CN110522440A

CN110522440A - Electrocardiosignal identification device based on grouping convolutional neural networks

Info

Publication number: CN110522440A
Application number: CN201910740289.3A
Authority: CN
Inventors: 王红梅
Original assignee: Guangzhou Shiyuan Electronics Thecnology Co Ltd
Current assignee: Guangzhou Shiyuan Electronics Thecnology Co Ltd
Priority date: 2019-08-12
Filing date: 2019-08-12
Publication date: 2019-12-03
Anticipated expiration: 2039-08-12
Also published as: WO2021027224A1; CN110522440B

Abstract

This application involves a kind of electrocardiosignal identification device, computer equipment and storage mediums based on grouping convolutional neural networks.Described device is used for: being obtained multi-lead electrocardiosignal, and is split the multi-lead electrocardiosignal, obtains independent leads electrocardiosignal；The independent leads electrocardiosignal is input to the convolution block respectively, obtains convolution feature；The convolution feature is grouped, obtains grouping feature, and the grouping feature is input to the grouping convolution block, obtains grouping convolution feature；The grouping convolution feature is combined, electrocardiosignal assemblage characteristic is obtained, and full connection processing is carried out to the electrocardiosignal assemblage characteristic, obtains heart infarction exception probability；According to the heart infarction exception probability, determine the multi-lead electrocardiosignal for heart infarction signal.Being able to solve current electrocardiosignal recognition methods using the present apparatus has heart infarction anomalous identification inaccuracy.

Description

Electrocardiosignal identification device based on grouping convolutional neural networks

Technical field

This application involves electrocardiosignals to identify field, believes more particularly to a kind of electrocardio based on grouping convolutional neural networks Number recognition methods, device, computer equipment and storage medium.

Background technique

Coronary heart disease (Coronary Heart Disease, CHD) is the number one killer of modern society's human health.Cardiac muscle stalk It is extremely coronary heart disease severest consequences.The heart infarction risk of patient is usually predicted by way of identifying electrocardiosignal at present.

Common electrocardiosignal recognition methods is mainly based upon the critical point detection of electrocardiosignal.For example, extracting electrocardio letter Number ST section, T wave, R wave feature detected.

However, the above method depends critically upon the detection to key points such as Q wave, P wave, J point, S point, T waves.Work as electrocardiosignal When quality is bad, possibly key point can not be accurately positioned, so that heart infarction risk can not be identified accurately from electrocardiosignal.

Therefore, heart infarction anomalous identification inaccuracy is had in current electrocardiosignal recognition methods.

Summary of the invention

Based on this, it is necessary in view of the above technical problems, provide a kind of electrocardiosignal based on grouping convolutional neural networks Recognition methods, device, computer equipment and storage medium.

In a first aspect, a kind of electrocardiosignal identification device based on grouping convolutional neural networks is provided, the grouping volume Product neural network includes convolution block and grouping convolution block, which includes:

Signal acquisition module for obtaining multi-lead electrocardiosignal, and splits the multi-lead electrocardiosignal, obtains multiple groups Independent leads electrocardiosignal；

Convolution module obtains corresponding for independent leads electrocardiosignal described in every group to be input to the convolution block respectively The convolution feature of the independent leads electrocardiosignal of group；

Feature grouping module is grouped for the convolution feature to every group of independent leads electrocardio electric signal, obtains The grouping feature of this group of independent leads electrocardio electric signal, and the grouping feature is input to the grouping convolution block, it is somebody's turn to do Multiple grouping convolution features of group independent leads electrocardiosignal；

Feature combination module is obtained for combining multiple grouping convolution features of every group of independent leads electrocardiosignal The electrocardiosignal assemblage characteristic of this group of independent leads electrocardiosignal；

Full connection processing module, carries out complete for the electrocardiosignal assemblage characteristic to each group independent leads electrocardiosignal Connection processing, obtains heart infarction exception probability；

Determination module, for determining the multi-lead electrocardiosignal for heart infarction signal according to the heart infarction exception probability.

In another embodiment, the feature grouping module, comprising:

Group number acquisition submodule, for obtaining the grouping convolution group number of the grouping convolution block；

It is grouped submodule, for the convolution according to the grouping convolution group number to every group of independent leads electrocardio electric signal Feature carries out impartial grouping, obtains the grouping feature of this group of independent leads electrocardio electric signal.

In another embodiment, the grouping convolution block includes grouping convolutional layer and the maximum pond layer of grouping, the spy Levy grouping module, comprising:

Be grouped convolution feature output sub-module, for by the grouping convolutional layer, to grouping feature progress convolution, Normalization and activation are criticized, grouping convolutional layer output feature is obtained；

Pond beggar's module, for being carried out most to grouping convolutional layer output feature by the maximum pond layer of the grouping Great Chiization obtains multiple grouping convolution features of this group of independent leads electrocardiosignal.

In another embodiment, the convolution block of the grouping convolutional neural networks includes first volume block and the second convolution Block, the convolution module, comprising:

First convolution submodule, for carrying out convolution to the independent leads electrocardiosignal by the first volume block With pond, convolution block output feature is obtained；

Second convolution submodule, for exporting special progress convolution sum pond to the convolution block by the volume Two block Change, obtains the convolution feature.

In another embodiment, the first convolution submodule, is specifically used for:

By first convolutional layer, convolution, batch normalization and activation are carried out to the independent leads electrocardiosignal, obtained First convolutional layer exports feature；By the described first maximum pond layer, maximum pond is carried out to first convolutional layer output feature Change, obtains the convolution block output feature.

In another embodiment, the volume Two block includes the second convolutional layer and the second maximum pond layer, and described the Two convolution submodules, are specifically used for:

By second convolutional layer, convolution, batch normalization and activation are carried out to convolution block output feature, obtain the Two convolutional layers export feature；By the described second maximum pond layer, maximum pond is carried out to second convolutional layer output feature, Obtain the convolution feature.

In another embodiment, the signal acquisition module, comprising:

Original signal receiving submodule, for receiving original signal；

Wavelet decomposition submodule obtains wavelet decomposition signal for carrying out wavelet decomposition to the original signal；It is described small Wave Decomposition signal is tieed up with X1；

Zero setting submodule obtains part zero setting signal for the signal zero setting to the X2 dimension in the wavelet decomposition signal； Wherein, X2 < X1；

Inverse transformation submodule obtains denoised signal for carrying out wavelet inverse transformation to the part zero setting signal；It is described to go Noise cancellation signal is the signal after high-frequency noise and baseline drift removal；

Multi-lead acquisition submodule, for obtaining the multi-lead electrocardiosignal according to the denoised signal.

In another embodiment, the multi-lead acquisition submodule, is specifically used for:

Determine the R wave position of the denoised signal；

Determine the preceding M1 position of R wave position, and, determine the rear M2 position of R wave position；

Using the denoised signal on R wave position, the preceding M1 position, the rear M2 position, structuring is formed Signal matrix, as the multi-lead electrocardiosignal.

Second aspect provides a kind of training method for being grouped convolutional neural networks, the grouping convolutional neural networks packet Convolution block and grouping convolution block are included, this method comprises:

Obtain the electrocardiosignal training sample for the grouping convolutional neural networks；

Machine training is carried out to the grouping convolutional neural networks using the electrocardiosignal training sample, after being trained It is grouped convolutional neural networks；It includes the convolution block trained in advance and grouping convolution block that convolutional neural networks are grouped after the training； Convolutional neural networks are grouped after the training, for every group of independent leads electrocardiosignal to be input to the convolution block respectively, are obtained To the convolution feature of the independent leads electrocardiosignal of respective sets；The independent leads electrocardiosignal is to split multi-lead electrocardiosignal It obtains；It is also used to be grouped the convolution feature of every group of independent leads electrocardio electric signal, obtains this group of independent leads The grouping feature of electrocardio electric signal, and the grouping feature is input to the grouping convolution block, obtain this group of independent leads heart Multiple grouping convolution features of electric signal；The multiple grouping convolution for being also used to combine every group of independent leads electrocardiosignal are special Sign, obtains the electrocardiosignal assemblage characteristic of this group of independent leads electrocardiosignal；It is also used to believe each group independent leads electrocardio Number electrocardiosignal assemblage characteristic carry out full connection processing, obtain heart infarction exception probability；It is also used to extremely general according to the heart infarction Rate determines the multi-lead electrocardiosignal for heart infarction signal.

The third aspect provides a kind of electrocardiosignal recognition methods for being grouped convolutional neural networks, the grouping convolution mind Through network include convolution block and grouping convolution block, this method comprises:

Multi-lead electrocardiosignal is obtained, and splits the multi-lead electrocardiosignal, obtains multiple groups independent leads electrocardiosignal；

Independent leads electrocardiosignal described in every group is input to the convolution block respectively, obtains the independent leads heart of respective sets The convolution feature of electric signal；

The convolution feature of every group of independent leads electrocardio electric signal is grouped, this group of independent leads electrocardio electricity is obtained The grouping feature of signal, and the grouping feature is input to the grouping convolution block, obtain this group of independent leads electrocardiosignal Multiple grouping convolution features；

The multiple grouping convolution features for combining every group of independent leads electrocardiosignal obtain this group of independent leads electrocardio letter Number electrocardiosignal assemblage characteristic；

Full connection processing is carried out to the electrocardiosignal assemblage characteristic of each group independent leads electrocardiosignal, it is different to obtain heart infarction Normal probability；

According to the heart infarction exception probability, determine the multi-lead electrocardiosignal for heart infarction signal.

