CN114841209A - Multi-target domain electrocardiosignal classification method based on depth field self-adaption - Google Patents
Multi-target domain electrocardiosignal classification method based on depth field self-adaption Download PDFInfo
- Publication number
- CN114841209A CN114841209A CN202210526548.4A CN202210526548A CN114841209A CN 114841209 A CN114841209 A CN 114841209A CN 202210526548 A CN202210526548 A CN 202210526548A CN 114841209 A CN114841209 A CN 114841209A
- Authority
- CN
- China
- Prior art keywords
- features
- layer
- branch unit
- convolution
- lead
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Signal Processing (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
A multi-target domain electrocardiosignal classification method based on depth field self-adaptation does not need feature alignment between two domains, focuses on a prediction method of the category accuracy and diversity of a target domain, improves the generalization capability of a model on the multi-target domain, has high convergence speed, and solves the problem of poor effect of a single target domain adaptation algorithm on the multi-target domain. In addition, the introduction of the deformable convolution kernel can consciously adjust the field of experience of convolution in the model training process, and better adapt to the lead characteristics of the multi-lead electrocardiosignals.
Description
Technical Field
The invention relates to the field of electrocardiosignal classification, in particular to a multi-target domain electrocardiosignal classification method based on depth field self-adaptation.
Background
At present, the existing intelligent electrocardiosignal classification method has good effect in labeling a large amount of electrocardiosignal data, however, in practical application, the labeling of the electrocardiosignal needs to depend on expert knowledge, and the method needs a large amount of manpower and time cost. The development of the field self-adaption in the transfer learning provides a thought for solving the problem, namely, a neural network trained on a source domain data set (labeled) can be highly suitable for a target domain data set (unlabeled) with characteristic distribution obviously different from that of a source domain, and the problem of characteristic distribution misalignment of the source domain and the target domain can be effectively solved.
Disclosure of Invention
In order to overcome the defects of the technology, the invention provides a method for improving the accuracy of the electrocardio classification of a plurality of target domains on a classification model trained by a source domain.
The technical scheme adopted by the invention for overcoming the technical problems is as follows:
a multi-target domain electrocardiosignal classification method based on depth field self-adaptation comprises the following steps:
a) acquiring a plurality of 12-lead electrocardiosignals acquired by different acquisition equipment and different individuals, labeling the 12-lead electrocardiosignals, dividing the labeled 12-lead electrocardiosignals into a source domain S, and dividing the unlabeled 12-lead electrocardiosignals into a target domain combination T, wherein T is ═ T { T } 1 ,T 2 ,...,T k ,...,T M },T k The number of target domains is k, wherein k is 1, 2.
b) Respectively and sequentially performing down-sampling processing, slicing processing and normalization processing on the source domain S and the target domain combination T to obtain a processed source domain S 'and a processed target domain combination T';
c) establishing a deep learning classification model of 12-lead electrocardiosignals, wherein the deep learning classification model of the 12-lead electrocardiosignals is characterized byThe extractor and the full-connection classifier are formed, the processed source domain S 'and the processed target domain combination T' are input into a deep learning classification model of 12-lead electrocardiosignals, and the label prediction value of the source domain is outputAnd label prediction value of target domain
d) Calculating a loss functionOptimizing a deep learning classification model of the 12-lead electrocardiosignals by an Adam optimization algorithm;
e) and respectively inputting the data of each target domain into the optimized deep learning classification model of the 12-lead electrocardiosignals, outputting to obtain a label predicted value of each target domain, and finishing the classification of the electrocardiosignals of the target domains.
Further, the method for processing the slices in the step b) comprises the following steps: the method comprises the steps of down-sampling all 12-lead electrocardiosignals in a source domain S or a target domain combination T to have the same frequency, randomly intercepting 30S of electrocardiosignals, intercepting 12-lead electrocardiosignals with the data length exceeding 30S, and repeatedly filling 12-lead electrocardiosignals with the data length being less than 30S.
