ECG signal multi-waveform detection method based on improved encoder-decoder structure

Abstract

A multi-waveform detection method for ECG signals based on an improved encoder-decoder structure relates to the technical field of ECG signal processing. In the encoder module, the combination of CNN and dilated convolutional network can realize multi-scale extraction of ECG signal features, effectively capture feature information at different levels, and reduce the number of parameters, thereby improving computational efficiency. In the residual connection module, the lower-level features transmitted from the encoder are nonlinearly changed and fused with the higher-level features from the decoder. This technology can effectively alleviate the gradient vanishing and gradient exploding problems of deep neural networks during training. In the decoder module, the use of unpooling operations combined with convolutional layers can fill the corresponding area of ​​the output upsampled feature map with the same value as the input feature map, thereby realizing refined extraction of ECG signal features and recovery of spatial information.

Description

Technical Field

[0001] The invention relates to the technical field of electrocardiogram signal processing, and in particular to an electrocardiogram signal multi-waveform detection method based on an improved encoder-decoder structure. Background Art

[0002] Cardiovascular diseases pose a huge threat to human life and health. Electrocardiogram examination is widely used as a non-invasive tool for diagnosing cardiovascular diseases. This painless and non-invasive method can record the electrical activity of myocardial cells on the body surface. Diagnostic experts can judge different types of heart diseases by analyzing the patient's electrocardiogram. However, manual analysis of electrocardiograms is not only time-consuming and laborious, but also requires strong professional knowledge. Therefore, patients with abnormal heart rhythms must consult diagnostic experts and may miss the best time for effective treatment. At present, automatic analysis of ECG signals has become an important means of cardiovascular disease detection. As a prerequisite for automatic analysis of ECG signals, waveform detection of ECG signals has always been a research hotspot in the field of ECG signal processing. Accurately identifying the various components of the ECG waveform is very important for the correct diagnosis of various heart diseases. ECG analysis includes detecting P waves, QRS waves and T waves, and then analyzing their shapes, amplitudes, relative positions, etc. Since different types of waveforms may overlap and interfere with each other, feature extraction is difficult and information loss is also likely to occur, which also poses a challenge to mixed waveform detection. Therefore, choosing a suitable method to accurately extract the features of mixed waveforms is of great significance for improving the early diagnosis and treatment of cardiovascular diseases. Summary of the invention

[0003] In order to overcome the deficiencies of the above technologies, the present invention provides an ECG signal multi-waveform detection method which can effectively restore the original ECG signal information from coarse to fine and improve the waveform segmentation performance.

[0004] The technical solution adopted by the present invention to overcome the technical problems is:

[0005] A method for detecting multiple waveforms of electrocardiogram signals based on an improved encoder-decoder structure comprises the following steps:

[0006] a) Obtain the first lead electrocardiogram record from the LUDB public dataset;

[0007] b) resampling the acquired electrocardiogram record of the first lead to obtain an electrocardiogram segment set U consisting of Z segments of duration Xs, where U = {U1, U2, ..., U t ,…,U Z},U t is the t-th signal segment, and the t-th signal segment U tThere are four categories: p wave, QRS wave, t wave, and no wave, t∈{1,2,…,Z}; c) establishing a waveform detection network ED_SegNet, which consists of an encoder, a first residual connection block Res Path, a second residual connection block Res Path, a third residual connection block Res Path, a fourth residual connection block Res Path, a decoder, a fully connected layer, and a softmax function;

[0008] d) The tth signal segment U t Input into the encoder of the waveform detection network ED_SegNet, and output features f1, f2, f3, and f4 respectively;

[0009] e) Input feature f1 into the first residual connection block Res Path of the waveform detection network ED_SegNet, and output feature f a , input the feature f2 into the second residual connection block Res Path of the waveform detection network ED_SegNet, and output the feature f b , input the feature f3 into the third residual connection block Res Path of the waveform detection network ED_SegNet, and output the feature f c , input the feature f4 into the fourth residual connection block Res Path of the waveform detection network ED_SegNet, and output the feature f d ;

[0010] f) The feature f a , feature f b , feature f c , feature f d They are respectively input into the decoder of the waveform detection network ED_SegNet, and the outputs are feature f8, feature f7, feature f6, and feature f5;

[0011] g) Input feature f8 into the waveform detection network ED_SegNet in sequence, and output the predicted probability of the category p wave by the fully connected layer and softmax function of the encoder. Predicted probability of QRS complex Predicted probability of category being t wave Predicted probability of class no wave By formula Calculate the average value P of the waveform prediction n .

[0012] Preferably, in step b), the acquired electrocardiogram record of the first lead is resampled at a frequency of 250 Hz to obtain an electrocardiogram signal segment set U consisting of 10 signal segments with a duration of 10 s.

[0013] Further, step d) comprises the following steps:

[0014] d-1) The encoder of the waveform detection network ED_SegNet is composed of the first convolution block, the second convolution block, the third convolution block, and the fourth convolution block. The first convolution block, the second convolution block, the third convolution block, and the fourth convolution block are all composed of the first convolution layer, the second convolution layer, the DAC module, and the maximum pooling layer. The DAC module is composed of a convolution layer and a hole convolution layer;

[0015] d-2) The tth signal segment U t Input into the first convolutional layer of the first convolutional block, and output the feature f 1_1 , the feature f 1_1 Input into the second convolutional layer of the first convolutional block, and output the feature f 1_2 , the feature f 1_2 Input to the convolution layer of the DAC module of the first convolution block, and output the feature f 1_2 ′, the feature f 1_2 Input to the hole convolution layer of the DAC module of the first convolutional block, and output the feature f 1_2 ″, the feature f 1_2 ′ and feature f 1_2 ″ performs concatenation operation to obtain feature f 11 , the feature f 11 With feature f 1_2 Perform concatenation operation to obtain feature f 1_3 , the feature f 1_3 Input into the maximum pooling layer of the first convolutional block, and output the feature f1;

[0016] d-3) Input feature f1 into the first convolutional layer of the second convolutional block and output feature f 2_1 , the feature f 2_1 Input to the second convolutional layer of the second convolutional block, and output the feature f 2_2 , the feature f 2_2 Input to the convolution layer of the DAC module of the second convolution block, and output the feature f 2_2 ′, the feature f 2_2 Input to the hole convolution layer of the DAC module of the second convolutional block, and output the feature f 2_2 ″, the feature f 2_2 ′ and feature f 2_2 ″ performs concatenation operation to obtain feature f 21 , the feature f 21 With feature f 2_2 Perform concatenation operation to obtain feature f 2_3 , the feature f 2_3Input to the maximum pooling layer of the second convolutional block, and output the feature f2;