Fourth aspect provides a kind of training device for being grouped convolutional neural networks, the grouping convolutional neural networks packet Convolution block and grouping convolution block are included, which includes:

Training sample obtains module, for obtaining the electrocardiosignal training sample for being directed to the grouping convolutional neural networks；

Machine training module, for carrying out machine to the grouping convolutional neural networks using the electrocardiosignal training sample Device training, is grouped convolutional neural networks after being trained；It includes the volume trained in advance that convolutional neural networks are grouped after the training Block and grouping convolution block；Convolutional neural networks are grouped after the training, for respectively that every group of independent leads electrocardiosignal is defeated Enter the convolution feature that the independent leads electrocardiosignal of respective sets is obtained to the convolution block；The independent leads electrocardiosignal is Split what multi-lead electrocardiosignal obtained；It is also used to divide the convolution feature of every group of independent leads electrocardio electric signal Group obtains the grouping feature of this group of independent leads electrocardio electric signal, and the grouping feature is input to the grouping convolution block, Obtain multiple grouping convolution features of this group of independent leads electrocardiosignal；It is also used to combine every group of independent leads electrocardiosignal Multiple grouping convolution features, obtain the electrocardiosignal assemblage characteristic of this group of independent leads electrocardiosignal；It is also used to described each The electrocardiosignal assemblage characteristic of group independent leads electrocardiosignal carries out full connection processing, obtains heart infarction exception probability；It is also used to root According to the heart infarction exception probability, determine the multi-lead electrocardiosignal for heart infarction signal.

5th aspect, provides a kind of electronic equipment characterized by comprising memory, one or more processors；

The memory, for storing one or more programs；

When one or more of programs are executed by one or more of processors, so that one or more of processing Device realizes following operation:

6th aspect, provides a kind of computer readable storage medium, is stored thereon with computer program, the computer It is performed the steps of when program is executed by processor

A kind of above-mentioned electrocardiosignal recognition methods, device, computer equipment and storage based on grouping convolutional neural networks Medium by obtaining multi-lead electrocardiosignal, and splits multi-lead electrocardiosignal, obtains independent leads electrocardiosignal；Then, first The independent leads electrocardiosignal is input in convolution block, convolution feature is obtained；Then, it is grouped to the convolution feature, Grouping feature is obtained, and grouping feature is input to grouping convolution block, obtains grouping convolution feature；Then, according to the grouping Convolution feature determines heart infarction exception probability；Finally, determining that multi-lead electrocardiosignal is heart infarction signal according to heart infarction exception probability； In this way, the parameter amount of convolution operation can be reduced, in the case where not reducing the recognition performance of grouping convolutional neural networks, subtract The light over-fitting of the grouping convolutional neural networks, so as to more accurately identify judge multi-lead electrocardiosignal whether be Heart infarction signal；To when identifying electrocardiosignal, without dependent on to electrocardiosignal key point Q wave, P wave, J point, S point, T wave Accurate positionin, even if electrocardiosignal quality it is bad, electrocardiosignal key point can not be accurately positioned in the case where, pass through input The grouping convolutional neural networks of the application, so as to more accurately identify heart infarction risk from electrocardiosignal.

Detailed description of the invention

Fig. 1 is a kind of electrocardiosignal identification device based on grouping convolutional neural networks that the embodiment of the present application one provides Structural schematic diagram；

Fig. 2 is a kind of nerve for electrocardiosignal recognition methods based on grouping convolutional network that the embodiment of the present application one provides Schematic network structure；

Fig. 3 is a kind of electrocardiosignal identification device based on grouping convolutional neural networks that the embodiment of the present application two provides Structural schematic diagram；

Fig. 4 A is a kind of schematic diagram of original electro-cardiologic signals in one embodiment；

Fig. 4 B is a kind of schematic diagram of denoised signal in one embodiment；

Fig. 5 is the network knot of electrocardiosignal recognition methods of one of the one embodiment based on grouping convolutional neural networks Structure schematic diagram；

Fig. 6 is a kind of characteristic dimension variation of the electrocardiosignal identification based on grouping convolutional neural networks in one embodiment Schematic diagram；

Fig. 7 is the flow chart that electrocardiosignal identification is carried out based on neural network；

Fig. 8 is a kind of process for electrocardiosignal method based on grouping convolutional neural networks that the embodiment of the present application four provides Figure；

Fig. 9 is the structural schematic diagram for a kind of electronic equipment that the embodiment of the present application six provides；

Figure 10 is a kind of flow chart of the training method for grouping convolutional neural networks that the embodiment of the present application three provides；

Figure 11 is a kind of structural schematic diagram of the training device for grouping convolutional neural networks that the embodiment of the present application five provides.

Specific embodiment

It is with reference to the accompanying drawings and embodiments, right in order to which the objects, technical solutions and advantages of the application are more clearly understood The application is further elaborated.It should be appreciated that specific embodiment described herein is only used to explain the application, not For limiting the application.

Embodiment one

Fig. 1 is a kind of electrocardiosignal identification device based on grouping convolutional neural networks that the embodiment of the present application one provides Structural schematic diagram.Specifically, with reference to Fig. 1, the electrocardiosignal identification based on grouping convolutional neural networks of the embodiment of the present application one Device, grouping convolutional neural networks include the convolution block trained in advance and grouping convolution block, are specifically included:

Signal acquisition module 110 for obtaining multi-lead electrocardiosignal, and splits multi-lead electrocardiosignal, obtains multiple groups Independent leads electrocardiosignal.

Wherein, multi-lead electrocardiosignal can be the signal data matrix of characterization multi-lead electrocardiosignal.Multi-lead electrocardio Signal can be the collected signal of multi-lead Electrocardiograph.

Wherein, independent leads electrocardiosignal can refer to the electrocardio of each independent leads in above-mentioned multi-lead electrocardiosignal Signal.

In the specific implementation, original signal can be acquired, by the pre- place for carrying out wavelet transformation, denoising etc. to original signal Reason, obtains above-mentioned multi-lead electrocardiosignal.Then, then by above-mentioned multi-lead electrocardiosignal split, it is only to obtain multiple groups Vertical lead electrocardiosignal.

In practical application, multi-lead electrocardiosignal can be acquired by multi-lead Electrocardiograph.Relatively conventional at present is more Lead Electrocardiograph is 12 lead.Wherein, the multi-lead electrocardiosignal of 12 lead, include lead signals V1, V2, V3, V4, V5, V6, aVF, aVR, aVL, I, II and III.In other words, the multi-lead electrocardiosignal of the embodiment of the present application, can refer to ten Two or more lead signals in the multi-lead electrocardiosignal of two leads.Then, by the multi-lead electrocardiosignal Progress is split according to respective lead type, obtains the independent leads electrocardiosignal of multiple groups, independent leads electrocardiosignal Group number can be Z.For example, the multi-lead electrocardiosignal of 12 lead is split, 12 groups of independent leads electrocardiosignals are obtained, i.e., Z is 12.

Convolution module 120 obtains the only of respective sets for every group of independent leads electrocardiosignal to be input to convolution block respectively The convolution feature of vertical lead electrocardiosignal.

Wherein, convolution block (Basic Convolutional Block) can be one-dimensional for carrying out to the feature of input A series of set of operations such as convolution, batch normalization, activation and pond.According to work of this series of operation in neural network With being named as convolution block.

Wherein, criticizing normalization can also be by GroupNormalization (group normalizes), Instance Normalization (example regularization), Layer Normalization (layer standardization) scheduling algorithm replace.

Wherein, common activation primitive includes ReLU (a kind of activation primitive), ELU (a kind of activation primitive), SELU (one kind Activation primitive), Sigmoid (a kind of activation primitive), tanh (a kind of activation primitive) etc..

Wherein, pondization can refer to a kind of operation by high dimensional feature dimensionality reduction at low-dimensional feature and the feature of getting rid of redundancy Method.

Wherein, pond, which can be, maximizes at least one of pond algorithm and average pond algorithm.

In the specific implementation, will be upper after obtaining multiple groups independent leads electrocardiosignal splitting above-mentioned multi-lead electrocardiosignal It states each group independent leads electrocardiosignal to be separately input into corresponding convolution block, using independent leads electrocardiosignal as convolution block Input, convolution block carry out convolution and export, and the data of output are special as the convolution for the independent leads electrocardiosignal for obtaining respective sets Sign.For example, above-mentioned grouping convolutional network has Z signal input channel, above-mentioned Z group independent leads electrocardio is believed respectively Number, it inputs in corresponding signal input channel, and then realize and Z group independent leads electrocardiosignal is input to corresponding convolution block, Obtain every group of convolution feature.In practical application, one-dimensional convolution can be carried out to the signal of input by convolution block, by one-dimensional volume Feature after product carries out batch normalization, is activated finally by activation primitive so that have to the expression of feature it is non-linear, no It is again only 0 or 1 output, to improve the ability to express of model.The feature of output is subjected to pond, i.e., by high dimensional feature Dimensionality reduction is at low-dimensional feature and gets rid of the feature of redundancy, finally obtains convolution feature.In above-mentioned grouping convolutional neural networks, Convolution block can be one, or multiple, those skilled in the art can be designed convolution block according to actual needs Quantity.

It should be noted that each group independent leads ECG's data compression process is similar to the above embodiments, herein no longer It repeats.

Feature grouping module 130 is grouped for the convolution feature to every group of independent leads electrocardio electric signal, is somebody's turn to do The grouping feature of group independent leads electrocardio electric signal, and grouping feature is input to grouping convolution block, obtain this group of independent leads Multiple grouping convolution features of electrocardiosignal.

Wherein, grouping feature can refer to the feature that feature grouping is carried out to convolution feature.

Wherein, grouping convolution block can refer to the convolution block that convolution algorithm is carried out to above-mentioned grouping feature.In practical application In, above-mentioned grouping convolution block can be for carrying out a series of operations such as one-dimensional convolution, pond to the grouping feature of input Set.

In the specific implementation, according to the packet count of setting, to every group of independent leads electrocardio electric signal of above-mentioned convolution block output Convolution feature carry out impartial grouping, obtain the grouping feature of this group of independent leads electrocardio electric signal；Meanwhile the grouping that will be obtained Feature is input in above-mentioned grouping convolution block, and using grouping feature as the input of grouping convolution block, grouping convolution block is rolled up It accumulates and exports, multiple grouping convolution features of the data of output as this group of independent leads electrocardiosignal.

In practical applications, one-dimensional convolution can be carried out to the grouping feature of input by grouping convolution block, by one-dimensional volume Feature after product carries out pond, and high dimensional feature dimensionality reduction at low-dimensional feature and is got rid of the feature of redundancy, finally obtains grouping volume Product feature.