Further, step c) comprises the steps of:
c-1) the feature extractor sequentially comprises a first residual error structure, a second residual error structure, a third residual error structure and a fourth residual error structure;
c-2.1) the first residual error structure is composed of a first branch unit, a second branch unit and a third branch unit, wherein the first branch unit is composed of a first convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a second convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and the processed source domain S 'and the processed target domain combination T' are respectively input into the first branch unit and then respectively output to obtain the lead internal characteristic of the 12-lead electrocardiosignalAnd features of
c-2.2) the second branch unit of the first residual error structure is composed of a convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a DeformConv variable convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and the processed source domain S 'and the processed target domain combination T' are respectively input into the second branch unit and then respectively output to obtain the inter-lead characteristics of the 12-lead electrocardiosignalsAnd features of
c-2.3) the third branch unit of the first residual error structure is composed of a convolution layer and a Maxpool layer in sequence, and the processed source domain S 'and the processed target domain combination T' are respectively input into the third branch unit and then respectively output to obtain the characteristics of 12-lead electrocardiosignalsAnd features of
c-2.4) in-lead characterizationAnd features between leadsInputting the spliced features into a Maxpool layer and outputting the Maxpool layer to obtain the featuresInner features of leadsAnd features between leadsAfter feature splicing, inputting the feature into a Maxpool layer of the Maxpool layer and outputting the feature to obtain features
c-2.5) characterization ofAnd features ofOverlap to form a new feature F 1 S Will be characterized byAnd features ofPerforming superposition to form new features
c-3.1) the second residual structure is composed of a first branch unit, a second branch unit and a third branch unit, wherein the first branch unit is composed of a first convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a second convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and the characteristics F are respectively formed by 1 S And features ofRespectively output the internal lead characteristics of the obtained 12-lead electrocardiosignals after being input into the first branch unitAnd features of
c-3.2) the second branch unit of the second residual structure sequentially comprises a convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a DeformConv variable convolution layer, a second batch of normalization layers and a second ReLu activation function layer, and respectively forms the characteristic F 1 S And featuresThe inter-lead characteristics of 12-lead electrocardiosignals are respectively output after being input into the second branch unitAnd features of
c-3.3) the third branch unit of the second residual structure is composed of a convolution layer and a Maxpool layer in sequence, and respectively converts the characteristic F 1 S And featuresRespectively outputting the signals to obtain the characteristics of 12-lead electrocardiosignals after being input into a third branch unitAnd features of
c-3.4) in-lead characterizationAnd features between leadsInputting the spliced features into a Maxpool layer and outputting the Maxpool layer to obtain the featuresInner features of leadsAnd features between leadsAfter feature splicing, inputting the feature into a Maxpool layer of the Maxpool layer and outputting the feature to obtain features
c-3.5) characterization ofAnd features ofPerforming superposition to form new featuresWill be characterized byAnd features ofPerforming superposition to form new features
c-4.1) the third residual error structure is composed of a first branch unit, a second branch unit and a third branch unit, wherein the first branch unit is composed of a first convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a second convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and the characteristics of the first branch unit, the first batch of normalization layers and the second batch of normalization layers are respectively characterizedAnd featuresRespectively output the internal lead characteristics of the obtained 12-lead electrocardiosignals after being input into the first branch unitAnd features of
c-4.2) the second branch unit of the third residual error structure sequentially comprises a convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a DeformConv variable convolution layer, a second batch of normalization layers and a second ReLu activation function layer, and the characteristics of the second branch unit are respectively represented byAnd featuresThe inter-lead characteristics of 12-lead electrocardiosignals are respectively output after being input into the second branch unitAnd features of
c-4.3) the third branch unit of the third residual structure is composed of a convolution layer and a Maxpool layer in sequence, and features of the convolution layer and the Maxpool layer are respectivelyAnd featuresRespectively outputting the signals to obtain the characteristics of 12-lead electrocardiosignals after being input into a third branch unitAnd features of
c-4.4) in-lead characterizationAnd features between leadsAfter feature splicing, inputting the feature into a Maxpool layer of the Maxpool layer and outputting the feature to obtain featuresInner features of leadsAnd features between leadsInputting the spliced features into a Maxpool layer and outputting the Maxpool layer to obtain the features
c-4.5) characterization ofAnd features ofOverlap to form a new feature F 3 S Will be characterized byAnd features ofPerforming superposition to form new features
c-5.1) the fourth residual structure is composed of a first branch unit, a second branch unit, and a third branch unit,the first branch unit consists of a first convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a second convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and the first branch unit respectively consists of a characteristic F 3 S And featuresRespectively output the internal lead characteristics of the obtained 12-lead electrocardiosignals after being input into the first branch unitAnd features of
c-5.2) the second branch unit of the fourth residual structure sequentially comprises a convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a DeformConv variable convolution layer, a second batch of normalization layers and a second ReLu activation function layer, and respectively forms the characteristic F 3 S And featuresThe inter-lead characteristics of 12-lead electrocardiosignals are respectively output after being input into the second branch unitAnd features of
c-5.3) the third branch unit of the fourth residual structure is composed of a convolution layer and a Maxpool layer in sequence, and the characteristics F are respectively obtained 3 S And featuresRespectively outputting the signals to obtain the characteristics of 12-lead electrocardiosignals after being input into a third branch unitAnd features of
c-5.4) in-lead characterizationAnd features between leadsInputting the spliced features into a Maxpool layer and outputting the Maxpool layer to obtain the featuresInner features of leadsAnd features between leadsAfter feature splicing, inputting the feature into a Maxpool layer of the Maxpool layer and outputting the feature to obtain features
c-5.5) characterization ofAnd features ofPerforming superposition to form new featuresWill be characterized byAnd features ofPerforming superposition to form new features
c-6) characterization ofAnd features ofInputting the data into a full-connection classifier, and outputting the data to obtain a label predicted value of a source domainAnd label prediction value of target domainFurther, in the step c-2.1), the number of channels of the first convolution layer and the second convolution layer of the first branch unit is 32, the size of the convolution kernel is 1 × 3, in the step c-2.2), the number of channels of the convolution layer of the second branch unit is 32, the size of the convolution kernel is 1 × 3, the number of channels of the DeformConv variable convolution layer is 32, the size of the convolution kernel is 3 × 3, in the step c-2.3), the size of the convolution kernel of the max pool layer is 1 × 4, and in the step c-2.4), the size of the convolution kernel of the max pool layer is 1 × 4.