[0017] d-4) Input feature f2 into the first convolutional layer of the third convolutional block and output feature f 3_1 , the feature f 3_1 Input to the second convolution layer of the third convolution block, and output the feature f 3_2 , the feature f 3_2 Input to the convolution layer of the DAC module of the third convolution block, and output the feature f 3_2 ′, the feature f 3_2 Input to the hole convolution layer of the DAC module of the third convolutional block, and output the feature f 3_2 ″, the feature f 3_2 ′ and feature f 3_2 ″ performs concatenation operation to obtain feature f 31 , the feature f 31 With feature f 3_2 Perform concatenation operation to obtain feature f 3_3 , the feature f 3_3 Input to the maximum pooling layer of the third convolutional block, and output the feature f3;

[0018] d-5) Input feature f3 into the first convolutional layer of the fourth convolutional block and output feature f 4_1 , the feature f 4_1 Input to the second convolution layer of the fourth convolution block, and output the feature f 4_2 , the feature f 4_2 Input to the convolution layer of the DAC module of the fourth convolutional block, and output the feature f 4_2 ′, the feature f 4_2 Input to the hole convolution layer of the DAC module of the fourth convolutional block, and output the feature f 4_2 ″, the feature f 4_2 ′ and feature f 4_2 ″ performs concatenation operation to obtain feature f 41 , the feature f 41 With feature f 4_2 Perform concatenation operation to obtain feature f 4_3 , the feature f 4_3 Input into the maximum pooling layer of the fourth convolutional block and output the feature f4.

[0019] Preferably, in step d-2), the convolution kernel size of the first convolution layer of the first convolution block is 3×1, the step size is 1, the padding is 1, and the number of channels is 64, the convolution kernel size of the second convolution layer of the first convolution block is 3×1, the step size is 1, the padding is 3, and the number of channels is 64, the convolution kernel size of the convolution layer of the DAC module of the first convolution block is 7×1, the step size is 3, and the padding is 1, the convolution kernel size of the hole convolution layer of the DAC module of the first convolution block is 3×1, the step size is 3, the padding is 2, and the hole convolution rate is 2; in step d-3), the convolution kernel size of the first convolution layer of the second convolution block is 3×1, the step size is 1, the padding is 1, and the number of channels is 128; the convolution kernel size of the second convolution layer of the second convolution block is 3×1, the step size is 1, the padding is 1, and the number of channels is 128; the convolution kernel size of the convolution layer of the DAC module of the second convolution block is 7×1, the step size is 3, and the padding is 1; the convolution kernel size of the hole convolution layer of the DAC module of the second convolution block is 3×1, the step size is 3, the padding is 2, and the hole convolution rate is 2; in d-4), the convolution kernel size of the first convolution layer of the third convolution block is 3×1, the stride is 1, the padding is 1, and the number of channels is 256; the convolution kernel size of the second convolution layer of the third convolution block is 3×1, the stride is 1, the padding is 1, and the number of channels is 256; the convolution kernel size of the convolution layer of the DAC module of the third convolution block is 7×1, the stride is 3, and the padding is 1; the convolution kernel size of the hole convolution layer of the DAC module of the third convolution block is 3×1, the stride is 3, the padding is 2, and the hole convolution rate is 2; d-5), the convolution kernel size of the first convolution layer of the fourth convolution block is 3×1, the stride is 1, the padding is 1, and the number of channels is 512; the convolution kernel size of the second convolution layer of the fourth convolution block is 3×1, the stride is 1, the padding is 1, and the number of channels is 512; the convolution kernel size of the convolution layer of the DAC module of the fourth convolution block is 7×1, the stride is 3, and the padding is 1; the convolution kernel size of the dilated convolution layer of the DAC module of the fourth convolution block is 3×1, the stride is 3, the padding is 2, and the dilated convolution rate is 2.

[0020] Further, step e) comprises the following steps:

[0021] e-1) The first residual connection block Res Path of the waveform detection network ED_SegNet is composed of the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer. The feature f1 is input into the first convolutional layer of the first residual connection block Res Path, and the feature f is output. a_1 , input the feature f1 into the second convolutional layer of the first residual connection block Res Path, and output the feature f a_2 , the feature f a_1 With feature f a_2 Perform concatenation operation to obtain feature f a_3 , the feature f a_3Input to the third convolutional layer of the first residual connection block Res Path, and output the feature f a_4 , the feature f a_3 Input to the fourth convolutional layer of the first residual connection block ResPath, and output the feature f a_5 , the feature f a_4 With feature f a_5 Perform concatenation operation to obtain feature f a ; e-2) The second residual connection block Res Path of the waveform detection network ED_SegNet is composed of the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer. The feature f2 is input into the first convolutional layer of the second residual connection block Res Path, and the feature f is output. b_1 , input the feature f2 into the second convolutional layer of the second residual connection block Res Path, and output the feature f b_2 , the feature f b_1 With feature f b_2 Perform concatenation operation to obtain feature f b_3 , the feature f b_3 Input to the third convolutional layer of the second residual connection block ResPath, and output the feature f b_4 , the feature f b_3 Input to the fourth convolutional layer of the second residual connection block Res Path, and output the feature f b_5 , the feature f b_4 With feature f b_5 Perform concatenation operation to obtain feature f b ; e-3) The third residual connection block Res Path of the waveform detection network ED_SegNet is composed of the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer. The feature f3 is input into the first convolutional layer of the third residual connection block Res Path, and the feature f is output. c_1 , input the feature f3 into the second convolutional layer of the third residual connection block Res Path, and output the feature f c_2 , the feature f c_1 With feature f c_2 Perform concatenation operation to obtain feature f c_3 , the feature f c_3 Input to the third convolutional layer of the third residual connection block Res Path, and output the feature f c_4 , the feature f c_3 Input to the fourth convolutional layer of the third residual connection block Res Path, and output the feature f c_5 , the feature f c_4 With feature f c_5 Perform concatenation operation to obtain feature f c; e-4) The fourth residual connection block Res Path of the waveform detection network ED_SegNet is composed of the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer. The feature f4 is input into the first convolutional layer of the fourth residual connection block Res Path, and the feature f is output. d_1 , input the feature f4 into the second convolutional layer of the fourth residual connection block Res Path, and output the feature f d_2 , the feature f d_1 With feature f d_2 Perform concatenation operation to obtain feature f d_3 , the feature f d_3 Input to the third convolutional layer of the fourth residual connection block Res Path, and output the feature f d_4 , the feature f d_3 Input to the fourth convolutional layer of the fourth residual connection block Res Path, and output the feature f d_5 , the feature f d_4 With feature f d_5 Perform concatenation operation to obtain feature f d .

