolution kernel performs convolution operation on the corresponding basic sub-feature map to obtain a plurality of grouped feature maps; splicing the plurality of grouped feature maps according to the channel dimension; performing information fusion on the spliced grouped feature maps through a channel rearrangement function to obtain grouped features; the feature fusion of the grouped features into the multi-scale feature selection and fusion module to obtain fused features comprises: extracting features from the grouped features through three convolution kernels of different sizes to obtain corresponding three feature vectors, and performing element summation on the three feature vectors to obtain a first feature; embedding global information into the first feature through a global average pooling layer to obtain channel statistics; performing linear operation on the channel statistics through a fully connected layer to obtain a second feature; performing multi-classification processing on the second feature through a softmax activation function to obtain three feature attention vectors; multiplying the elements in the three feature attention vectors respectively to obtain corresponding three fused sub-features; performing element summation on the three fused sub-features to obtain the fused features; Wherein, the calculation formula of the first feature obtained by element sum of three feature vectors is: U is the first feature, U1 is a feature vector obtained by feature extraction of the grouped feature using a 1x3 convolution kernel, U2 is a feature vector obtained by feature extraction of the grouped feature using a 1x5 convolution kernel, and U3 is a feature vector obtained by feature extraction of the grouped feature using a 1x7 convolution kernel. The formula for calculating the channel statistics is: , S is the channel statistics, denotes a global average pooling function with parameter U, H is the convolution height, W is the convolution width, c is the total number of channels, U is the first feature, i is the selected convolution height serial number, j is the selected convolution width serial number, is the domain; The calculation formula of the second feature is: , Z is the second feature, a full connection function with parameter S, δ represents a ReLU function, B is a batch normalization function, S is the channel statistics, W is the convolution width, c is the total number of channels, d is the number of channels after linear compression, is the domain; the calculation formula of the three feature attention vectors is: ， , wherein a represents a feature attention vector of U1, b represents a feature attention vector of U2, and c represents a feature attention vector of U3. 2.The ECG identity recognition method based on multi-scale feature selection fusion residual network according to claim 1, characterized in that, the method for obtaining sample electrocardio data and pre-processing the sample electrocardio data to obtain fixed-length electrocardio data comprises: The sample electrocardio data is decomposed by a Mallet algorithm to obtain first wavelet coefficients of each layer; The first wavelet coefficients of each layer are input into a dynamic soft threshold function to obtain second wavelet coefficients through noise coefficient discarding processing; Wavelet reconstruction is performed according to the second wavelet coefficients to obtain denoised electrocardio data; The denoised electrocardio data is subjected to sampling frequency normalization to obtain resampled electrocardio data; R-wave peaks in the resampled electrocardio data are located by a Pan-Tompkins algorithm; The resampled electrocardio data is subjected to heartbeat segmentation with the R-wave peaks as reference points; The resampled electrocardio data after heartbeat segmentation is normalized by a min-max normalization formula to obtain fixed-length electrocardio data. 3.The ECG identity recognition method based on multi-scale feature selection fusion residual network according to claim 2, characterized in that, The first wavelet coefficients of each layer are compared with preset critical threshold values of the corresponding layers, and when the first wavelet coefficients are greater than or equal to the critical threshold values, the first wavelet coefficients of the corresponding layers are subjected to dynamic soft threshold calculation to obtain the second wavelet coefficients. The resampled electrocardio data is input into a band-pass filter for signal filtering processing; wherein the calculation formula for obtaining the second wavelet coefficient is: , Meanwhile, when the condition is satisfied, for the second wavelet coefficients, for the first wavelet coefficients, for the critical threshold, j is the wavelet decomposition scale, is a general threshold. 4.The ECG identity recognition method based on multi-scale feature selection fusion residual network according to claim 2, characterized in that, The resampled electrocardio data is subjected to differential equation enhancement to obtain a differential signal; The differential signal is subjected to nonlinear operation to obtain an integral signal; When the differential signal satisfies a preset first signal threshold and the integral signal satisfies a preset second signal threshold, the maximum value in the QRS complex is determined as an R-wave peak. The fixed-length electrocardio data is divided into multiple data sets; One data set is selected from the multiple data sets as the test set, and the remaining data sets are determined as the training set; 5.The ECG identity recognition method based on multi-scale feature selection fusion residual network according to claim 1, characterized in that, The ECG identity recognition network model is trained by the training set and tested by the trained ECG identity recognition network model to obtain a test sub-result, until all test sub-results obtained by taking each data set as the test set are obtained; An average value of all test sub-results is obtained, and the average value is determined as a target test result of the ECG identity recognition network model; The target test result is compared with a preset model standard value, and when the target test result is greater than the model standard value, it is determined that the ECG identity recognition network model is trained. ​ ​ ​ 6. An ECG identity recognition device based on a multi-scale feature selection fusion residual network, characterized in that, The ECG identity recognition device based on the multi-scale feature selection and fusion residual network is provided with a basic convolution block, a residual network and a multi-scale feature selection and fusion module, the residual network comprises a first grouped convolution block and a second grouped convolution block, and the device is used for: Obtaining sample electrocardio data, pre-processing the sample electrocardio data to obtain fixed-length electrocardio data, the pre-processing comprising wavelet denoising, R-point positioning, heart beat segmentation and normalization; Dividing the fixed-length electrocardio data to obtain a training set and a test set; Inputting the training set into a preset ECG identity recognition network model for training, inputting the test set into the trained ECG identity recognition network model for testing, and determining a final ECG identity recognition network model according to the test result, the ECG identity recognition network model comprising a basic convolution block, a residual network and a multi-scale feature selection and fusion module, the residual network comprising a first grouped convolution block and a second grouped convolution block; Obtaining to-be-recognized electrocardio data, and inputting the to-be-recognized electrocardio data into the trained ECG identity recognition network model; Inputting the to-be-recognized electrocardio data into the basic convolution block to extract a basic feature map, performing grouped feature extraction on the basic feature map through the first grouped convolution block to obtain grouped features, inputting the grouped features into the multi-scale feature selection and fusion module for feature fusion to obtain fusion features, inputting the fusion features into the second grouped convolution block for dimension lifting to obtain a target identity label, and determining an identity recognition result according to the target identity label; The first grouped convolution block comprises a plurality of convolution kernels, and the grouped feature extraction on the basic feature map through the first grouped convolution block comprises: Dividing the basic feature map into a plurality of basic sub-feature maps, wherein the number of the basic sub-feature maps is the same as the number of the convolution kernels; Each convolution kernel respectively performs convolution operation on the corresponding basic sub-feature map to obtain a plurality of grouped feature maps; Splicing the plurality of grouped feature maps according to the channel dimension; Performing information fusion on the spliced grouped feature maps through a channel rearrangement function to obtain grouped features; The feature fusion of the grouped features into the multi-scale feature selection and fusion module to obtain fusion features comprises: Extracting features from the grouped features through three convolution kernels of different sizes to obtain three corresponding feature vectors, and performing element summation on the three feature vectors to obtain a first feature; Embedding global information into the first feature through a global average pooling layer to obtain channel statistics; Performing linear operation on the channel statistics through a full connection layer to obtain a second feature; Performing multi-classification processing on the second feature through a softmax activation function to obtain three feature attention vectors; Multiplying the elements in the three feature attention vectors respectively to obtain three corresponding fusion sub-features; Performing element summation on the three fusion sub-features to obtain the fusion features; Wherein, the calculation formula of the first feature obtained by element sum of three feature vectors is: U is the first feature, U1 is a feature vector obtained by feature extraction of the grouped feature using a 1x3 convolution kernel, U2 is a feature vector obtained by feature extraction of the grouped feature using a 1x5 convolution kernel, and U3 is a feature vector obtained by feature extraction of the grouped feature using a 1x7 convolution kernel. The formula for calculating the channel statistics is: , S is the channel statistics, is a global average pooling function with parameter U, H is the convolution height, W is the convolution width, c is the total number of channels, U is the first feature, i is the selected convolution height sequence number, j is the selected convolution width sequence number, is a domain; The calculation formula of the second feature is: , Z is the second feature, a full connection function with parameter S, δ represents a ReLU function, B is a batch normalization function, S is the channel statistics, W is the convolution width, c is the total number of channels, and d is the number of channels after linear compression, is the domain; The calculation formula of the three feature attention vectors is: ， , Wherein a represents the characteristic attention vector of U1, b represents the characteristic attention vector of U2, and c represents the characteristic attention vector of U3.

7. An electronic device, comprising: The ECG identity recognition device based on the multi-scale feature selection and fusion residual network of claim 6 is included.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer executable instructions for causing a computer to execute the ECG identity recognition method based on the multi-scale feature selection and fusion residual network of any one of claims 1 to 5.

Images