Further, in step c-3.1), the number of channels of the first convolution layer and the second convolution layer of the first branch unit is 64, the size of the convolution kernel is 1 × 3, in step c-3.2), the number of channels of the convolution layer of the second branch unit is 64, the size of the convolution kernel is 1 × 3, the number of channels of the DeformConv variable convolution layer is 64, the size of the convolution kernel is 3 × 3, in step c-3.3), the size of the convolution kernel of the max pool layer is 1 × 4, and in step c-3.4), the size of the convolution kernel of the max pool layer is 1 × 4.
Further, in step c-4.1), the number of channels of the first convolution layer and the second convolution layer of the first branch unit is 128, the size of the convolution kernel is 1 × 3, in step c-4.2), the number of channels of the convolution layer of the second branch unit is 128, the size of the convolution kernel is 1 × 3, the number of channels of the DeformConv variable convolution layer is 128, the size of the convolution kernel is 3 × 3, in step c-4.3), the size of the convolution kernel of the Maxpool maximum pooling layer is 1 × 4, and in step c-4.4), the size of the convolution kernel of the Maxpool maximum pooling layer is 1 × 4.
Further, in step c-5.1), the number of channels of the first convolution layer and the second convolution layer of the first branch unit is 256, the size of the convolution kernel is 1 × 3, in step c-5.2), the number of channels of the convolution layer of the second branch unit is 256, the size of the convolution kernel is 1 × 3, the number of channels of the DeformConv variable convolution layer is 256, the size of the convolution kernel is 3 × 3, in step c-5.3), the size of the convolution kernel of the Maxpool layer is 1 × 4, and in step c-5.3), the number of channels of the first convolution layer and the number of channels of the second convolution layer are 256, the size of the convolution kernel is 1 × 3
c-5.4) Maxpool layer has a convolution kernel size of 1 × 4.
Further, step d) comprises the following steps:
d-1) by the formulaCalculating to obtain cross entropy classification lossIn the formulaN S Is the total number of samples in the source domain S,is the number of samples with class j in the source domain S, j is 1,2 b In order to train the batch size value,for the true tag value of the ith sample in the source domain S,the label prediction value of the ith sample in the source domain S is obtained;
d-2) by the formulaComputationally derived class certainty lossWhere C is the number of all classes, ω 1 And omega 2 Are sample deterministic weighting coefficients, T is a transpose,tag prediction values for target domainsPrediction probability matrix obtained by SoftmaxThe (c) th column of (a),tag prediction values for target domainsPrediction probability matrix obtained by SoftmaxColumn j';
d-3) by the formulaCalculating to obtain class diversity loss To predict a probability matrixThe rank of (d);
The invention has the beneficial effects that: the generalization capability of the model on multiple target domains is improved, the convergence speed is high, and the problem that the effect of a single target domain adaptive algorithm on the multiple target domains is poor is solved. In addition, the introduction of the deformable convolution kernel can consciously adjust the field of experience of convolution in the model training process, and better adapt to the lead characteristics of the multi-lead electrocardiosignals.
Drawings
Fig. 1 is a schematic diagram of a network structure according to the present invention.
Detailed Description
The invention is further described below with reference to fig. 1.
A multi-target domain electrocardiosignal classification method based on depth field self-adaptation comprises the following steps:
a) acquiring a plurality of 12-lead electrocardiosignals acquired by different acquisition equipment and different individuals, labeling the 12-lead electrocardiosignals, dividing the labeled 12-lead electrocardiosignals into a source domain S, and dividing the unlabeled 12-lead electrocardiosignals into a target domain combination T, wherein T is ═ T { T } 1 ,T 2 ,...,T k ,...,T M },T k For the kth target field, k is 1, 2.
b) And respectively and sequentially performing down-sampling processing, slicing processing and normalization processing on the source domain S and the target domain combination T to obtain a processed source domain S 'and a processed target domain combination T'.
c) Establishing a deep learning classification model of 12-lead electrocardiosignals, wherein the deep learning classification model of the 12-lead electrocardiosignals consists of a feature extractor and a full-connection classifier, inputting the processed source domain S 'and the processed target domain combination T' into the deep learning classification model of the 12-lead electrocardiosignals, and outputting to obtain a label prediction value of the source domainAnd label prediction of target domainsValue of
d) Calculating a loss functionAnd optimizing a deep learning classification model of the 12-lead electrocardiosignals by an Adam optimization algorithm.
e) And (3) respectively inputting the data of each target domain into the optimized deep learning classification model of the 12-lead electrocardiosignals, outputting to obtain a label predicted value of each target domain, and completing classification of the electrocardiosignals of the target domains.
The method is non-antagonistic, does not need characteristic alignment between two domains, focuses on the prediction method of the category accuracy and diversity of the target domain, improves the generalization capability of the model on the multi-target domain, has high convergence rate, and solves the problem of poor effect of a single target domain adaptive algorithm on the multi-target domain. In addition, the introduction of the deformable convolution kernel can consciously adjust the field of experience of convolution in the model training process, and better adapt to the lead characteristics of the multi-lead electrocardiosignals. The method is not only oriented to a single target domain, but also highly applicable to a plurality of target domains, improves the characteristic distribution difference phenomenon existing among different domain samples, and solves the problem of difficult labeling of the electrical data of a center in practical application.