[0022] Preferably, the convolution kernel size of the first convolution layer of the first residual connection block Res Path, the second residual connection block Res Path, the third residual connection block Res Path, and the fourth residual connection block Res Path is 7×1, the step size is 3, and the padding is 1. The convolution kernel size of the second convolution layer of the first residual connection block Res Path, the second residual connection block Res Path, the third residual connection block Res Path, and the fourth residual connection block Res Path is 3×1, the step size is 1, and the padding is 1. The convolution kernel size of the third convolution layer of the first residual connection block Res Path, the second residual connection block Res Path, the third residual connection block Res Path, and the fourth residual connection block Res Path is 7×1, the step size is 3, and the padding is 1. The first residual connection block Res Path, the second residual connection block Res Path, and the third residual connection block Res The convolution kernel size of the fourth convolutional layer of Path and the fourth residual connection block ResPath is 3×1, the step size is 1, and the padding is 1.

[0023] Further, step f) comprises the following steps:

[0024] f-1) The decoder of the waveform detection network ED_SegNet consists of the first convolution block, the second convolution block, the third convolution block, and the fourth convolution block;

[0025] f-2) The fourth convolutional block of the decoder consists of an anti-pooling layer, a first convolutional layer, and a second convolutional layer, which transforms the feature fd Input to the depooling layer of the fourth convolutional block, and output the feature f 5_1 , the feature f 5_1 Input into the first convolutional layer of the fourth convolutional block, and output the feature f 5_2 , the feature f 5_2 Input into the second convolutional layer of the fourth convolutional block, and output feature f5;

[0026] f-3) The third convolutional block of the decoder consists of an unpooling layer, a first convolutional layer, and a second convolutional layer. c Concatenate with feature f5 to get feature f 6_1 ′, the feature f 6_1 ′ is input to the depooling layer of the third convolutional block, and the output is feature f 6_1 , the feature f 6_1 Input into the first convolutional layer of the third convolutional block, and output the feature f 6_2 , the feature f 6_2 Input into the second convolution layer of the third convolution block, and output feature f6;

[0027] f-4) The second convolutional block of the decoder consists of an anti-pooling layer, a first convolutional layer, and a second convolutional layer. b Concatenate with feature f6 to get feature f 7_1 ′, the feature f 7_1 ′ is input to the unpooling layer of the second convolutional block, and the output is feature f 7_1 , the feature f 7_1 Input into the first convolutional layer of the second convolutional block, and output the feature f 7_2 , the feature f 7_2 Input into the second convolutional layer of the second convolutional block, and output feature f7;

[0028] f-5) The first convolutional block of the decoder consists of an unpooling layer, a first convolutional layer, and a second convolutional layer. a Concatenate with feature f7 to get feature f 8_1 ′, the feature f 8_1 ′ is input into the unpooling layer of the first convolutional block, and the output is feature f 8_1 , the feature f 8_1 Input into the first convolutional layer of the first convolutional block, and output the feature f 8_2 , the feature f 8_2 Input into the second convolutional layer of the first convolutional block, and output feature f8.

[0029] Preferably, in step f-2), the convolution kernel size of the first convolution layer and the second convolution layer of the fourth convolution block of the decoder are both 5×1, the step size is 1, the padding is 2, and the number of channels is 512; in step f-3), the convolution kernel size of the first convolution layer and the second convolution layer of the third convolution block of the decoder are both 5×1, the step size is 1, the padding is 2, and the number of channels is 256; in f-4), the convolution kernel size of the first convolution layer and the second convolution layer of the second convolution block of the decoder are both 5×1, the step size is 1, the padding is 2, and the number of channels is 128; in f-5), the convolution kernel size of the first convolution layer and the second convolution layer of the first convolution block of the decoder are both 5×1, the step size is 1, the padding is 2, and the number of channels is 64.

[0030] Furthermore, it also includes using an Adam optimizer to train the waveform detection network ED_SegNet using a cross entropy loss function to obtain an optimized waveform detection network ED_SegNet.

[0031] Preferably, when training the waveform detection network ED_SegNet, the number of training rounds is set to 60, the learning rate is set to le-4, and the batch size is set to 32.

[0032] The beneficial effects of the present invention are as follows: the ECG signal waveform detection network ED_SegNet uses a combination of encoder, decoder and residual connection, connects the feature set of ECG signals at all levels in the coding structure to the decoder part for multi-scale decoding, can effectively restore the original ECG signal from coarse to fine information, and improve the waveform segmentation performance. In the encoder module, the combination of CNN and hole convolution network can realize multi-scale extraction of ECG signal features, effectively capture feature information at different levels, while reducing the number of parameters and improving computational efficiency. This combination can also improve the model's ability to capture local details and global context in ECG signals, and help to more accurately characterize important features in ECG signals. In the residual connection module, the lower-level features transmitted from the encoder are nonlinearly changed and fused with the higher-level features from the decoder. This technology can effectively alleviate the gradient vanishing and gradient explosion problems of deep neural networks during training, is conducive to the training and optimization of deeper network models, and improves the network's feature transmission and information flow efficiency. In the decoder module, the combination of unpooling operation and convolution layer can fill the corresponding area of ​​the output upsampled feature map with the same value as the input feature map, realize the refined extraction of ECG signal features and the recovery of spatial information, which helps to enhance the model's abstract expression ability of ECG signal features. This method uses the improved encoder-decoder structure and residual connection, combined with CNN, dilated convolution and unpooling operation, to achieve accurate detection of P, QRS and T waveforms in ECG signals, thus providing an efficient and accurate solution for ECG waveform detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a structural diagram of the waveform detection network ED_SegNet of the present invention;

[0034] Figure 2 is a structural diagram of a DAC module of the present invention;

[0035] Figure 3 It is a structural diagram of the residual connection block Res Path of the present invention;

[0036] Figure 4 This is the confusion matrix of the model tested on the LU database of the present invention. DETAILED DESCRIPTION

[0037] The following is combined with Figure 1 , Attachment Figure 2 , Attachment Figure 3 , Attachment Figure 4 The present invention is further described.

[0038] A method for detecting multiple waveforms of electrocardiogram signals based on an improved encoder-decoder structure comprises the following steps:

[0039] a) Obtain the first lead ECG record from the LUDB public dataset.