For example, as it is known that preset packet count is G, above-mentioned convolution feature has M characteristic face, convolution feature is divided Group obtains the grouping feature that group number is G；Wherein, every group of above-mentioned grouping feature all has M/G characteristic face.

It should be noted that by the corresponding convolution characteristic processing process of each independent leads electrocardiosignal and above-mentioned implementation Example is similar, and details are not described herein.

Feature combination module 140 is somebody's turn to do for combining multiple grouping convolution features of every group of independent leads electrocardiosignal The electrocardiosignal assemblage characteristic of group independent leads electrocardiosignal.

Wherein, electrocardiosignal assemblage characteristic, which can refer to be combined by above-mentioned each grouping convolution feature, obtains feature.

In the specific implementation, after obtaining multiple grouping convolution features of every group of independent leads electrocardiosignal, by each group independence Multiple grouping convolution features of lead electrocardiosignal are combined, to obtain the electrocardio of this group of independent leads electrocardiosignal Signal assemblage characteristic.

For example, as it is known that Z group independent leads electrocardiosignal is corresponding with Z component group convolution feature, each grouping convolution feature Intrinsic dimensionality is A*B, then combines the electrocardiosignal that the group number that above-mentioned grouping convolution feature is combined by feature is 1 Feature.Wherein, the intrinsic dimensionality of the electrocardiosignal assemblage characteristic after connection is (ZA) * B.

Full connection processing module 150, carries out complete for the electrocardiosignal assemblage characteristic to each group independent leads electrocardiosignal Connection processing, obtains heart infarction exception probability.

Wherein, full connection processing can be finger and be handled using full Connection Neural Network classifier.

In the specific implementation, multiple grouping convolution features of each group independent leads electrocardiosignal are combined to obtain electrocardiosignal After assemblage characteristic, above-mentioned electrocardiosignal assemblage characteristic is input in full Connection Neural Network classifier, uses full connection Neural network classifier carries out full connection processing to electrocardiosignal assemblage characteristic, obtains heart infarction exception probability.

Determination module 160, for determining that multi-lead electrocardiosignal is heart infarction signal according to heart infarction exception probability.

In the specific implementation, the input cell number of above-mentioned full Connection Neural Network classifier and electrocardiosignal assemblage characteristic Feature vector number is equal, and the output cell number of full Connection Neural Network classifier is 2, and then represents two kinds of prediction results.I.e. The predicted value that each input heart is clapped can be obtained, it is complete to connect when obtained heart infarction exception probability is higher than preset abnormal probability threshold value The predicted value of neural network classifier output is 1, and representing this heart bat sample has the performance of heart infarction relevant abnormalities；When obtained heart infarction is different When normal probability is lower than preset abnormal probability threshold value, the predicted value of full Connection Neural Network classifier output is 0, represents the bat of this heart Sample health.

It should be noted that the grouping convolutional neural networks stated in use carry out spy to above-mentioned multi-lead electrocardiosignal Before sign identification, need to lead using the various multi-lead electrocardiosignals with abnormal signal and known heart infarction type with normally more Join electrocardiosignal as training sample, above-mentioned grouping convolutional neural networks are trained, optimizes above-mentioned heart infarction identification mind Through network.

In practical application, it can be trained and test by the public database of such as PTB etc..More specifically, can incite somebody to action Patients with myocardial infarction and non-salary motivation patient data collection, are randomly divided into training set and test set, two datasets are not in proportion It simultaneously include same person's data.The multi-lead electrocardiosignal of structuring is labeled as X, it will " there are the relevant characteristics of heart infarction The exception of variation ", " there is no the exceptions of the relevant characteristic variation of the heart infarction " output of label as grouping convolutional neural networks Y.(X, the Y) of training set collectively constitutes the training sample of the more topology convergence networks of multi-lead.X is by certain batch size by batch input Convolutional neural networks are grouped, the predicted value Pred_Y of Y is obtained by propagated forward, Y and Pred_Y is calculated by loss function and is damaged It loses, backpropagation will be lost, using gradient descent method training network, obtain optimal grouping convolutional neural networks.

The embodiment of the present application is deeply understood for the ease of those skilled in the art, is carried out below with reference to a specific example Explanation.

Fig. 2 is a kind of neural network structure of electrocardiosignal identification device based on grouping convolutional network of one embodiment Schematic diagram.As shown in Fig. 2, obtaining multi-lead electrocardiosignal first, and above-mentioned multi-lead electrocardiosignal is split, obtains multiple groups independence Lead electrocardiosignal；Then, every group of above-mentioned independent leads electrocardiosignal is input to convolution block respectively, obtains the only of respective sets The convolution feature of vertical lead electrocardiosignal；Subsequently, the convolution feature of every group of independent leads electrocardio electric signal is grouped, is obtained It is input to grouping convolution block to the grouping feature of this group of independent leads electrocardio electric signal, and by grouping feature, is rolled up as the grouping The input of block obtains multiple grouping convolution features of this group of independent leads electrocardiosignal；Subsequently, every group of independent leads are combined The grouping convolution feature of multiple grouping convolution blocks output of electrocardiosignal, obtains the electrocardiosignal of this group of independent leads electrocardiosignal Assemblage characteristic, and full connection processing is carried out to the electrocardiosignal assemblage characteristic of each group independent leads electrocardiosignal, it is different to obtain heart infarction Normal probability；Finally, determining that multi-lead electrocardiosignal is heart infarction signal further according to heart infarction exception probability.

In the above-mentioned electrocardiosignal identification device based on grouping convolutional neural networks, by obtaining multi-lead electrocardiosignal, And multi-lead electrocardiosignal is split, obtain independent leads electrocardiosignal；Then, the independent leads electrocardiosignal is first input to volume In block, convolution feature is obtained；Then, it is grouped to the convolution feature, obtains grouping feature, and grouping feature is inputted To grouping convolution block, grouping convolution feature is obtained；Then, heart infarction exception probability is being determined according to the grouping convolution feature；Finally, According to heart infarction exception probability, determine that multi-lead electrocardiosignal is heart infarction signal；In this way, the parameter of convolution operation can be reduced Amount alleviates the excessively quasi- of the grouping convolutional neural networks in the case where not reducing the recognition performance of grouping convolutional neural networks It closes, judges whether multi-lead electrocardiosignal is heart infarction signal so as to more accurately identify；To in identification electrocardiosignal When, without the accurate positionin depended on to electrocardiosignal key point Q wave, P wave, J point, S point, T wave, even if in electrocardiosignal quality It is bad, electrocardiosignal key point can not be accurately positioned in the case where, by inputting the grouping convolutional neural networks of the application, thus Heart infarction risk can be more accurately identified from electrocardiosignal.

Embodiment two

Fig. 3 is a kind of electrocardiosignal identification device based on grouping convolutional neural networks that the embodiment of the present application two provides Structural schematic diagram.Specifically, with reference to Fig. 3, the electrocardiosignal identification based on grouping convolutional neural networks of the embodiment of the present application two Method specifically includes:

Signal acquisition module 310 for obtaining multi-lead electrocardiosignal, and splits multi-lead electrocardiosignal, obtains multiple groups Independent leads electrocardiosignal.

Optionally, the group number of independent leads electrocardiosignal can be Z.More specifically, Z=12.12 groups of independent leads electrocardios Signal can be V1 lead electrocardiosignal, V2 lead electrocardiosignal, V3 lead electrocardiosignal, V4 lead electrocardiosignal, V5 lead Electrocardiosignal, V6 lead electrocardiosignal, aVF lead electrocardiosignal, aVR lead electrocardiosignal, aVL lead electrocardiosignal, I lead Electrocardiosignal, II lead electrocardiosignal and III lead electrocardiosignal.

Optionally, above-mentioned signal acquisition module 310, comprising:

Original signal receiving submodule, for receiving original signal；Wavelet decomposition submodule, for the original signal Wavelet decomposition is carried out, wavelet decomposition signal is obtained；The wavelet decomposition signal is tieed up with X1；Zero setting submodule, for described The signal zero setting of X2 dimension in wavelet decomposition signal, obtains part zero setting signal；Wherein, X2 < X1；Inverse transformation submodule, is used for Wavelet inverse transformation is carried out to the part zero setting signal, obtains denoised signal；The denoised signal is that high-frequency noise and baseline float Remove the signal after removing；Multi-lead acquisition submodule, for obtaining the multi-lead electrocardiosignal according to the denoised signal.

Wherein, original signal can be the collected original signal of multi-lead Electrocardiograph.

Wherein, wavelet decomposition signal can be for obtained signal after original signal progress wavelet decomposition.

Wherein, the signal that zero setting signal in part can be zeroed out for the signal of partial dimensional.It, can be with after wavelet decomposition The wavelet decomposition signal of X1 dimension is decomposited, the signal zero setting to wherein X2 dimension has obtained part zero setting signal.

In the specific implementation, can be to the signal of original signal progress resampling to certain frequency, for example, being resampled to The signal of 1000Hz.

Then, using the wavelet basis function of certain db (power gain unit), X1 is carried out to the signal of resampling and ties up small echo It decomposes, obtains the wavelet decomposition signal of X1 dimension.For example, can preferably 6db wavelet basis function carry out wavelet decomposition.

Zero setting is carried out to the X2 dimension wavelet decomposition signal in X1 dimension, obtains part zero setting signal.For example, when X1 is 10, X2 It can be 3, specifically zero setting can be carried out to the wavelet decomposition signal of the 0th dimension, the 9th dimension, the 10th dimension.

After obtaining part zero setting signal, zero setting signal in part can be converted, obtained by way of wavelet inverse transformation Signal, as denoised signal, denoised signal eliminates high-frequency noise and baseline drift, can finally be based on the denoised signal, obtain To multi-lead electrocardiosignal.