Example 1:
further, the method for processing the slices in the step b) comprises the following steps: the method comprises the steps of down-sampling all 12-lead electrocardiosignals in a source domain S or a target domain combination T to have the same frequency, randomly intercepting 30S of electrocardiosignals, intercepting 12-lead electrocardiosignals with the data length exceeding 30S, and repeatedly filling 12-lead electrocardiosignals with the data length being less than 30S.
Example 2:
further, step c) comprises the following steps:
c-1) the feature extractor sequentially comprises a first residual error structure, a second residual error structure, a third residual error structure and a fourth residual error structure.
c-2.1) the first residual error structure is composed of a first branch unit, a second branch unit and a third branch unit, wherein the first branch unit is composed of a first convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a second convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and the processed source domain S 'and the processed target domain combination T' are respectively input into the first branch unit and then respectively output to obtain the lead internal characteristic of the 12-lead electrocardiosignalAnd features of
c-2.2) the second branch unit of the first residual error structure is composed of a convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a DeformConv variable convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and the processed source domain S 'and the processed target domain combination T' are respectively input into the second branch unit and then respectively output to obtain the inter-lead characteristics of the 12-lead electrocardiosignalsAnd features of
c-2.3) the third branch unit of the first residual error structure is composed of a convolution layer and a Maxpool layer in sequence, and the processed source domain S 'and the processed target domain combination T' are respectively input into the third branch unit and then respectively output to obtain the characteristics of 12-lead electrocardiosignalsAnd features of
c-2.4) in-lead characterizationAnd features between leadsInputting the spliced features into a Maxpool layer and outputting the Maxpool layer to obtain the featuresInner features of leadsAnd features between leadsAfter feature splicing, inputting the feature into a Maxpool layer of the Maxpool layer and outputting the feature to obtain features
c-2.5) characterization ofAnd features ofOverlap to form a new feature F 1 S Will be characterized byAnd features ofPerforming superposition to form new features
c-3.1) second residual error StructureThe first branch unit consists of a first convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a second convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and respectively forms a characteristic F 1 S And features ofRespectively output the internal lead characteristics of the obtained 12-lead electrocardiosignals after being input into the first branch unitAnd features of
c-3.2) the second branch unit of the second residual structure sequentially comprises a convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a DeformConv variable convolution layer, a second batch of normalization layers and a second ReLu activation function layer, and respectively forms the characteristic F 1 S And featuresThe inter-lead characteristics of 12-lead electrocardiosignals are respectively output after being input into the second branch unitAnd features of
c-3.3) the third branch unit of the second residual structure is composed of a convolution layer and a Maxpool layer in sequence, and respectively converts the characteristic F 1 S And featuresRespectively outputting the signals to obtain the characteristics of 12-lead electrocardiosignals after being input into a third branch unitAnd features of
c-3.4) in-lead characterizationAnd inter-lead characteristicsInputting the spliced features into a Maxpool layer and outputting the Maxpool layer to obtain the featuresInner features of leadsAnd features between leadsInputting the spliced features into a Maxpool layer with the Maxpool to output the spliced features to obtain the features
c-3.5) characterization ofAnd features ofPerforming superposition to form new featuresWill be characterized byAnd features ofAre superposed to form a newFeature(s)
c-4.1) the third residual error structure is composed of a first branch unit, a second branch unit and a third branch unit, wherein the first branch unit is composed of a first convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a second convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and the characteristics of the first branch unit, the first batch of normalization layers and the second batch of normalization layers are respectively characterizedAnd featuresRespectively output the internal lead characteristics of the obtained 12-lead electrocardiosignals after being input into the first branch unitAnd features of
c-4.2) the second branch unit of the third residual error structure sequentially comprises a convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a DeformConv variable convolution layer, a second batch of normalization layers and a second ReLu activation function layer, and the characteristics of the second branch unit are respectively represented byAnd featuresThe inter-lead characteristics of 12-lead electrocardiosignals are respectively output after being input into the second branch unitAnd features of
c-4.3) second of the third residual StructureThe three-branch unit is composed of a convolution layer and a Maxpool layer in sequence, and features of the three-branch unit are respectivelyAnd featuresRespectively outputting the signals to obtain the characteristics of 12-lead electrocardiosignals after being input into a third branch unitAnd features of
c-4.4) in-lead characterizationAnd features between leadsAfter feature splicing, inputting the feature into a Maxpool layer of the Maxpool layer and outputting the feature to obtain featuresInner features of leadsAnd features between leadsInputting the spliced features into a Maxpool layer and outputting the Maxpool layer to obtain the features
c-4.5 characterization ofAnd features ofOverlap to form a new feature F 3 S Will be characterized byAnd features ofPerforming superposition to form new features