[0040] b) resampling the acquired electrocardiogram record of the first lead to obtain an electrocardiogram segment set U consisting of Z segments of duration Xs, where U = {U1, U2, ..., U t ,…,U Z},U t is the t-th signal segment, and the t-th signal segment U t It includes four categories: p wave, QRS wave, t wave, and no wave, t∈{1,2,…,Z}. c) Establish a waveform detection network ED_SegNet, which consists of an encoder, a first residual connection block Res Path, a second residual connection block Res Path, a third residual connection block Res Path, a fourth residual connection block Res Path, a decoder, a fully connected layer, and a softmax function.

[0041] d) The tth signal segment U t Input into the encoder of the waveform detection network ED_SegNet, and output features f1, f2, f3, and f4 respectively.

[0042] e) Input feature f1 into the first residual connection block Res Path of the waveform detection network ED_SegNet, and output feature f a, input the feature f2 into the second residual connection block Res Path of the waveform detection network ED_SegNet, and output the feature f b , input the feature f3 into the third residual connection block Res Path of the waveform detection network ED_SegNet, and output the feature f c , input the feature f4 into the fourth residual connection block Res Path of the waveform detection network ED_SegNet, and output the feature f d .

[0043] f) The feature f a , feature f b , feature f c , feature f d They are respectively input into the decoder of the waveform detection network ED_SegNet, and the outputs are feature f8, feature f7, feature f6, and feature f5.

[0044] g) Input feature f8 into the waveform detection network ED_SegNet in sequence, and output the predicted probability of the category p wave by the fully connected layer and softmax function of the encoder. Predicted probability of QRS complex Predicted probability of category being t wave Predicted probability of class no wave By formula Calculate the average value P of the waveform prediction n .

[0045] The encoder part extracts low-level features, among which the DAC module uses dilated convolution and standard convolution with different convolution kernel sizes. The dilated convolution can obtain receptive fields of different sizes by setting convolution operations with different dilation rates. The combination of the two can extract a large number of important ECG signal features. The residual connection can make the model pay more attention to important features and contextual information, thereby reducing the difference between low-level features and high-level features. The decoder part extracts high-level features and can upsample and reconstruct features, so it can pay more attention to local details and small feature changes than the encoder. In addition, a multi-scale decoding structure is also adopted, and the feature set output results of the ECG signals at all levels in the encoding structure are directly short-circuited to the decoding structure. The hierarchical decoding is merged as the final feature output. This operation can effectively reduce information loss during the encoding process.

[0046] This paper proposes a new waveform detection network ED_SegNet, which can better capture the features in the ECG signal sequence, retain important boundary positions and detail information, and reduce the loss of information in the feature extraction process. The network can better capture the important feature information of the ECG signal, which is conducive to improving the detection accuracy of P, QRS, and T waveforms in the ECG signal.

[0047] In order to verify the reliability of this patented method, a systematic evaluation of the final prediction segmentation accuracy is conducted. The comparison of the ECG waveform (P, QRS, T and NW) classification Acc of the waveform detection network ED_SegNet with other state-of-the-art methods on the LU database is shown in Table 1:

[0048] Table 1 Comparison of different experimental methods

[0049]

[0050] In our experimental method, we compared a variety of segmentation methods, namely 2BiLSTM, 1CNN_2BiLSTM, 3CNN_2BiLSTM, 4CNN_2BiLSTM, 2-UNet cascade, 3-UNet cascade, and ECG_SegNet. Each method segmented the LU dataset, and the Acc evaluation index was used to count the accuracy values ​​of segmenting P waves, QRS waves, T waves, and NW waves. The higher the accuracy value, the better the performance. From the data in Table 1, it can be seen that in the classification of P waves, QRS waves, T waves, and NW waves in the LU database, ED_SegNet has a higher Acc, and the average accuracy of ECG waveform classification is 96.53%. It can be seen that ED_SegNet has a better performance in segmentation, followed by ECG_SegNet, 3-UNet cascade, 2-UNet cascade, 3CNN_2BiLSTM, and 4CNN_1BiLSTM. However, 1CNN_2BiLSTM and 2BiLSTM have poor performance, especially for P wave and NW classification. It can be seen that the Acc evaluation index of P wave detection by the method of the present invention is improved by 0.2 compared with the previous optimal method, the Acc evaluation index of QRS wave detection is improved by 0.81 compared with the previous optimal method, the Acc evaluation index of T wave detection is improved by 1.54 compared with the previous optimal method, and the Acc evaluation index of NW wave detection is improved by 1.95 compared with the previous optimal method, all of which are better than the previous methods. Figure 1 This is the confusion matrix of the model tested on the LU database. From Table 1, it can be seen that the average Acc of P, QRS, T and NW waveform classification are 98.02%, 98.54%, 96.67% and 92.88% respectively. The model has the best classification effect on QRS wave, followed by P wave, T wave and NW wave. The above results show that the method of the present invention has certain advantages and the segmentation effect is relatively ideal.

[0051] In one embodiment of the present invention, the acquired electrocardiogram record of the first lead is resampled at a frequency of 250 Hz in step b) to obtain an electrocardiogram signal segment set U consisting of 10 signal segments with a duration of 10 s. In one embodiment of the present invention, step d) includes the following steps:

[0052] d-1) The encoder of the waveform detection network ED_SegNet is composed of the first convolution block, the second convolution block, the third convolution block, and the fourth convolution block. The first convolution block, the second convolution block, the third convolution block, and the fourth convolution block are all composed of the first convolution layer, the second convolution layer, the DAC module, and the maximum pooling layer. The DAC module is composed of a convolution layer and a hole convolution layer.

[0053] d-2) The tth signal segment U t Input into the first convolutional layer of the first convolutional block, and output the feature f 1_1 , the feature f 1_1 Input into the second convolutional layer of the first convolutional block, and output the feature f 1_2 , the feature f 1_2 Input to the convolution layer of the DAC module of the first convolution block, and output the feature f 1_2 ′, the feature f 1_2 Input to the hole convolution layer of the DAC module of the first convolutional block, and output the feature f 1_2 ″, the feature f 1_2 ′ and feature f 1_2 ″ performs concatenation operation to obtain feature f 11 , the feature f 11 With feature f 1_2 Perform concatenation operation to obtain feature f 1_3 , the feature f 1_3 Input into the maximum pooling layer of the first convolutional block and output the feature f1.

[0054] d-3) Input feature f1 into the first convolutional layer of the second convolutional block and output feature f 2_1 , the feature f 2_1 Input to the second convolutional layer of the second convolutional block, and output the feature f 2_2 , the feature f 2_2 Input to the convolution layer of the DAC module of the second convolution block, and output the feature f 2_2 ′, the feature f 2_2 Input to the hole convolution layer of the DAC module of the second convolutional block, and output the feature f 2_2 ″, the feature f 2_2 ′ and feature f 2_2 ″ performs concatenation operation to obtain feature f 21 , the feature f 21 With feature f2_2 Perform concatenation operation to obtain feature f 2_3 , the feature f 2_3 Input into the maximum pooling layer of the second convolutional block and output the feature f2.