Fig. 4 A is a kind of schematic diagram of original electro-cardiologic signals of one embodiment.Fig. 4 B is a kind of denoising of one embodiment The schematic diagram of signal.As shown, X-axis and Y-axis respectively indicate acquisition time (second, s) and signal strength (mV, the milli of signal Volt), original electro-cardiologic signals are compared with denoised signal as it can be seen that the signal base line of denoised signal becomes to tend to be smooth, after being more advantageous to The extraction and detection of continuous feature.

It is inverse by progress wavelet decomposition, the signal zero setting of partial dimensional, small echo according to the technical solution of the embodiment of the present application The preprocessing means such as transformation have obtained the denoised signal of removal high-frequency noise and baseline drift, obtain lead based on denoised signal more Join electrocardiosignal, signal quality more preferably multi-lead electrocardiosignal can be obtained to avoid the interference of high-frequency noise and baseline drift, Improve the accuracy of electrocardiosignal identification.

Optionally, above-mentioned multi-lead acquisition submodule, is used for, specifically for the R wave position of the determination denoised signal； Determine the preceding M1 position of R wave position, and, determine the rear M2 position of R wave position；Using R wave position, Denoised signal on the preceding M1 position, the rear M2 position, forms structured signal matrix, as the multi-lead heart Electric signal.

Wherein, the position that R wave position can occur for R wave maximum value in signal.

Wherein, structured signal matrix can be the matrix formed by the value arrangements of characterization signal.

In the specific implementation, a kind of improved Pan-Tompkins (algorithm for detecting QRS complex) algorithm, detection can be passed through The R wave position of each denoised signal out.Wherein, Pan-Tompkins algorithm can specifically include low-pass filtering, high-pass filtering, micro- Point, square, integral, the calculating processes such as adaptive threshold and search.

Then, on the basis of each R wave position, the preceding M1 position and rear M2 position of R wave position are determined, using R wave Position, preceding M1 position, the denoised signal on rear M2 position, form a pair being made of (M1+M2+1) a denoised signal The signal data that the Ying Yuyi heart is clapped is directed to the same patient, the available signal data clapped to N number of heart, and forms knot Structure signal matrix.

Matrix structure can be N*L* (M1+M2+1), wherein L represents the quantity of lead, and the specific value of M1 and M2 can To set according to actual needs.

Convolution module 320 obtains the only of respective sets for every group of independent leads electrocardiosignal to be input to convolution block respectively The convolution feature of vertical lead electrocardiosignal.

In the specific implementation, will be upper after obtaining multiple groups independent leads electrocardiosignal splitting above-mentioned multi-lead electrocardiosignal It states each group independent leads electrocardiosignal to be separately input into corresponding convolution block, using independent leads electrocardiosignal as convolution block Input, convolution block carry out convolution and export, and the data of output are special as the convolution for the independent leads electrocardiosignal for obtaining respective sets Sign.

Optionally, the convolution block for being grouped convolutional neural networks includes first volume block and volume Two block, above-mentioned convolution Module 320, comprising:

Wherein, first volume block can be for carrying out one-dimensional convolution, batch normalization, activation and pond to the feature of input Etc. a series of set of operations.

First convolution submodule, for carrying out convolution sum pond to independent leads electrocardiosignal, obtaining by first volume block Feature is exported to convolution block；Second convolution submodule, for exporting special progress convolution sum pond to convolution block by volume Two block Change, obtains convolution feature.

In the specific implementation, splitting above-mentioned multi-lead electrocardiosignal, it, will be above-mentioned each after obtaining independent leads electrocardiosignal A independent leads electrocardiosignal is separately input into corresponding first volume block, using independent leads electrocardiosignal as the first convolution The input of block, first volume block carry out one-dimensional convolution to the independent leads electrocardiosignal of input, by the feature after one-dimensional convolution into Row batch normalization, is activated finally by activation primitive, pond is carried out in the feature that will be exported, for example, maximizing Chi Huahuo Average pond, makes high dimensional feature dimensionality reduction at low-dimensional feature and gets rid of the feature of redundancy, finally obtains convolution block output feature.

Then, convolution block output feature is input in corresponding volume Two block, using convolution block output feature as the The input of a roll of block, volume Two block carries out one-dimensional convolution to the convolution block output feature of input, by the spy after one-dimensional convolution Sign carries out batch normalization, is activated finally by activation primitive, pond is carried out in the feature that will be exported, for example, maximizing pond Change or average pond, make high dimensional feature dimensionality reduction at low-dimensional feature and get rid of the feature of redundancy, finally obtain convolution feature.Its In, activation primitive can be at least one of ReLU, ELU, SELU, Sigmoid, tanh.

It should be noted that when convolution block includes 2 even more convolution blocks, treatment process and the above embodiments phase Seemingly, details are not described herein.

Optionally, first volume block includes the first convolutional layer and the first maximum pond layer, and the first convolution submodule is specific to use In: by the first convolutional layer, convolution, batch normalization and activation are carried out to independent leads electrocardiosignal, it is defeated to obtain the first convolutional layer Feature out；By the first maximum pond layer, maximum pond is carried out to the first convolutional layer output feature, it is special to obtain the output of convolution block Sign.

Wherein, maximum pond layer (Max Pool) can be the operation for the feature maximizing pond to input.Root According to effect of the operation in neural network, it is named as maximum pond layer.

In the specific implementation, independent leads electrocardiosignal is input to the first convolutional layer, as the input of the first convolutional layer, so One-dimensional convolution is carried out by independent leads electrocardiosignal of first convolutional layer to input afterwards, the feature after one-dimensional convolution is criticized Normalization, is activated finally by activation primitive, obtains the first convolutional layer output feature.

Then, above-mentioned the first convolutional layer output feature is input to the first maximum pond layer, as the first maximum pond Then the input of layer exports feature to the first convolutional layer by the first maximum pond layer and carries out maximum pond, drops high dimensional feature It ties up into low-dimensional feature and gets rid of the feature of redundancy, finally obtain convolution block output feature.Wherein, activation primitive can be At least one of ReLU, ELU, SELU, Sigmoid, tanh.

In practical application, the characteristic dimension of independent leads electrocardiosignal can be 1*600, the structural parameters of the first convolutional layer It can be with are as follows: convolution kernel size k=61, sliding step s=1, complementary element p=0 export characteristic face number f=12, the first convolution The characteristic dimension of the first convolutional layer output feature of layer output can be 12*540.And the sliding step s of the first maximum pond layer The characteristic dimension of the convolution block output feature of=3. first maximum pond layer output can be 12*180.

The technical solution of the embodiment of the present application, by using the first convolutional layer, to independent leads electrocardiosignal carry out convolution, Normalization and activation are criticized, the first convolutional layer output feature is obtained；By the first maximum pond layer, feature is exported to the first convolutional layer Maximum pond is carried out, so that realizing reduces the characteristic dimension of independent leads electrocardiosignal to obtain convolution block output feature, And then the parameter processing amount of subsequent heart infarction signal identification is reduced, improve the recognition efficiency of heart infarction signal.

Optionally, volume Two block includes the second convolutional layer and the second maximum pond layer, the second above-mentioned convolution submodule, It is specifically used for: by the second convolutional layer, convolution, batch normalization and activation is carried out to convolution block output feature, obtain the second convolution Layer output feature；By the second maximum pond layer, maximum pond is carried out to the second convolutional layer output feature, obtains convolution feature.

In the specific implementation, convolution block output feature is input to the second convolutional layer, as the input of the second convolutional layer, then Feature is exported by convolution block of second convolutional layer to input and carries out one-dimensional convolution, and the feature after one-dimensional convolution is subjected to batch normalizing Change, activated finally by activation primitive, obtains the second convolutional layer output feature.

Then, above-mentioned the second convolutional layer output feature is input to the second maximum pond layer, as the second maximum pond Then the input of layer exports feature to the second convolutional layer by the second maximum pond layer and carries out maximum pond, drops high dimensional feature It ties up into low-dimensional feature and gets rid of the feature of redundancy, finally obtain convolution feature.Wherein, activation primitive can be ReLU, ELU, At least one of SELU, Sigmoid, tanh.

In practical application, the characteristic dimension of convolution block output feature can be 12*180.The structural parameters of second convolutional layer It can be with are as follows: convolution kernel size k=31, sliding step s=1, complementary element p=0 export characteristic face number f=16, the second convolution The characteristic dimension of the second convolutional layer output feature of layer output can be 16*150.And the sliding step s of the second maximum pond layer The characteristic dimension of the convolution feature of=3. second maximum pond layer output can be 16*50.

The technical solution of the embodiment of the present application is right by the second convolutional layer after the convolution algorithm of the first convolutional layer Convolution block exports feature and carries out convolution, batch normalization and activation, obtains the second convolutional layer output feature；Pass through the second maximum pond Layer carries out maximum pond to the second convolutional layer output feature, obtains convolution feature；To realize the spy to convolution block output feature Sign dimension further decreases, and then further reduces the parameter processing amount of subsequent heart infarction signal identification process, improves heart infarction Recognition speed；Meanwhile convolution algorithm is carried out to independent leads electrocardiosignal by using the first convolutional layer and the first convolutional layer, it can To avoid during carrying out reduction characteristic dimension, dimensionality reduction gradient is excessive and causes Character losing, and then improve heart infarction knowledge Other accuracy.

Feature grouping module 330 is grouped for the convolution feature to every group of independent leads electrocardio electric signal, is somebody's turn to do The grouping feature of group independent leads electrocardio electric signal, and grouping feature is input to grouping convolution block, obtain this group of independent leads Multiple grouping convolution features of electrocardiosignal.

Optionally, above-mentioned feature grouping module 330, comprising: group number acquisition submodule, for obtaining grouping convolution block It is grouped convolution group number；It is grouped submodule, for special according to convolution of the grouping convolution group number to every group of independent leads electrocardio electric signal Sign carries out impartial grouping, obtains the grouping feature of this group of independent leads electrocardio electric signal.

In the specific implementation, before being grouped to convolution feature, it is necessary first to obtain the grouping convolution of grouping convolution block Group number；Then, impartial grouping is being carried out to convolution feature according to group convolution group number, is obtaining grouping feature.For example, as it is known that grouping volume Product group number is 4, and convolution feature has 16 characteristic faces, then above-mentioned convolution feature is divided into four groups of grouping features；Wherein, it is grouped Feature has 4 characteristic faces.