c-5.1) the fourth residual structure is composed of a first branch unit, a second branch unit and a third branch unit, wherein the first branch unit is composed of a first convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a second convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and the characteristics F are respectively formed by 3 S And featuresRespectively output the internal lead characteristics of the obtained 12-lead electrocardiosignals after being input into the first branch unitAnd features of
c-5.2) the second branch unit of the fourth residual structure sequentially comprises a convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a DeformConv variable convolution layer, a second batch of normalization layers and a second ReLu activation function layer, and respectively forms the characteristic F 3 S And featuresThe inter-lead characteristics of 12-lead electrocardiosignals are respectively output after being input into the second branch unitAnd is characterized bySign for
c-5.3) the third branch unit of the fourth residual structure is composed of a convolution layer and a Maxpool layer in sequence, and the characteristics F are respectively obtained 3 S And featuresRespectively outputting the signals to obtain the characteristics of 12-lead electrocardiosignals after being input into a third branch unitAnd features of
c-5.4) in-lead characterizationAnd features between leadsAfter feature splicing, inputting the feature into a Maxpool layer of the Maxpool layer and outputting the feature to obtain featuresInner features of leadsAnd features between leadsInputting the spliced features into a Maxpool layer and outputting the Maxpool layer to obtain the features
c-5.5) characterization ofAnd features ofPerforming superposition to form new featuresWill be characterized byAnd features ofPerforming superposition to form new features
c-6) characterization ofAnd features ofInputting the data into a full-connection classifier, and outputting the data to obtain a label predicted value of a source domainAnd label prediction value of target domain
Example 3:
further, in the step c-2.1), the number of channels of the first convolution layer and the second convolution layer of the first branch unit is 32, the size of the convolution kernel is 1 × 3, in the step c-2.2), the number of channels of the convolution layer of the second branch unit is 32, the size of the convolution kernel is 1 × 3, the number of channels of the DeformConv variable convolution layer is 32, the size of the convolution kernel is 3 × 3, in the step c-2.3), the size of the convolution kernel of the Maxpool maximum pooling layer is 1 × 4, and in the step c-2.4), the size of the convolution kernel of the Maxpool maximum pooling layer is 1 × 4.
Example 4:
further, in step c-3.1), the number of channels of the first convolution layer and the second convolution layer of the first branch unit is 64, the size of the convolution kernel is 1 × 3, in step c-3.2), the number of channels of the convolution layer of the second branch unit is 64, the size of the convolution kernel is 1 × 3, the number of channels of the DeformConv variable convolution layer is 64, the size of the convolution kernel is 3 × 3, in step c-3.3), the size of the convolution kernel of the max pool layer is 1 × 4, and in step c-3.4), the size of the convolution kernel of the max pool layer is 1 × 4.
Example 5:
further, in step c-4.1), the number of channels of the first convolution layer and the second convolution layer of the first branch unit is 128, the size of the convolution kernel is 1 × 3, in step c-4.2), the number of channels of the convolution layer of the second branch unit is 128, the size of the convolution kernel is 1 × 3, the number of channels of the DeformConv variable convolution layer is 128, the size of the convolution kernel is 3 × 3, in step c-4.3), the size of the convolution kernel of the Maxpool maximum pooling layer is 1 × 4, and in step c-4.4), the size of the convolution kernel of the Maxpool maximum pooling layer is 1 × 4.
Example 6:
further, in step c-5.1), the number of channels of the first convolution layer and the second convolution layer of the first branch unit is 256, the size of the convolution kernel is 1 × 3, in step c-5.2), the number of channels of the convolution layer of the second branch unit is 256, the size of the convolution kernel is 1 × 3, the number of channels of the DeformConv variable convolution layer is 256, the size of the convolution kernel is 3 × 3, in step c-5.3), the size of the convolution kernel of the max pool layer is 1 × 4, and in step c-5.4), the size of the convolution kernel of the max pool layer is 1 × 4.
Example 7:
further, step d) comprises the following steps:
the training process is divided into three branches, the first branch is supervised training of the source domain data of the label, and the label prediction value of the source domain dataAnd label prediction value of target domainComputing cross entropy classification lossThe second branch is a pseudo label prediction matrix for each target domain dataClass-deterministic loss function forming a target domainThe third branch is a pseudo label prediction matrix of the target domain dataConstructed class diversity loss functionThe three components form a combined lossThe parameters of the network model are optimized through an Adam optimization algorithm, so that the joint loss is minimum, the model has good classification performance, the class prediction accuracy of a target domain test set is better, and the model trained by the target domain in a source domain is highly applicable. Specifically, the method comprises the following steps:
d-1) by the formulaCalculating to obtain cross entropy classification lossIn the formulaN S Is the total number of samples in the source domain S,the number of samples of the source domain S with the category j, j 1,2Amount, N b In order to train the batch size value,for the true tag value of the ith sample in the source domain S,and predicting the label of the ith sample in the source domain S.
d-2) by the formulaComputationally derived class certainty lossWhere C is the number of all classes, ω 1 And omega 2 Are sample deterministic weighting coefficients, T is a transpose,tag prediction values for target domainsPrediction probability matrix obtained by SoftmaxThe (c) th column of (a),tag prediction values for target domainsPrediction probability matrix obtained by SoftmaxColumn j' of (1).
d-3) by the formulaCalculating to obtain category diversityLoss of power To predict a probability matrixIs determined.