[0055] d-4) Input feature f2 into the first convolutional layer of the third convolutional block and output feature f 3_1 , the feature f 3_1 Input to the second convolution layer of the third convolution block, and output the feature f 3_2 , the feature f 3_2 Input to the convolution layer of the DAC module of the third convolution block, and output the feature f 3_2 ′, the feature f 3_2 Input to the hole convolution layer of the DAC module of the third convolutional block, and output the feature f 3_2 ″, the feature f 3_2 ′ and feature f 3_2 ″ performs concatenation operation to obtain feature f 31 , the feature f 31 With feature f 3_2 Perform concatenation operation to obtain feature f 3_3 , the feature f 3_3 Input into the maximum pooling layer of the third convolutional block and output the feature f3.

[0056] d-5) Input feature f3 into the first convolutional layer of the fourth convolutional block and output feature f 4_1 , the feature f 4_1 Input to the second convolution layer of the fourth convolution block, and output the feature f 4_2 , the feature f 4_2 Input to the convolution layer of the DAC module of the fourth convolutional block, and output the feature f 4_2 ′, the feature f 4_2 Input to the hole convolution layer of the DAC module of the fourth convolutional block, and output the feature f 4_2 ″, the feature f 4_2 ′ and feature f 4_2 ″ performs concatenation operation to obtain feature f 41 , the feature f 41 With feature f 4_2 Perform concatenation operation to obtain feature f 4_3 , the feature f 4_3 Input into the maximum pooling layer of the fourth convolutional block and output the feature f4.

[0057] In this embodiment, in step d-2), the convolution kernel size of the first convolution layer of the first convolution block is 3×1, the step size is 1, the padding is 1, and the number of channels is 64; the convolution kernel size of the second convolution layer of the first convolution block is 3×1, the step size is 1, the padding is 3, and the number of channels is 64; the convolution kernel size of the convolution layer of the DAC module of the first convolution block is 7×1, the step size is 3, and the padding is 1; the convolution kernel size of the hole convolution layer of the DAC module of the first convolution block is 3×1, the step size is 3, the padding is 2, and the hole convolution kernel size ... The product rate is 2; in step d-3), the convolution kernel size of the first convolution layer of the second convolution block is 3×1, the stride is 1, the padding is 1, and the number of channels is 128; the convolution kernel size of the second convolution layer of the second convolution block is 3×1, the stride is 1, the padding is 1, and the number of channels is 128; the convolution kernel size of the convolution layer of the DAC module of the second convolution block is 7×1, the stride is 3, and the padding is 1; the convolution kernel size of the hole convolution layer of the DAC module of the second convolution block is 3×1, the stride is 3, the padding is 2, and the hole convolution kernel size is The product rate is 2; in d-4), the convolution kernel size of the first convolution layer of the third convolution block is 3×1, the stride is 1, the padding is 1, and the number of channels is 256. The convolution kernel size of the second convolution layer of the third convolution block is 3×1, the stride is 1, the padding is 1, and the number of channels is 256. The convolution kernel size of the convolution layer of the DAC module of the third convolution block is 7×1, the stride is 3, and the padding is 1. The convolution kernel size of the hole convolution layer of the DAC module of the third convolution block is 3×1, the stride is 3, the padding is 2, and the hole convolution rate is is 2; in d-5), the convolution kernel size of the first convolution layer of the fourth convolution block is 3×1, the stride is 1, the padding is 1, and the number of channels is 512; the convolution kernel size of the second convolution layer of the fourth convolution block is 3×1, the stride is 1, the padding is 1, and the number of channels is 512; the convolution kernel size of the convolution layer of the DAC module of the fourth convolution block is 7×1, the stride is 3, and the padding is 1; the convolution kernel size of the dilated convolution layer of the DAC module of the fourth convolution block is 3×1, the stride is 3, the padding is 2, and the dilated convolution rate is 2.

[0058] In one embodiment of the present invention, step e) comprises the following steps:

[0059] e-1) The first residual connection block Res Path of the waveform detection network ED_SegNet is composed of the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer. The feature f1 is input into the first convolutional layer of the first residual connection block Res Path, and the feature f is output. a_1 , input the feature f1 into the second convolutional layer of the first residual connection block Res Path, and output the feature f a_2 , the feature f a_1 With feature f a_2 Perform concatenation operation to obtain feature f a_3 , the feature f a_3Input to the third convolutional layer of the first residual connection block Res Path, and output the feature f a_4 , the feature f a_3 Input to the fourth convolutional layer of the first residual connection block ResPath, and output the feature f a_5 , the feature f a_4 With feature f a_5 Perform concatenation operation to obtain feature f a e-2) The second residual connection block Res Path of the waveform detection network ED_SegNet is composed of the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer. The feature f2 is input into the first convolutional layer of the second residual connection block Res Path, and the feature f is output. b_1 , input the feature f2 into the second convolutional layer of the second residual connection block Res Path, and output the feature f b_2 , the feature f b_1 With feature f b_2 Perform concatenation operation to obtain feature f b_3 , the feature f b_3 Input to the third convolutional layer of the second residual connection block ResPath, and output the feature f b_4 , the feature f b_3 Input to the fourth convolutional layer of the second residual connection block Res Path, and output the feature f b_5 , the feature f b_4 With feature f b_5 Perform concatenation operation to obtain feature f b e-3) The third residual connection block Res Path of the waveform detection network ED_SegNet is composed of the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer. The feature f3 is input into the first convolutional layer of the third residual connection block Res Path, and the feature f is output. c_1 , input the feature f3 into the second convolutional layer of the third residual connection block Res Path, and output the feature f c_2 , the feature f c_1 With feature f c_2 Perform concatenation operation to obtain feature f c_3 , the feature f c_3 Input to the third convolutional layer of the third residual connection block Res Path, and output the feature f c_4 , the feature f c_3 Input to the fourth convolutional layer of the third residual connection block Res Path, and output the feature f c_5 , the feature f c_4 With feature f c_5 Perform concatenation operation to obtain feature f ce-4) The fourth residual connection block Res Path of the waveform detection network ED_SegNet is composed of the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer. The feature f4 is input into the first convolutional layer of the fourth residual connection block Res Path, and the feature f is output. d_1 , input the feature f4 into the second convolutional layer of the fourth residual connection block Res Path, and output the feature f d_2 , the feature f d_1 With feature f d_2 Perform concatenation operation to obtain feature f d_3 , the feature f d_3 Input to the third convolutional layer of the fourth residual connection block Res Path, and output the feature f d_4 , the feature f d_3 Input to the fourth convolutional layer of the fourth residual connection block Res Path, and output the feature f d_5 , the feature f d_4 With feature f d_5 Perform concatenation operation to obtain feature f d .