In practical application, the grouping convolution group number of grouping convolution block is g, then the characteristic face of input feature vector is divided into g group, The characteristic face for exporting feature simultaneously is also divided into g group, wherein i-th group of feature of output passes through convolution by the i-th group of feature exported It is calculated；For example, as it is known that grouping convolution group number is 4, convolution feature has 16 characteristic faces, then by above-mentioned convolution feature point For four groups of grouping features；Wherein, grouping feature has 4 characteristic faces.Then, four groups of grouping feature inputs are grouped convolution blocks, The grouping convolution feature finally exported has 24 characteristic faces, is divided into four groups, every group of 6 characteristic faces, then the 1st group of characteristic face is by dividing First group of convolution group in group convolution block is obtained by convolutional calculation, other 3 groups and so on, details are not described herein.

Optionally, grouping convolution block includes grouping convolutional layer and the maximum pond layer of grouping, above-mentioned feature grouping module 330, comprising: grouping convolution feature output sub-module, for carrying out convolution, batch normalizing to grouping feature by grouping convolutional layer Change and activate, obtains grouping convolutional layer output feature；Pond beggar's module is grouped maximum pond layer for passing through, to grouping convolution Layer output feature carries out maximum pond, obtains multiple grouping convolution features of this group of independent leads electrocardiosignal.

In the specific implementation, grouping feature is input to grouping convolutional layer, as the input of grouping convolutional layer, then by dividing Group convolutional layer carries out one-dimensional convolution to the grouping feature of input, and the feature after one-dimensional convolution is carried out batch normalization, finally by Activation primitive is activated, and grouping convolutional layer output feature is obtained.

Then, above-mentioned grouping convolutional layer output feature is input to the maximum pond layer of grouping, as the maximum pond of grouping Then the input of layer exports feature to grouping convolutional layer by the maximum pond layer of grouping and carries out maximum pond, drops high dimensional feature It ties up into low-dimensional feature and gets rid of the feature of redundancy, finally obtain grouping convolution feature.Wherein, activation primitive can be ReLU, At least one of ELU, SELU, Sigmoid, tanh.

In practical application, the characteristic dimension of the convolution block output feature of the first maximum pond layer output can be 16*50, then The characteristic dimension of grouping feature can be 4*50.The structural parameters for being grouped convolutional layer can be with are as follows: convolution kernel size k=9, sliding step Long s=1, complementary element p=0 export characteristic face number f=24, are grouped convolution group number g=4,；It is grouped point of convolutional layer output The characteristic dimension of group convolutional layer output feature can be 24*42.And it is grouped the sliding step s=2 of maximum pond layer, grouping is maximum The characteristic dimension of the grouping convolution feature of pond layer output can be 24*14.

The technical solution of the embodiment of the present application carries out convolution, batch normalization to grouping feature and swashs by being grouped convolutional layer It is living, obtain grouping convolutional layer output feature；By being grouped maximum pond layer, maximum pond is carried out to grouping convolutional layer output feature Change, obtains multiple grouping convolution features of this group of independent leads electrocardiosignal；In this way, to be grouped convolution block in progress convolution fortune There is less parameter processing amount during calculation, to improve the processing speed of heart infarction signal recognition device, and then improves Heart infarction recognition efficiency.

For the ease of the understanding of those skilled in the art, table 1 provides a kind of electrocardio based on grouping convolutional neural networks The network architecture parameters table of signal recognition method.

1 network architecture parameters table of table

Wherein, k is convolution kernel size, and s is sliding step, and p is complementary element, and f is output characteristic face number, and g is grouping Convolution group number.

Specifically, independent leads electrocardiosignal, that is, input feature vector characteristic dimension can be 1*600, the knot of the first convolutional layer Structure parameter can be with are as follows: and convolution kernel size k=61, sliding step s=1, complementary element p=0 export characteristic face number f=12, the The characteristic dimension of the first convolutional layer output feature of one convolutional layer output can be 12*540.First convolutional layer output feature is made For the input of the first maximum pond layer, the convolution of the maximum pond layer output of the sliding step s=3. first of the first maximum pond layer The characteristic dimension of block output feature can be 12*180.

Then, using the characteristic dimension of convolution block output feature as the input of the second convolutional layer.The structure of second convolutional layer Parameter can be with are as follows: and convolution kernel size k=31, sliding step s=1, complementary element p=0 export characteristic face number f=16, and second The characteristic dimension of the second convolutional layer output feature of convolutional layer output can be 16*150.Using the second convolutional layer output feature as The input of second maximum pond layer.And the convolution of the maximum pond layer output of sliding step s=3. second of the second maximum pond layer The characteristic dimension of block output feature can be 16*50.

Finally, using the characteristic dimension of convolution block output feature as the input of grouping convolutional layer.The then feature of grouping feature Dimension can be 4*50.The structural parameters for being grouped convolutional layer can be with are as follows: convolution kernel size k=9, sliding step s=1, supplement member Plain p=0 exports characteristic face number f=24, is grouped convolution group number g=4,；The grouping convolutional layer for being grouped convolutional layer output exports spy The characteristic dimension of sign can be 24*42.Convolutional layer output feature will be grouped as the input for being grouped maximum pond layer, grouping is maximum The sliding step s=2 of pond layer, the characteristic dimension for being grouped the convolution block output feature of maximum pond layer output can be 24*14.

Feature combination module 340 is somebody's turn to do for combining multiple grouping convolution features of every group of independent leads electrocardiosignal The electrocardiosignal assemblage characteristic of group independent leads electrocardiosignal.

Full connection processing module 350, carries out complete for the electrocardiosignal assemblage characteristic to each group independent leads electrocardiosignal Connection processing, obtains heart infarction exception probability.

Heart infarction determination module 360, for determining multi-lead when heart infarction exception probability is higher than preset abnormal probability threshold value Electrocardiosignal is heart infarction signal.

In the specific implementation, the input cell number of above-mentioned full Connection Neural Network classifier with position heart infarction identification feature Feature vector number it is equal, the output cell number of full Connection Neural Network classifier is 2, and then represents two kinds of prediction results. The predicted value that each input heart is clapped can be obtained, when obtained heart infarction exception probability is higher than preset abnormal probability threshold value, Quan Lian The predicted value for connecing neural network classifier output is 1, and representing this heart bat sample has the performance of heart infarction relevant abnormalities；When obtained heart infarction When abnormal probability is lower than preset abnormal probability threshold value, the predicted value of full Connection Neural Network classifier output is 0, represents this heart Clap sample health.

Modules in the above-mentioned electrocardiosignal identification device based on grouping convolutional neural networks can completely or partially lead to Software, hardware and combinations thereof are crossed to realize.Above-mentioned each module can be embedded in the form of hardware or independently of in computer equipment In processor, can also be stored in a software form in the memory in computer equipment, in order to processor call execute with The corresponding operation of upper modules.

Fig. 5 is a kind of schematic network structure of electrocardiosignal identification device based on grouping convolutional neural networks；Such as Fig. 5 It is shown, including first volume block, volume Two block and grouping convolution block；Wherein, first volume block includes the first convolutional layer and the One maximum pond layer；Volume Two block includes the second convolutional layer and the second maximum pond layer；Being grouped convolution block includes grouping convolution Layer and the maximum pond layer of grouping；Firstly, by the first convolutional layer, to independent leads electrocardiosignal carry out convolution, batch normalization and Activation obtains the first convolutional layer output feature；Again by the first maximum pond layer, the first convolutional layer output feature is carried out maximum Chi Hua obtains convolution block output feature.

Then, by the second convolutional layer, convolution, batch normalization and activation is carried out to convolution block output feature, obtain second Convolutional layer exports feature；Again by the second maximum pond layer, maximum pond is carried out to the second convolutional layer output feature, obtains convolution Feature.

Subsequently, by being grouped convolutional layer, convolution, batch normalization and activation is carried out to grouping feature, obtain grouping convolution Layer output feature；Again by being grouped maximum pond layer, maximum pond is carried out to grouping convolutional layer output feature, obtains grouping convolution Feature；

Finally, being grouped convolution feature to combination, electrocardiosignal assemblage characteristic is obtained, and carry out to electrocardiosignal assemblage characteristic Full connection processing, obtains heart infarction exception probability；Wherein, using full Connection Neural Network classifier output electrocardiosignal identification knot Fruit；The output cell number of full Connection Neural Network classifier is 2, and then represents two kinds of prediction results.Each input heart can be obtained The predicted value of bat, when obtained heart infarction exception probability is higher than preset abnormal probability threshold value, full Connection Neural Network classifier The predicted value of output is 1, and representing this heart bat sample has the performance of heart infarction relevant abnormalities；When obtained heart infarction exception probability is lower than default Abnormal probability threshold value when, the predicted value of full Connection Neural Network classifier output is 0, represents this heart and claps sample health.

Embodiment three

Figure 11 is a kind of structure chart of the training device for grouping convolutional neural networks that the embodiment of the present application three provides.Specifically , with reference to Figure 11, the training method of the grouping convolutional neural networks of the embodiment of the present application three is specifically included:

Training sample obtains module 1110, for obtaining the electrocardiosignal training sample for being directed to grouping convolutional neural networks.

In the specific implementation, passing through the public database of such as PTB, the electrocardiosignal for grouping convolutional neural networks is obtained Training sample.Wherein, electrocardiosignal training sample includes the multi-lead electrocardiosignal with abnormal signal and known heart infarction type With normal multi-lead electrocardiosignal.