Finally, it should be noted that: although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that changes may be made in the embodiments and/or equivalents thereof without departing from the spirit and scope of the invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (8)
1. A multi-target domain electrocardiosignal classification method based on depth field self-adaptation is characterized by comprising the following steps:
a) acquiring a plurality of 12-lead electrocardiosignals acquired by different acquisition equipment and different individuals, labeling the 12-lead electrocardiosignals, dividing the labeled 12-lead electrocardiosignals into a source domain S, and dividing the unlabeled 12-lead electrocardiosignals into a target domain combination T, wherein T is ═ T { T } 1 ,T 2 ,...,T k ,...,T M },T k The number of target domains is k, wherein k is 1, 2.
b) Respectively and sequentially performing down-sampling processing, slicing processing and normalization processing on the source domain S and the target domain combination T to obtain a processed source domain S 'and a processed target domain combination T';
c) establishing a deep learning classification model of 12-lead electrocardiosignals, wherein the deep learning classification model of the 12-lead electrocardiosignals consists of a feature extractor and a full-connection classifier, inputting the processed source domain S 'and the processed target domain combination T' into the deep learning classification model of the 12-lead electrocardiosignals, and outputting to obtain a label prediction value of the source domainAnd label prediction value of target domain
d) Calculating a loss functionOptimizing a deep learning classification model of the 12-lead electrocardiosignal by an Adam optimization algorithm;
e) and (3) respectively inputting the data of each target domain into the optimized deep learning classification model of the 12-lead electrocardiosignals, outputting to obtain a label predicted value of each target domain, and completing classification of the electrocardiosignals of the target domains.
2. The multi-target domain electrocardiosignal classification method based on the depth domain self-adaptation as claimed in claim 1, wherein the slicing processing method in step b) is as follows: the method comprises the steps of down-sampling all 12-lead electrocardiosignals in a source domain S or a target domain combination T to have the same frequency, randomly intercepting 30S of electrocardiosignals, intercepting 12-lead electrocardiosignals with the data length exceeding 30S, and repeatedly filling 12-lead electrocardiosignals with the data length being less than 30S.
3. The method for classifying multi-target domain electrocardiosignals based on depth domain self-adaptation according to claim 1, wherein the step c) comprises the following steps:
c-1) the feature extractor sequentially comprises a first residual error structure, a second residual error structure, a third residual error structure and a fourth residual error structure;
c-2.1) the first residual error structure is composed of a first branch unit, a second branch unit and a third branch unit, wherein the first branch unit is composed of a first convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a second convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and the processed source domain S 'and the processed target domain combination T' are respectively input into the first branch unit and then respectively output to obtain the lead internal characteristic of the 12-lead electrocardiosignalAnd features of
c-2.2) the second branch unit of the first residual error structure is composed of a convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a DeformConv variable convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and the processed source domain S 'and the processed target domain combination T' are respectively input into the second branch unit and then respectively output to obtain the inter-lead characteristics of the 12-lead electrocardiosignalsAnd features of
c-2.3) the third branch unit of the first residual error structure is composed of a convolution layer and a Maxpool layer in sequence, and the processed source domain S 'and the processed target domain combination T' are respectively input into the third branch unit and then respectively output to obtain the characteristics of 12-lead electrocardiosignalsAnd features of
c-2.4) in-lead characterizationAnd features between leadsInputting the spliced features into a Maxpool layer and outputting the Maxpool layer to obtain the featuresInner features of leadsAnd features between leadsInputting the spliced features into a Maxpool layer with the Maxpool to output the spliced features to obtain the features
c-2.5) characterization ofAnd features ofOverlap to form a new feature F 1 S Will be characterized byAnd features ofPerforming superposition to form new features
c-3.1) the second residual structure is composed of a first branch unit, a second branch unit and a third branch unit, wherein the first branch unit is composed of a first convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a second convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and the characteristics F are respectively formed by 1 S And featuresRespectively output the internal lead characteristics of the obtained 12-lead electrocardiosignals after being input into the first branch unitAnd features of
c-3.2) the second branch unit of the second residual structure sequentially comprises a convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a DeformConv variable convolution layer, a second batch of normalization layers and a second ReLu activation function layer, and respectively forms the characteristic F 1 S And featuresThe inter-lead characteristics of 12-lead electrocardiosignals are respectively output after being input into the second branch unitAnd features of
c-3.3) the third branch unit of the second residual structure is composed of a convolution layer and a Maxpool layer in sequence, and respectively converts the characteristic F 1 S And featuresRespectively outputting the signals to obtain the characteristics of 12-lead electrocardiosignals after being input into a third branch unitAnd features of
c-3.4) in-lead characterizationAnd features between leadsInputting the spliced features into a Maxpool layer to output the obtained featuresInner features of leadsAnd features between leadsAfter feature splicing, inputting the feature into a Maxpool layer of the Maxpool layer and outputting the feature to obtain features
c-3.5) characterization ofAnd features ofPerforming superposition to form new featuresWill be characterized byAnd features ofPerforming superposition to form new features