[0060] In this embodiment, the convolution kernel size of the first convolution layer of the first residual connection block Res Path, the second residual connection block Res Path, the third residual connection block Res Path, and the fourth residual connection block Res Path is 7×1, the step size is 3, and the padding is 1. The convolution kernel size of the second convolution layer of the first residual connection block Res Path, the second residual connection block Res Path, the third residual connection block Res Path, and the fourth residual connection block Res Path is 3×1, the step size is 1, and the padding is 1. The convolution kernel size of the third convolution layer of the first residual connection block Res Path, the second residual connection block Res Path, the third residual connection block Res Path, and the fourth residual connection block Res Path is 7×1, the step size is 3, and the padding is 1. The convolution kernel size of the fourth convolutional layer of Path is 3×1, the stride is 1, and the padding is 1.

[0061] In one embodiment of the present invention, step f) comprises the following steps:

[0062] f-1) The decoder of the waveform detection network ED_SegNet consists of the first convolution block, the second convolution block, the third convolution block, and the fourth convolution block.

[0063] f-2) The fourth convolutional block of the decoder consists of an anti-pooling layer, a first convolutional layer, and a second convolutional layer, which transforms the feature f d Input to the depooling layer of the fourth convolutional block, and output the feature f 5_1 , the feature f 5_1 Input into the first convolutional layer of the fourth convolutional block, and output the feature f 5_2 , the feature f 5_2 Input into the second convolutional layer of the fourth convolutional block, and output the feature f5.

[0064] f-3) The third convolutional block of the decoder consists of an unpooling layer, a first convolutional layer, and a second convolutional layer. c Concatenate with feature f5 to get feature f 6_1 ′, the feature f 6_1 ′ is input to the depooling layer of the third convolutional block, and the output is feature f 6_1 , the feature f 6_1 Input into the first convolutional layer of the third convolutional block, and output the feature f 6_2 , the feature f 6_2 Input into the second convolution layer of the third convolution block, and output feature f6.

[0065] f-4) The second convolutional block of the decoder consists of an anti-pooling layer, a first convolutional layer, and a second convolutional layer. b Concatenate with feature f6 to get feature f 7_1 ′, the feature f 7_1 ′ is input to the unpooling layer of the second convolutional block, and the output is feature f 7_1 , the feature f 7_1 Input into the first convolutional layer of the second convolutional block, and output the feature f 7_2 , the feature f 7_2 Input into the second convolutional layer of the second convolutional block, and output feature f7.

[0066] f-5) The first convolutional block of the decoder consists of an unpooling layer, a first convolutional layer, and a second convolutional layer. a Concatenate with feature f7 to get feature f 8_1 ′, the feature f 8_1 ′ is input into the unpooling layer of the first convolutional block, and the output is feature f 8_1 , the feature f 8_1 Input into the first convolutional layer of the first convolutional block, and output the feature f 8_2 , the feature f 8_2 Input into the second convolutional layer of the first convolutional block, and output feature f8.

[0067] In this embodiment, in step f-2), the convolution kernel size of the first convolution layer and the second convolution layer of the fourth convolution block of the decoder are both 5×1, the step size is 1, the padding is 2, and the number of channels is 512; in step f-3), the convolution kernel size of the first convolution layer and the second convolution layer of the third convolution block of the decoder are both 5×1, the step size is 1, the padding is 2, and the number of channels is 256; in f-4), the convolution kernel size of the first convolution layer and the second convolution layer of the second convolution block of the decoder are both 5×1, the step size is 1, the padding is 2, and the number of channels is 128; in f-5), the convolution kernel size of the first convolution layer and the second convolution layer of the first convolution block of the decoder are both 5×1, the step size is 1, the padding is 2, and the number of channels is 64.

[0068] In one embodiment of the present invention, it also includes using an Adam optimizer to train the waveform detection network ED_SegNet using a cross entropy loss function to obtain an optimized waveform detection network ED_SegNet.

[0069] In this embodiment, when training the waveform detection network ED_SegNet, the number of training rounds is set to 60, the learning rate is set to le-4, and the batch size is set to 32.