Machine training module 1120, for carrying out machine instruction to grouping convolutional neural networks using electrocardiosignal training sample Practice, is grouped convolutional neural networks after being trained；Convolutional neural networks are grouped after training to include the convolution block trained in advance and divide Group convolution block；It is grouped convolutional neural networks after training, for every group of independent leads electrocardiosignal to be input to convolution block respectively, obtains To the convolution feature of the independent leads electrocardiosignal of respective sets；Independent leads electrocardiosignal is to split multi-lead electrocardiosignal to obtain 's；It is also used to be grouped the convolution feature of every group of independent leads electrocardio electric signal, obtains this group of independent leads electrocardio telecommunications Number grouping feature, and by grouping feature be input to grouping convolution block, obtain multiple groupings of this group of independent leads electrocardiosignal Convolution feature；It is also used to combine multiple grouping convolution features of every group of independent leads electrocardiosignal, obtains this group of independent leads heart The electrocardiosignal assemblage characteristic of electric signal；It is also used to carry out the electrocardiosignal assemblage characteristic of each group independent leads electrocardiosignal complete Connection processing, obtains heart infarction exception probability；It is also used to determine multi-lead electrocardiosignal for heart infarction letter according to heart infarction exception probability Number.

In the specific implementation, carrying out machine training to grouping convolutional neural networks, more after getting electrocardiosignal training sample Specifically, patients with myocardial infarction and non-salary motivation patient data can be collected, is randomly divided into training set and test set in proportion, Two datasets do not include same person's data simultaneously.The multi-lead electrocardiosignal of structuring is labeled as X, it will " there are heart infarctions The exception of relevant characteristic variation ", " there is no the exceptions of the relevant characteristic variation of heart infarction " label are as grouping convolution mind Output Y through network.(X, the Y) of training set collectively constitutes the training sample of the more topology convergence networks of multi-lead.X is by certain batch Size is obtained the predicted value Pred_Y of Y by propagated forward, is calculated by loss function by batch input grouping convolutional neural networks Y and Pred_Y loss, will lose backpropagation, using gradient descent method training network, obtain optimal grouping convolutional Neural net Convolutional neural networks are grouped after network i.e. training.

It includes the convolution block trained in advance and grouping convolution block that convolutional neural networks are grouped after training；Using grouping convolution Neural network carries out in the identification process of electrocardiosignal, multi-lead electrocardiosignal available first, and splits above-mentioned multi-lead Electrocardiosignal obtains multiple groups independent leads electrocardiosignal；Then, every group of above-mentioned independent leads electrocardiosignal is input to respectively Convolution block obtains the convolution feature of the independent leads electrocardiosignal of respective sets；Subsequently, to every group of independent leads electrocardio electric signal Convolution feature be grouped, obtain the grouping feature of this group of independent leads electrocardio electric signal, and grouping feature is input to point Group convolution block obtains multiple grouping convolution features of this group of independent leads electrocardiosignal as the input of the grouping convolution block；Again Then, the grouping convolution feature for combining multiple grouping convolution blocks output of every group of independent leads electrocardiosignal, obtains group independence The electrocardiosignal assemblage characteristic of lead electrocardiosignal, and the electrocardiosignal assemblage characteristic of each group independent leads electrocardiosignal is carried out Full connection processing, obtains heart infarction exception probability；Finally, determining that multi-lead electrocardiosignal is heart infarction further according to heart infarction exception probability Signal.

Technical solution provided by the embodiments of the present application, by using the electrocardiosignal training for grouping convolutional neural networks Sample carries out machine training to grouping convolutional neural networks, is grouped convolutional neural networks after being trained；Convolution is grouped after training Neural network includes the convolution block trained in advance and grouping convolution block；By obtaining multi-lead electrocardiosignal, and split multi-lead Electrocardiosignal obtains independent leads electrocardiosignal；Then, first the independent leads electrocardiosignal is input in convolution block, is obtained Convolution feature；Then, it is grouped to the convolution feature, obtains grouping feature, and grouping feature is input to grouping convolution Block obtains grouping convolution feature；Then, heart infarction exception probability is being determined according to the grouping convolution feature；Finally, different according to heart infarction Normal probability determines that multi-lead electrocardiosignal is heart infarction signal；In this way, the parameter amount of convolution operation can be reduced, do not reducing In the case where the recognition performance for being grouped convolutional neural networks, the over-fitting of the grouping convolutional neural networks is alleviated, so as to More accurately identification judges whether multi-lead electrocardiosignal is heart infarction signal.

Further, when identifying electrocardiosignal, without dependent on to electrocardiosignal key point Q wave, P wave, J point, S point, T The accurate positionin of wave, though electrocardiosignal quality it is bad, electrocardiosignal key point can not be accurately positioned in the case where, by defeated Enter the grouping convolutional neural networks of the application, so as to more accurately identify heart infarction risk from electrocardiosignal.

Modules in the training device of above-mentioned grouping convolutional neural networks can fully or partially through software, hardware and A combination thereof is realized.Above-mentioned each module can be embedded in the form of hardware or independently of in the processor in computer equipment, can also Be stored in the memory in computer equipment in a software form, in order to which processor calls the above modules of execution corresponding Operation.

Fig. 6 is that a kind of characteristic dimension of electrocardiosignal identification based on grouping convolutional neural networks changes schematic diagram.Such as Fig. 6 It is shown, it is input in first volume block using multiple groups independent leads electrocardiosignal as input feature vector by more autonomous channels first, Wherein, the characteristic dimension of independent leads electrocardiosignal can be 1*600, and then, the convolution pond by first volume block obtains Convolution block exports feature；Wherein, the characteristic dimension of convolution block output feature can be 12*180；Subsequently, convolution block is exported Input of the feature as volume Two block, the convolution pond by volume Two block, obtains convolution feature；Wherein, convolution feature Characteristic dimension can be 16*50；Subsequently, it is known that grouping convolution group number is 4, and convolution feature has 16 characteristic faces, then will Above-mentioned convolution feature is divided into four groups of grouping features；Wherein, grouping feature has 4 characteristic faces.Then, by four groups of grouping features Input grouping convolution block is grouped convolution pond, the grouping convolution feature finally exported；Wherein, it is grouped the feature of convolution feature Dimension can be 24*14；Specifically, grouping convolution feature is divided into four groups, every group of 6 characteristic faces.Due to 12 groups of independent leads Electrocardiosignal is corresponding with 12 groups of grouping convolution features, and the intrinsic dimensionality of each grouping convolution feature is 24*14, then will be above-mentioned The electrocardiosignal assemblage characteristic that the group number that grouping convolution feature is combined by feature is 1.Wherein, the electrocardiosignal after connection The intrinsic dimensionality of assemblage characteristic is 268*12.Finally, electrocardiosignal assemblage characteristic is input to full Connection Neural Network classifier In, wherein the output cell number of full Connection Neural Network classifier is 2, and then represent two kinds of prediction results.It can be obtained each defeated The predicted value for entering heart bat, when obtained heart infarction exception probability is higher than preset abnormal probability threshold value, full Connection Neural Network point The predicted value of class device output is 1, and representing this heart bat sample has the performance of heart infarction relevant abnormalities；When obtained heart infarction exception probability is lower than When preset exception probability threshold value, the predicted value of full Connection Neural Network classifier output is 0, represents this heart and claps sample health.

Fig. 7 is the flow chart that electrocardiosignal identification is carried out based on neural network.As shown, passing through multi-lead first Electrocardiograph acquires the multi-lead electrocardiosignal of patient, stores multi-lead electrocardiosignal, then carries out to multi-lead electrocardiosignal Wavelet decomposition, partial dimensional signal zero setting etc. pretreatment, and structuring processing is carried out to signal, obtains the signal of structuring Matrix, the input as the more topology convergence networks of multi-lead.The more topology convergence networks of multi-lead according to the input data, export more Network polymerization recognition result, and recognition result is polymerize according to Multi net voting, final report is generated, reflection patient whether there is heart infarction Risk.

Example IV

Fig. 8 is a kind of electrocardiosignal recognition methods based on grouping convolutional neural networks that the embodiment of the present application four provides Flow chart.With reference to Fig. 8, the electrocardiosignal recognition methods provided in this embodiment based on grouping convolutional neural networks is specifically included:

S810 obtains multi-lead electrocardiosignal, and splits the multi-lead electrocardiosignal, obtains multiple groups independent leads electrocardio Signal；

Independent leads electrocardiosignal described in every group is input to the convolution block respectively, obtains the independence of respective sets by S820 The convolution feature of lead electrocardiosignal；

S830 is grouped the convolution feature of every group of independent leads electrocardio electric signal, obtains this group of independent leads The grouping feature of electrocardio electric signal, and the grouping feature is input to the grouping convolution block, obtain this group of independent leads heart Multiple grouping convolution features of electric signal；

S840 combines multiple grouping convolution features of every group of independent leads electrocardiosignal, obtains this group of independent leads The electrocardiosignal assemblage characteristic of electrocardiosignal；

S850 carries out full connection processing to the electrocardiosignal assemblage characteristic of each group independent leads electrocardiosignal, obtains Heart infarction exception probability；

S860 determines the multi-lead electrocardiosignal for heart infarction signal according to the heart infarction exception probability.

Technical solution provided by the embodiments of the present application by obtaining multi-lead electrocardiosignal, and splits multi-lead electrocardio letter Number, obtain independent leads electrocardiosignal；Then, first the independent leads electrocardiosignal is input in convolution block, obtains convolution spy Sign；Then, it is grouped to the convolution feature, obtains grouping feature, and grouping feature is input to grouping convolution block, obtain It is grouped convolution feature；Then, heart infarction exception probability is being determined according to the grouping convolution feature；Finally, according to heart infarction exception probability, Determine that multi-lead electrocardiosignal is heart infarction signal；In this way, the parameter amount of convolution operation can be reduced, grouping convolution is not being reduced In the case where the recognition performance of neural network, the over-fitting of the grouping convolutional neural networks is alleviated, so as to more accurate Ground identification judges whether multi-lead electrocardiosignal is heart infarction signal.

In another embodiment, the convolution feature to every group of independent leads electrocardio electric signal is grouped, Obtain the grouping feature of this group of independent leads electrocardio electric signal, comprising:

Obtain the grouping convolution group number of the grouping convolution block；

Equal equal part is carried out according to convolution feature of the grouping convolution group number to every group of independent leads electrocardio electric signal Group obtains the grouping feature of this group of independent leads electrocardio electric signal.

In another embodiment, the grouping convolution block includes grouping convolutional layer and is grouped maximum pond layer, it is described by The grouping feature is input to the grouping convolution block, obtains multiple grouping convolution features of this group of independent leads electrocardiosignal, Include:

By the grouping convolutional layer, convolution, batch normalization and activation are carried out to the grouping feature, obtain grouping convolution Layer output feature；

By the maximum pond layer of the grouping, maximum pond is carried out to grouping convolutional layer output feature, obtains the group Multiple grouping convolution features of independent leads electrocardiosignal.

In another embodiment, the convolution block of the grouping convolutional neural networks includes first volume block and the second convolution Block, it is described that independent leads electrocardiosignal described in every group is input to the convolution block respectively, obtain the independent leads heart of respective sets The convolution feature of electric signal, comprising:

By the first volume block, convolution sum pond is carried out to the independent leads electrocardiosignal, it is defeated to obtain convolution block Feature out；

By the volume Two block, special progress convolution sum pond is exported to the convolution block, obtains the convolution feature.

In another embodiment, the first volume block includes the first convolutional layer and the first maximum pond layer, described logical The first volume block is crossed, convolution sum pond is carried out to the independent leads electrocardiosignal, obtains convolution block output feature, packet It includes:

By first convolutional layer, convolution, batch normalization and activation are carried out to the independent leads electrocardiosignal, obtained First convolutional layer exports feature；

By the described first maximum pond layer, maximum pond is carried out to first convolutional layer output feature, is obtained described Convolution block exports feature.

In another embodiment, the volume Two block includes the second convolutional layer and the second maximum pond layer, described logical The volume Two block is crossed, special progress convolution sum pond is exported to the convolution block, obtains the convolution feature, comprising:

By second convolutional layer, convolution, batch normalization and activation are carried out to convolution block output feature, obtain the Two convolutional layers export feature；

By the described second maximum pond layer, maximum pond is carried out to second convolutional layer output feature, is obtained described Convolution feature.

In another embodiment, the acquisition multi-lead electrocardiosignal, comprising:

Receive original signal；

Wavelet decomposition is carried out to the original signal, obtains wavelet decomposition signal；The wavelet decomposition signal is tieed up with X1；

Signal zero setting to the X2 dimension in the wavelet decomposition signal, obtains part zero setting signal；Wherein, X2 < X1；

Wavelet inverse transformation is carried out to the part zero setting signal, obtains denoised signal；The denoised signal is high-frequency noise With the signal after baseline drift removal；

According to the denoised signal, the multi-lead electrocardiosignal is obtained.

In another embodiment, described according to the denoised signal, obtain the multi-lead electrocardiosignal, comprising:

Determine the R wave position of the denoised signal；

The electrocardiosignal recognition methods based on grouping convolutional neural networks of above-mentioned offer can be used for executing above-mentioned any reality The electrocardiosignal identification device based on grouping convolutional neural networks for applying example offer, has corresponding function and beneficial effect.

It is above right that specific restriction about the electrocardiosignal recognition methods based on grouping convolutional neural networks may refer to In the restriction of the electrocardiosignal identification device based on grouping convolutional neural networks, details are not described herein.

Embodiment five

Figure 10 is a kind of flow chart of the training method for grouping convolutional neural networks that the embodiment of the present application five provides.With reference to The training method of Figure 10, grouping convolutional neural networks provided in this embodiment specifically include:

S1010 obtains the electrocardiosignal training sample for the grouping convolutional neural networks；

S1020 carries out machine training to the grouping convolutional neural networks using the electrocardiosignal training sample, obtains Convolutional neural networks are grouped after training；It includes the convolution block trained in advance and grouping volume that convolutional neural networks are grouped after the training Block；Convolutional neural networks are grouped after the training, for every group of independent leads electrocardiosignal to be input to the convolution respectively Block obtains the convolution feature of the independent leads electrocardiosignal of respective sets；The independent leads electrocardiosignal is to split the multi-lead heart What electric signal obtained；It is also used to be grouped the convolution feature of every group of independent leads electrocardio electric signal, it is only to obtain the group The grouping feature of vertical lead electrocardio electric signal, and the grouping feature is input to the grouping convolution block, obtain group independence Multiple grouping convolution features of lead electrocardiosignal；It is also used to combine multiple groupings volume of every group of independent leads electrocardiosignal Product feature, obtains the electrocardiosignal assemblage characteristic of this group of independent leads electrocardiosignal；It is also used to each group independent leads heart The electrocardiosignal assemblage characteristic of electric signal carries out full connection processing, obtains heart infarction exception probability；It is also used to different according to the heart infarction Normal probability determines the multi-lead electrocardiosignal for heart infarction signal.

The training method of the grouping convolutional neural networks of above-mentioned offer can be used for executing point that above-mentioned any embodiment provides The training device of group convolutional neural networks, has corresponding function and beneficial effect.

The specific restriction of training method about grouping convolutional neural networks may refer to above for grouping convolution mind Electrocardiosignal identification device through network limits, and details are not described herein.

It should be understood that although each step in the flow chart of Fig. 8 and Figure 10 is successively shown according to the instruction of arrow, But these steps are not that the inevitable sequence according to arrow instruction successively executes.Unless expressly state otherwise herein, these There is no stringent sequences to limit for the execution of step, these steps can execute in other order.Moreover, in Fig. 8 and Figure 10 At least part step may include multiple sub-steps perhaps these sub-steps of multiple stages or stage be not necessarily Synchronization executes completion, but can execute at different times, and the execution sequence in these sub-steps or stage also need not Be so successively carry out, but can at least part of the sub-step or stage of other steps or other steps in turn or Person alternately executes.

Embodiment six

Fig. 9 is the structural schematic diagram for a kind of electronic equipment that the embodiment of the present application six provides.As shown in the figure, which sets Standby includes: processor 40, memory 41, the display screen 42 with touch function, input unit 43, output device 44 and communication Device 45.The quantity of processor 40 can be one or more in the electronic equipment, in figure by taking a processor 40 as an example.It should The quantity of memory 41 can be one or more in electronic equipment, in figure by taking a memory 41 as an example.The electronic equipment Processor 40, memory 41, display screen 42, input unit 43, output device 44 and communication device 45 can pass through bus Or other modes connect, in figure for being connected by bus.In embodiment, electronic equipment can be computer, and mobile phone is put down Plate, projector or interactive intelligent tablet computer etc..In embodiment, by taking electronic equipment is interactive intelligent tablet computer as an example, it is described.

Memory 41 is used as a kind of computer readable storage medium, can be used for storing software program, journey can be performed in computer Sequence and module, the corresponding program instruction/module of electrocardiosignal recognition methods as described in the application any embodiment.Memory 41 can mainly include storing program area and storage data area, wherein storing program area can storage program area, at least one function Required application program；Storage data area, which can be stored, uses created data etc. according to equipment.In addition, memory 41 can be with It can also include nonvolatile memory, for example, at least disk memory, a flash memory including high-speed random access memory Device or other non-volatile solid state memory parts.In some instances, memory 41 can further comprise relative to processor 40 remotely located memories, these remote memories can pass through network connection to equipment.The example of above-mentioned network include but It is not limited to internet, intranet, local area network, mobile radio communication and combinations thereof.

Display screen 42 is the display screen 42 with touch function, can be capacitance plate, electromagnetic screen or infrared screen.Generally For, display screen 42 is used to show data according to the instruction of processor 40, is also used to receive the touch behaviour for acting on display screen 42 Make, and corresponding signal is sent to processor 40 or other devices.Optionally, it when display screen 42 is infrared screen, also wraps Infrared touch frame is included, which is arranged in the surrounding of display screen 42, can be also used for receiving infrared signal, and should Infrared signal is sent to processor 40 or other equipment.

Communication device 45 communicates to connect for establishing with other equipment, can be wire communication device and/or channel radio T unit.

Input unit 43 can be used for receiving the number or character information of input, and generates and set with the user of electronic equipment It sets and the related key signals of function control inputs, can also be the camera for obtaining image and obtain audio data Pick up facility.Output device 44 may include the audio frequency apparatuses such as loudspeaker.It should be noted that input unit 43 and output device 44 concrete composition may be set according to actual conditions.

Software program, instruction and the module that processor 40 is stored in memory 41 by operation, thereby executing equipment Various function application and data processing, that is, realize it is above-mentioned based on grouping convolutional neural networks electrocardiosignal identification side Method.

Specifically, the grouping convolutional neural networks include convolution block and grouping convolution block in embodiment, processor 40 is held When the one or more programs stored in line storage 41, it is implemented as follows operation:

Multi-lead electrocardiosignal is obtained, and splits the multi-lead electrocardiosignal, obtains independent leads electrocardiosignal；

The independent leads electrocardiosignal is input to the convolution block respectively, obtains convolution feature；

The convolution feature is grouped, obtains grouping feature, and the grouping feature is input to the grouping and is rolled up Block obtains grouping convolution feature；

Combine the grouping convolution feature, obtain electrocardiosignal assemblage characteristic, and to the electrocardiosignal assemblage characteristic into The full connection processing of row, obtains heart infarction exception probability；

On the basis of the above embodiments, described that the convolution feature is grouped, obtain grouping feature, comprising:

Impartial grouping is carried out to the convolution feature according to the grouping convolution group number, obtains the grouping feature.

On the basis of the above embodiments, the grouping convolution block includes grouping convolutional layer and the maximum pond layer of grouping, institute It states and the grouping feature is input to the grouping convolution block, obtain grouping convolution feature, comprising:

By the maximum pond layer of the grouping, maximum pond is carried out to grouping convolutional layer output feature, is obtained described It is grouped convolution feature.

On the basis of the above embodiments, the convolution block of the grouping convolutional neural networks includes first volume block and second Convolution block, it is described that the independent leads electrocardiosignal is input to the convolution block respectively, obtain convolution feature, comprising:

On the basis of the above embodiments, the first volume block includes the first convolutional layer and the first maximum pond layer, institute It states through the first volume block, convolution sum pond is carried out to the independent leads electrocardiosignal, obtain convolution block output feature, Include:

On the basis of the above embodiments, the volume Two block includes the second convolutional layer and the second maximum pond layer, institute It states through the volume Two block, special progress convolution sum pond is exported to the convolution block, obtains the convolution feature, comprising:

On the basis of the above embodiments, the acquisition multi-lead electrocardiosignal, comprising:

Receive original signal；

On the basis of the above embodiments, described that the multi-lead electrocardiosignal is obtained according to the denoised signal, packet It includes:

Determine the R wave position of the denoised signal；

Specifically, in embodiment, it is also specific real when processor 40 executes the one or more programs stored in memory 41 Now following operation:

Embodiment seven

The embodiment of the present application seven also provides a kind of storage medium comprising computer executable instructions, and the computer can be held Row instruction by computer processor when being executed for executing a kind of electrocardiosignal identification side based on grouping convolutional neural networks Method, the grouping convolutional neural networks include the convolution block trained in advance and are grouped convolution block, this method comprises:

The computer executable instructions are also used to execute when being executed by computer processor a kind of based on grouping convolution The electrocardiosignal recognition methods of neural network, this method comprises:

Certainly, a kind of storage medium comprising computer executable instructions, computer provided by the embodiment of the present application The operation for the electrocardiosignal recognition methods based on grouping convolutional neural networks that executable instruction is not limited to the described above, can be with Execute the related behaviour in the electrocardiosignal recognition methods based on grouping convolutional neural networks provided by the application any embodiment Make, and has corresponding function and beneficial effect.

It should be noted that term involved in the embodiment of the present invention " first second third " be only be that difference is similar Object, do not represent the particular sorted for object, it is possible to understand that ground, " Yi Er third " can be in the case where permission Exchange specific sequence or precedence.It should be understood that the object that " first second third " is distinguished in the appropriate case can be mutual It changes, so that the embodiment of the present invention described herein can be real with the sequence other than those of illustrating or describing herein It applies.

Those of ordinary skill in the art will appreciate that realizing all or part of the process in above-described embodiment method, being can be with Relevant hardware is instructed to complete by computer program, the computer program can be stored in a non-volatile computer In read/write memory medium, the computer program is when being executed, it may include such as the process of the embodiment of above-mentioned each method.Wherein, To any reference of memory, storage, database or other media used in each embodiment provided herein, Including non-volatile and/or volatile memory.Nonvolatile memory may include read-only memory (ROM), programming ROM (PROM), electrically programmable ROM (EPROM), electrically erasable ROM (EEPROM) or flash memory.Volatile memory may include Random access memory (RAM) or external cache.By way of illustration and not limitation, RAM is available in many forms, Such as static state RAM (SRAM), dynamic ram (DRAM), synchronous dram (SDRAM), double data rate sdram (DDRSDRAM), enhancing Type SDRAM (ESDRAM), synchronization link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic ram (DRDRAM) and memory bus dynamic ram (RDRAM) etc..

Each technical characteristic of above embodiments can be combined arbitrarily, for simplicity of description, not to above-described embodiment In each technical characteristic it is all possible combination be all described, as long as however, the combination of these technical characteristics be not present lance Shield all should be considered as described in this specification.

The several embodiments of the application above described embodiment only expresses, the description thereof is more specific and detailed, but simultaneously It cannot therefore be construed as limiting the scope of the patent.It should be pointed out that coming for those of ordinary skill in the art It says, without departing from the concept of this application, various modifications and improvements can be made, these belong to the protection of the application Range.Therefore, the scope of protection shall be subject to the appended claims for the application patent.

Claims

1. a kind of electrocardiosignal identification device based on grouping convolutional neural networks, which is characterized in that the grouping convolutional Neural Network includes the convolution block trained in advance and grouping convolution block, and described device includes:

Signal acquisition module for obtaining multi-lead electrocardiosignal, and splits the multi-lead electrocardiosignal, obtains multiple groups independence Lead electrocardiosignal；

Convolution module obtains respective sets for independent leads electrocardiosignal described in every group to be input to the convolution block respectively The convolution feature of independent leads electrocardiosignal；

Feature grouping module is grouped for the convolution feature to every group of independent leads electrocardio electric signal, obtains the group The grouping feature of independent leads electrocardio electric signal, and the grouping feature is input to the grouping convolution block, it is only to obtain the group Multiple grouping convolution features of vertical lead electrocardiosignal；

Feature combination module obtains the group for combining multiple grouping convolution features of every group of independent leads electrocardiosignal The electrocardiosignal assemblage characteristic of independent leads electrocardiosignal；

Full connection processing module, is connected entirely for the electrocardiosignal assemblage characteristic to each group independent leads electrocardiosignal Processing, obtains heart infarction exception probability；

2. the apparatus according to claim 1, which is characterized in that the feature grouping module, comprising:

It is grouped submodule, for the convolution feature according to the grouping convolution group number to every group of independent leads electrocardio electric signal Impartial grouping is carried out, the grouping feature of this group of independent leads electrocardio electric signal is obtained.

3. the apparatus according to claim 1, which is characterized in that the grouping convolution block includes grouping convolutional layer and is grouped most Great Chiization layer, the feature grouping module, comprising:

It is grouped convolution feature output sub-module, for convolution being carried out to the grouping feature, criticizing and return by the grouping convolutional layer One changes and activates, and obtains grouping convolutional layer output feature；

Pond beggar's module, for carrying out maximum pond to grouping convolutional layer output feature by the maximum pond layer of the grouping Change, obtain multiple grouping convolution features of this group of independent leads electrocardiosignal, to reduce the parameter processing of the grouping convolution block Amount.

4. the apparatus according to claim 1, which is characterized in that the convolution block of the grouping convolutional neural networks includes first Convolution block and volume Two block, the convolution module, comprising:

First convolution submodule, for carrying out convolution sum pond to the independent leads electrocardiosignal by the first volume block Change, convolution block output feature is obtained, to reduce the parameter processing amount of the volume Two block；

Second convolution submodule, for exporting special progress convolution sum pond to the convolution block, obtaining by the volume Two block It is excessive to avoid dimensionality reduction gradient and cause Character losing to the convolution feature.

5. device according to claim 4, which is characterized in that the first convolution submodule is specifically used for:

By first convolutional layer, convolution, batch normalization and activation are carried out to the independent leads electrocardiosignal, obtain first Convolutional layer exports feature；By the described first maximum pond layer, maximum pond is carried out to first convolutional layer output feature, is obtained Feature is exported to the convolution block.

6. device according to claim 4, which is characterized in that the volume Two block include the second convolutional layer and second most Great Chiization layer, the second convolution submodule, is specifically used for:

By second convolutional layer, convolution, batch normalization and activation are carried out to convolution block output feature, obtain volume Two Lamination exports feature；By the described second maximum pond layer, maximum pond is carried out to second convolutional layer output feature, is obtained The convolution feature.

7. the apparatus according to claim 1, which is characterized in that the signal acquisition module, comprising:

Original signal receiving submodule, for receiving original signal；

Wavelet decomposition submodule obtains wavelet decomposition signal for carrying out wavelet decomposition to the original signal；The small wavelength-division Signal is solved to tie up with X1；

Zero setting submodule obtains part zero setting signal for the signal zero setting to the X2 dimension in the wavelet decomposition signal；Its In, X2 < X1；

Inverse transformation submodule obtains denoised signal for carrying out wavelet inverse transformation to the part zero setting signal；The denoising letter Number for high-frequency noise and baseline drift removal after signal；

8. device according to claim 7, which is characterized in that the multi-lead acquisition submodule is specifically used for:

Determine the R wave position of the denoised signal；

Using the denoised signal on R wave position, the preceding M1 position, the rear M2 position, structured signal is formed Matrix, as the multi-lead electrocardiosignal.

9. a kind of training device for being grouped convolutional neural networks, which is characterized in that described device includes:

Machine training module, for carrying out machine instruction to the grouping convolutional neural networks using the electrocardiosignal training sample Practice, is grouped convolutional neural networks after being trained；It includes the convolution block trained in advance that convolutional neural networks are grouped after the training With grouping convolution block；Convolutional neural networks are grouped after the training, for being respectively input to every group of independent leads electrocardiosignal The convolution block obtains the convolution feature of the independent leads electrocardiosignal of respective sets；The independent leads electrocardiosignal is to split Multi-lead electrocardiosignal obtains；It is also used to be grouped the convolution feature of every group of independent leads electrocardio electric signal, obtain It is input to the grouping convolution block to the grouping feature of this group of independent leads electrocardio electric signal, and by the grouping feature, is obtained Multiple grouping convolution features of this group of independent leads electrocardiosignal；It is also used to combine the more of every group of independent leads electrocardiosignal A grouping convolution feature, obtains the electrocardiosignal assemblage characteristic of this group of independent leads electrocardiosignal；It is also used to only to each group The electrocardiosignal assemblage characteristic of vertical lead electrocardiosignal carries out full connection processing, obtains heart infarction exception probability；It is also used to according to institute Heart infarction exception probability is stated, determines the multi-lead electrocardiosignal for heart infarction signal.

10. a kind of electronic equipment characterized by comprising memory, one or more processors；

The memory, for storing one or more programs；

When one or more of programs are executed by one or more of processors, so that one or more of processors are held Row following steps；

Independent leads electrocardiosignal described in every group is input to the convolution block respectively, obtains the independent leads electrocardio letter of respective sets Number convolution feature；

The convolution feature of every group of independent leads electrocardio electric signal is grouped, this group of independent leads electrocardio electric signal is obtained Grouping feature, and the grouping feature is input to the grouping convolution block, obtains the more of this group of independent leads electrocardiosignal A grouping convolution feature；

The multiple grouping convolution features for combining every group of independent leads electrocardiosignal, obtain this group of independent leads electrocardiosignal Electrocardiosignal assemblage characteristic；

Full connection processing is carried out to the electrocardiosignal assemblage characteristic of each group independent leads electrocardiosignal, it is extremely general to obtain heart infarction Rate；

11. a kind of storage medium comprising computer executable instructions, which is characterized in that the computer executable instructions by For executing following steps when computer processor executes；