c-4.1) the third residual error structure is composed of a first branch unit, a second branch unit and a third branch unit, wherein the first branch unit is composed of a first convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a second convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and the characteristics of the first branch unit, the first batch of normalization layers and the second batch of normalization layers are respectively characterizedAnd featuresRespectively output the internal lead characteristics of the obtained 12-lead electrocardiosignals after being input into the first branch unitAnd features of
c-4.2) the second branch unit of the third residual error structure sequentially comprises a convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a DeformConv variable convolution layer, a second batch of normalization layers and a second ReLu activation function layer, and the characteristics of the second branch unit are respectively represented byAnd featuresThe inter-lead characteristics of 12-lead electrocardiosignals are respectively output after being input into the second branch unitAnd features of
c-4.3) the third branch unit of the third residual structure is composed of a convolution layer and a Maxpool layer in sequence, and features of the convolution layer and the Maxpool layer are respectivelyAnd featuresRespectively outputting the signals to obtain the characteristics of 12-lead electrocardiosignals after being input into a third branch unitAnd features of
c-4.4) in-lead characterizationAnd features between leadsInputting the spliced features into a Maxpool layer and outputting the Maxpool layer to obtain the featuresInner features of leadsAnd features between leadsAfter feature splicing, inputting the feature into a Maxpool layer of the Maxpool layer and outputting the feature to obtain features
c-4.5) characterization ofAnd features ofPerforming superposition to form new featuresWill be characterized byAnd features ofPerforming superposition to form new features
c-5.1) the fourth residual structure is composed of a first branch unit, a second branch unit and a third branch unit, wherein the first branch unit is composed of a first convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a second convolution layer, a second batch of normalization layers and a second ReLu activation function layer in sequence, and the characteristics are respectively formedAnd featuresRespectively output the internal lead characteristics of the obtained 12-lead electrocardiosignals after being input into the first branch unitAnd features of
c-5.2) the second branch unit of the fourth residual structure sequentially comprises a convolution layer, a first batch of normalization layers, a first ReLu activation function layer, a DeformConv variable convolution layer, a second batch of normalization layers and a second ReLu activation function layer, and the characteristics are respectively formedAnd featuresThe inter-lead characteristics of 12-lead electrocardiosignals are respectively output after being input into the second branch unitAnd features of
c-5.3) the third branch unit of the fourth residual structure is composed of a convolution layer and a Maxpool layer in sequence, and the characteristics are respectivelyAnd featuresRespectively outputting the signals to obtain the characteristics of 12-lead electrocardiosignals after being input into a third branch unitAnd features of
c-5.4) characterization of leadsAnd features between leadsInputting the spliced features into a Maxpool layer to output the obtained featuresInner features of leadsAnd features between leadsAfter feature splicing, inputting the feature into a Maxpool layer of the Maxpool layer and outputting the feature to obtain features
c-5.5) characterization ofAnd features ofOverlap to form new featuresWill be characterized byAnd features ofPerforming superposition to form new features
4. The method for classifying multi-target domain electrocardiosignals based on depth domain self-adaptation according to claim 1, which is characterized in that: in the step c-2.1), the number of channels of the first convolution layer and the second convolution layer of the first branch unit is 32, the size of convolution kernels is 1 × 3, in the step c-2.2), the number of channels of the convolution layer of the second branch unit is 32, the size of convolution kernels is 1 × 3, in the DeformConv variable convolution layer, the number of channels is 32, the size of convolution kernels is 3 × 3, in the step c-2.3), the size of convolution kernels of the Maxpool maximum pooling layer is 1 × 4, and in the step c-2.4), the size of convolution kernels of the Maxpool maximum pooling layer is 1 × 4.
5. The method for classifying multi-target domain electrocardiosignals based on depth domain self-adaptation according to claim 1, which is characterized in that: in the step c-3.1), the number of channels of the first convolution layer and the second convolution layer of the first branch unit is 64, the size of convolution kernels is 1 × 3, the number of channels of the convolution layer of the second branch unit in the step c-3.2) is 64, the size of convolution kernels is 1 × 3, the number of channels of the DeformConv variable convolution layer is 64, the size of convolution kernels is 3 × 3, the size of convolution kernels of the Maxpool maximum pooling layer in the step c-3.3) is 1 × 4, and the size of convolution kernels of the Maxpool maximum pooling layer in the step c-3.4) is 1 × 4.
6. The method for classifying multi-target domain electrocardiosignals based on depth domain self-adaptation according to claim 1, which is characterized in that: the number of channels of the first convolution layer and the second convolution layer of the first branch unit in the step c-4.1) is 128, the size of convolution kernels is 1 x 3, the number of channels of the convolution layer of the second branch unit in the step c-4.2) is 128, the size of convolution kernels is 1 x 3, the number of channels of the DeformConv variable convolution layer is 128, the size of convolution kernels is 3 x 3, the size of convolution kernels of the Maxpool maximum pooling layer in the step c-4.3) is 1 x 4, and the size of convolution kernels of the Maxpool maximum pooling layer in the step c-4.4) is 1 x 4.
7. The method for classifying multi-target domain electrocardiosignals based on depth domain self-adaptation according to claim 1, which is characterized in that: the number of channels of the first convolution layer and the second convolution layer of the first branch unit in the step c-5.1) is 256, the sizes of convolution kernels are 1 × 3, the number of channels of the convolution layer of the second branch unit in the step c-5.2) is 256, the sizes of convolution kernels are 1 × 3, the number of channels of the DeformConv variable convolution layer is 256, the sizes of convolution kernels are 3 × 3, the size of convolution kernel of the Maxpool layer in the step c-5.3) is 1 × 4, and the size of convolution kernel of the Maxpool layer in the step c-5.4) is 1 × 4.