[0070] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting multiple waveforms of electrocardiogram signals based on an improved encoder-decoder structure, characterized in that: The steps include: a) Obtain the first lead electrocardiogram record from the LUDB public dataset; b) resampling the acquired electrocardiogram record of the first lead to obtain an electrocardiogram segment set U consisting of Z segments of duration Xs, where U = {U1, U2, ..., U t ,…,U Z },U t is the t-th signal segment, and the t-th signal segment U t There are four categories: p wave, QRS wave, t wave, and no wave, t∈{1,2,…,Z}; c) establishing a waveform detection network ED_SegNet, which consists of an encoder, a first residual connection block Res Path, a second residual connection block Res Path, a third residual connection block Res Path, a fourth residual connection block ResPath, a decoder, a fully connected layer, and a softmax function; d) The tth signal segment U t Input into the encoder of the waveform detection network ED_SegNet, and output features f1, f2, f3, and f4 respectively; e) Input feature f1 into the first residual connection block Res Path of the waveform detection network ED_SegNet, and output feature f a , input the feature f2 into the second residual connection block Res Path of the waveform detection network ED_SegNet, and output the feature f b , input the feature f3 into the third residual connection block Res Path of the waveform detection network ED_SegNet, and output the feature f c , input the feature f4 into the fourth residual connection block Res Path of the waveform detection network ED_SegNet, and output the feature f d ; f) The feature f a , feature f b , feature f c , feature f d They are respectively input into the decoder of the waveform detection network ED_SegNet, and the outputs are feature f8, feature f7, feature f6, and feature f5; g) Input feature f8 into the waveform detection network ED_SegNet in sequence, and output the predicted probability of the category p wave by the fully connected layer and softmax function of the encoder. Predicted probability of QRS complex Predicted probability of category being t wave Predicted probability of class no wave By formula Calculate the average value P of the waveform prediction n ; Step d) comprises the following steps: d-1) The encoder of the waveform detection network ED_SegNet is composed of the first convolution block, the second convolution block, the third convolution block, and the fourth convolution block. The first convolution block, the second convolution block, the third convolution block, and the fourth convolution block are all composed of the first convolution layer, the second convolution layer, the DAC module, and the maximum pooling layer. The DAC module is composed of a convolution layer and a hole convolution layer; d-2) The tth signal segment U t Input into the first convolutional layer of the first convolutional block, and output the feature f 1_1 , the feature f 1_1 Input into the second convolutional layer of the first convolutional block, and output the feature f 1_2 , the feature f 1_2 Input to the convolution layer of the DAC module of the first convolution block, and output the feature f 1_2 ′, the feature f 1_2 Input to the hole convolution layer of the DAC module of the first convolutional block, and output the feature f 1_2 ″, the feature f 1_2 ′ and feature f 1_2 ″ performs concatenation operation to obtain feature f 11 , the feature f 11 With feature f 1_2 Perform concatenation operation to obtain feature f 1_3 , the feature f 1_3 Input into the maximum pooling layer of the first convolutional block, and output the feature f1; d-3) Input feature f1 into the first convolutional layer of the second convolutional block and output feature f 2_1 , the feature f 2_1 Input to the second convolutional layer of the second convolutional block, and output the feature f 2_2 , the feature f 2_2 Input to the convolution layer of the DAC module of the second convolution block, and output the feature f 2_2 ′, the feature f 2_2 Input to the hole convolution layer of the DAC module of the second convolutional block, and output the feature f 2_2 ″, the feature f 2_2 ′ and feature f 2_2 ″ performs concatenation operation to obtain feature f 21 , the feature f 21 With feature f 2_2 Perform concatenation operation to obtain feature f 2_3 , the feature f 2_3 Input to the maximum pooling layer of the second convolutional block, and output the feature f2; d-4) Input feature f2 into the first convolutional layer of the third convolutional block and output feature f 3_1 , the feature f 3_1 Input to the second convolution layer of the third convolution block, and output the feature f 3_2 , the feature f 3_2 Input to the convolution layer of the DAC module of the third convolution block, and output the feature f 3_2 ′, the feature f 3_2 Input to the hole convolution layer of the DAC module of the third convolutional block, and output the feature f 3_2 ″, the feature f 3_2 ′ and feature f 3_2 ″ performs concatenation operation to obtain feature f 31 , the feature f 31 With feature f 3_2 Perform concatenation operation to obtain feature f 3_3 , the feature f 3_3 Input to the maximum pooling layer of the third convolutional block, and output the feature f3; d-5) Input feature f3 into the first convolutional layer of the fourth convolutional block and output feature f 4_1 , the feature f 4_1 Input to the second convolution layer of the fourth convolution block, and output the feature f 4_2 , the feature f 4_2 Input to the convolution layer of the DAC module of the fourth convolutional block, and output the feature f 4_2 ′, the feature f 4_2 Input to the hole convolution layer of the DAC module of the fourth convolutional block, and output the feature f 4_2 ″, the feature f 4_2 ′ and feature f 4_2 ″ performs concatenation operation to obtain feature f 41 , the feature f 41 With feature f 4_2 Perform concatenation operation to obtain feature f 4_3 , the feature f 4_3 Input to the maximum pooling layer of the fourth convolutional block, and output the feature f4; Step e) comprises the following steps: e-1) The first residual connection block Res Path of the waveform detection network ED_SegNet is composed of the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer. The feature f1 is input into the first convolutional layer of the first residual connection block Res Path, and the feature f is output. a_1 , input the feature f1 into the second convolutional layer of the first residual connection block Res Path, and output the feature f a_2 , the feature f a_1 With feature f a_2 Perform concatenation operation to obtain feature f a_3 , the feature f a_3 Input to the third convolutional layer of the first residual connection block Res Path, and output the feature f a_4 , the feature f a_3 Input to the fourth convolutional layer of the first residual connection block Res Path, and output the feature f a_5 , the feature f a_4 With feature f a_5 Perform concatenation operation to obtain feature f a ; e-2) The second residual connection block Res Path of the waveform detection network ED_SegNet is composed of the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer. The feature f2 is input into the first convolutional layer of the second residual connection block Res Path, and the feature f is output. b_1 , input the feature f2 into the second convolutional layer of the second residual connection block Res Path, and output the feature f b_2 , the feature f b_1 With feature f b_2 Perform concatenation operation to obtain feature f b_3 , the feature f b_3 Input to the third convolutional layer of the second residual connection block Res Path, and output the feature f b_4 , the feature f b_3 Input to the fourth convolutional layer of the second residual connection block Res Path, and output the feature f b_5 , the feature f b_4 With feature f b_5 Perform concatenation operation to obtain feature f b ; e-3) The third residual connection block Res Path of the waveform detection network ED_SegNet is composed of the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer. The feature f3 is input into the first convolutional layer of the third residual connection block Res Path, and the feature f is output. c_1 , input the feature f3 into the second convolutional layer of the third residual connection block Res Path, and output the feature f c_2 , the feature f c_1 With feature f c_2 Perform concatenation operation to obtain feature f c_3 , the feature f c_3 Input to the third convolutional layer of the third residual connection block Res Path, and output the feature f c_4 , the feature f c_3 Input to the fourth convolutional layer of the third residual connection block Res Path, and output the feature f c_5 , the feature f c_4 With feature f c_5 Perform concatenation operation to obtain feature f c ; e-4) The fourth residual connection block Res Path of the waveform detection network ED_SegNet is composed of the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer. The feature f4 is input into the first convolutional layer of the fourth residual connection block Res Path, and the feature f is output. d_1 , input the feature f4 into the second convolutional layer of the fourth residual connection block Res Path, and output the feature f d_2 , the feature f d_1 With feature f d_2 Perform concatenation operation to obtain feature f d_3 , the feature f d_3 Input to the third convolutional layer of the fourth residual connection block Res Path, and output the feature f d_4 , the feature f d_3 Input to the fourth convolutional layer of the fourth residual connection block Res Path, and output the feature f d_5 , the feature f d_4 With feature f d_5 Perform concatenation operation to obtain feature f d ; Step f) comprises the following steps: f-1) The decoder of the waveform detection network ED_SegNet consists of the first convolution block, the second convolution block, the third convolution block, and the fourth convolution block; f-2) The fourth convolutional block of the decoder consists of an anti-pooling layer, a first convolutional layer, and a second convolutional layer, which transforms the feature f d Input to the depooling layer of the fourth convolutional block, and output the feature f 5_1 , the feature f 5_1 Input into the first convolutional layer of the fourth convolutional block, and output the feature f 5_2 , the feature f 5_2 Input into the second convolutional layer of the fourth convolutional block, and output feature f5; f-3) The third convolutional block of the decoder consists of an unpooling layer, a first convolutional layer, and a second convolutional layer. c Concatenate with feature f5 to get feature f 6_1 ′, the feature f 6_1 ′ is input to the depooling layer of the third convolutional block, and the output is feature f 6_1 , the feature f 6_1 Input into the first convolutional layer of the third convolutional block, and output the feature f 6_2 , the feature f 6_2 Input into the second convolution layer of the third convolution block, and output feature f6; f-4) The second convolutional block of the decoder consists of an anti-pooling layer, a first convolutional layer, and a second convolutional layer. b Concatenate with feature f6 to get feature f 7_1 ′, the feature f 7_1 ′ is input to the unpooling layer of the second convolutional block, and the output is feature f 7_1 , the feature f 7_1 Input into the first convolutional layer of the second convolutional block, and output the feature f 7_2 , the feature f 7_2 Input into the second convolutional layer of the second convolutional block, and output feature f7; f-5) The first convolutional block of the decoder consists of an unpooling layer, a first convolutional layer, and a second convolutional layer. a Concatenate with feature f7 to get feature f 8_1 ′, the feature f 8_1 ′ is input into the unpooling layer of the first convolutional block, and the output is feature f 8_1 , the feature f 8_1 Input into the first convolutional layer of the first convolutional block, and output the feature f 8_2 , the feature f 8_2 Input into the second convolutional layer of the first convolutional block, and output feature f8.

2. The method for detecting multiple waveforms of electrocardiographic signals based on an improved encoder-decoder structure according to claim 1, characterized in that: In step b), the acquired electrocardiogram record of the first lead is resampled at a frequency of 250 Hz to obtain an electrocardiogram signal segment set U consisting of 10 signal segments with a duration of 10 s.

3. The method for detecting multiple waveforms of electrocardiographic signals based on an improved encoder-decoder structure according to claim 1, characterized in that: In step d-2), the convolution kernel size of the first convolution layer of the first convolution block is 3×1, the step size is 1, the padding is 1, and the number of channels is 64; the convolution kernel size of the second convolution layer of the first convolution block is 3×1, the step size is 1, the padding is 3, and the number of channels is 64; the convolution kernel size of the convolution layer of the DAC module of the first convolution block is 7×1, the step size is 3, and the padding is 1; the convolution kernel size of the hole convolution layer of the DAC module of the first convolution block is 3×1, the step size is 3, the padding is 2, and the hole convolution rate is 2; In step d-3), the convolution kernel size of the first convolution layer of the second convolution block is 3×1, the step size is 1, the padding is 1, and the number of channels is 128. The convolution kernel size of the second convolution layer of the second convolution block is 3×1, the step size is 1, the padding is 1, and the number of channels is 128. The convolution kernel size of the convolution layer of the DAC module of the second convolution block is 7×1, the step size is 3, and the padding is 1. The convolution kernel size of the hole convolution layer of the DAC module of the second convolution block is 3×1, the step size is 3, the padding is 2, and the hole convolution rate is 2 ; In d-4), the convolution kernel size of the first convolution layer of the third convolution block is 3×1, the stride is 1, the padding is 1, and the number of channels is 256. The convolution kernel size of the second convolution layer of the third convolution block is 3×1, the stride is 1, the padding is 1, and the number of channels is 256. The convolution kernel size of the convolution layer of the DAC module of the third convolution block is 7×1, the stride is 3, and the padding is 1. The convolution kernel size of the hole convolution layer of the DAC module of the third convolution block is 3×1, the stride is 3, the padding is 2, and the hole convolution rate is 2 ; In d-5), the convolution kernel size of the first convolution layer of the fourth convolution block is 3×1, the stride is 1, the padding is 1, and the number of channels is 512; the convolution kernel size of the second convolution layer of the fourth convolution block is 3×1, the stride is 1, the padding is 1, and the number of channels is 512; the convolution kernel size of the convolution layer of the DAC module of the fourth convolution block is 7×1, the stride is 3, and the padding is 1; the convolution kernel size of the dilated convolution layer of the DAC module of the fourth convolution block is 3×1, the stride is 3, the padding is 2, and the dilated convolution rate is 2.

4. The method for detecting multiple waveforms of electrocardiographic signals based on an improved encoder-decoder structure according to claim 1, characterized in that: The convolution kernel size of the first convolution layer of the first residual connection block Res Path, the second residual connection block Res Path, the third residual connection block Res Path, and the fourth residual connection block Res Path is 7×1, the step size is 3, and the padding is 1. The convolution kernel size of the second convolution layer of the first residual connection block Res Path, the second residual connection block Res Path, the third residual connection block Res Path, and the fourth residual connection block Res Path is 3×1, the step size is 1, and the padding is 1. The convolution kernel size of the third convolution layer of the first residual connection block Res Path, the second residual connection block Res Path, the third residual connection block Res Path, and the fourth residual connection block Res Path is 7×1, the step size is 3, and the padding is 1. The convolution kernel size of the fourth convolution layer of the first residual connection block Res Path, the second residual connection block Res Path, the third residual connection block Res Path, and the fourth residual connection block Res Path is 3×1, the step size is 1, and the padding is 1.

5. The method for detecting multiple waveforms of electrocardiographic signals based on an improved encoder-decoder structure according to claim 1, characterized in that: In step f-2), the convolution kernel size of the first convolution layer and the second convolution layer of the fourth convolution block of the decoder are both 5×1, the step size is 1, the padding is 2, and the number of channels is 512; in step f-3), the convolution kernel size of the first convolution layer and the second convolution layer of the third convolution block of the decoder are both 5×1, the step size is 1, the padding is 2, and the number of channels is 256; in f-4), the convolution kernel size of the first convolution layer and the second convolution layer of the second convolution block of the decoder are both 5×1, the step size is 1, the padding is 2, and the number of channels is 128; in f-5), the convolution kernel size of the first convolution layer and the second convolution layer of the first convolution block of the decoder are both 5×1, the step size is 1, the padding is 2, and the number of channels is 64.

6. The method for detecting multiple waveforms of electrocardiographic signals based on an improved encoder-decoder structure according to claim 1, characterized in that: It also includes using the Adam optimizer to train the waveform detection network ED_SegNet using the cross entropy loss function to obtain the optimized waveform detection network ED_SegNet.

7. The method for detecting multiple waveforms of electrocardiographic signals based on an improved encoder-decoder structure according to claim 6, characterized in that: When training the waveform detection network ED_SegNet, the number of training rounds is set to 60, the learning rate is set to le-4, and the batch size is set to 32.

Images