8. The method for classifying multi-target domain electrocardiosignals based on depth domain self-adaptation according to claim 1, wherein the step d) comprises the following steps:
d-1) by the formulaCalculating to obtain cross entropy classification lossIn the formulaN S Is the total number of samples in the source domain S,is the number of samples with class j in the source domain S, j is 1,2 b In order to train the batch size value,for the true tag value of the ith sample in the source domain S,the label prediction value of the ith sample in the source domain S is obtained;
d-2) by the formulaComputationally derived class certainty lossWhere C is the number of all classes, ω 1 And omega 2 Are sample deterministic weighting coefficients, T is a transpose,tag prediction values for target domainsPrediction probability matrix obtained by SoftmaxThe (c) th column of (a),tag prediction values for target domainsPrediction by SoftmaxProbability matrixColumn j';
d-3) by the formulaCalculating to obtain class diversity loss To predict a probability matrixThe rank of (d);
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210526548.4A CN114841209A (en) | 2022-05-13 | 2022-05-13 | Multi-target domain electrocardiosignal classification method based on depth field self-adaption |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210526548.4A CN114841209A (en) | 2022-05-13 | 2022-05-13 | Multi-target domain electrocardiosignal classification method based on depth field self-adaption |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114841209A true CN114841209A (en) | 2022-08-02 |
Family
ID=82569616
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210526548.4A Pending CN114841209A (en) | 2022-05-13 | 2022-05-13 | Multi-target domain electrocardiosignal classification method based on depth field self-adaption |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114841209A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116778969A (en) * | 2023-06-25 | 2023-09-19 | 山东省人工智能研究院 | Domain-adaptive heart sound classification method based on double-channel cross attention |
CN117257322A (en) * | 2023-09-21 | 2023-12-22 | 齐鲁工业大学(山东省科学院) | Multi-label electrocardiosignal classification method based on dual-branch network |
-
2022
- 2022-05-13 CN CN202210526548.4A patent/CN114841209A/en active Pending
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116778969A (en) * | 2023-06-25 | 2023-09-19 | 山东省人工智能研究院 | Domain-adaptive heart sound classification method based on double-channel cross attention |
CN116778969B (en) * | 2023-06-25 | 2024-03-01 | 山东省人工智能研究院 | Domain-adaptive heart sound classification method based on double-channel cross attention |
CN117257322A (en) * | 2023-09-21 | 2023-12-22 | 齐鲁工业大学(山东省科学院) | Multi-label electrocardiosignal classification method based on dual-branch network |
CN117257322B (en) * | 2023-09-21 | 2024-04-19 | 齐鲁工业大学(山东省科学院) | Multi-label electrocardiosignal classification method based on dual-branch network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107066759B (en) | Steam turbine rotor vibration fault diagnosis method and device | |
CN114841209A (en) | Multi-target domain electrocardiosignal classification method based on depth field self-adaption | |
CN108764317B (en) | Residual convolutional neural network image classification method based on multipath feature weighting | |
CN109190464B (en) | Mechanical fault intelligent diagnosis method based on transfer learning under variable working conditions | |
CN111414942A (en) | Remote sensing image classification method based on active learning and convolutional neural network | |
CN113095409B (en) | Hyperspectral image classification method based on attention mechanism and weight sharing | |
CN112561910A (en) | Industrial surface defect detection method based on multi-scale feature fusion | |
CN113076920B (en) | Intelligent fault diagnosis method based on asymmetric domain confrontation self-adaptive model | |
CN112381763A (en) | Surface defect detection method | |
CN109472733A (en) | Image latent writing analysis method based on convolutional neural networks | |
CN112766340B (en) | Depth capsule network image classification method and system based on self-adaptive spatial mode | |
CN114355240A (en) | Power distribution network ground fault diagnosis method and device | |
CN114636975A (en) | LPI radar signal identification method based on spectrogram fusion and attention mechanism | |
CN111553186A (en) | Electromagnetic signal identification method based on depth long-time and short-time memory network | |
CN113221614A (en) | Power grid insulator damage image identification method based on hybrid neural network | |
CN113239949A (en) | Data reconstruction method based on 1D packet convolutional neural network | |
CN116011507A (en) | Rare fault diagnosis method for fusion element learning and graph neural network | |
CN111768792A (en) | Audio steganalysis method based on convolutional neural network and domain confrontation learning | |
CN113109782B (en) | Classification method directly applied to radar radiation source amplitude sequence | |
CN113592008B (en) | System, method, device and storage medium for classifying small sample images | |
CN112989955B (en) | Human body action recognition method based on space-time double-flow heterogeneous grafting convolutional neural network | |
CN113052099B (en) | SSVEP classification method based on convolutional neural network | |
CN116312860B (en) | Agricultural product soluble solid matter prediction method based on supervised transfer learning | |
CN107886108A (en) | Fruit and vegetable classification method and device based on AlexNet network models | |
CN109460788B (en) | Hyperspectral image classification method based on low-rank-sparse information combination network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |