An ECG identity recognition method based on quality assessment and deep transfer learning

Abstract

The present invention discloses an ECG identity recognition method based on quality assessment and deep transfer learning. First, an ECG signal quality assessment algorithm based on Support Vector Machine (SVM) is proposed to classify the original ECG signal into three different quality levels. Second, a wavelet transform denoising algorithm is proposed to denoise the "suspicious" signal, thereby obtaining a denoised "qualified" signal. Then, a generalized S transform is used to perform time-frequency domain analysis on the "qualified" signal and the denoised "qualified" signal, converting the one-dimensional ECG signal into a two-dimensional ECG trajectory diagram as a model input. Finally, by improving and optimizing the original GoogleNet network model, a secondary deep transfer recognition model based on GoogleNet is constructed. The GoogleNet model with a deeper network layer is selected for transfer learning recognition training based on ECG signals, which improves recognition accuracy while reducing recognition and authentication time.

Description

Technical Field

[0001] The present invention belongs to the field of information security under the construction of smart cities, and specifically relates to individual identity recognition and authentication technology in the field of information security. It proposes an electrocardiogram identity recognition method based on quality assessment and deep transfer learning. Background Art

[0002] The rapid development of information technology has promoted the rapid construction of smart cities, and also promoted the development and application of information security systems in the fields of financial security, public security and national defense. Network information security has become increasingly important in people’s daily lives, and the rapid development of intelligent life has also prompted people to put forward higher demands on current identity security authentication technologies. Common biometric identity recognition technologies such as fingerprint recognition and face recognition have the characteristics of weak anti-counterfeiting and easy prevention. Therefore, it is particularly important to seek an identity recognition technology with high anti-counterfeiting and high security. The electrocardiogram (ECG) signal, which is an endogenous signal of the human body, not only has high anti-counterfeiting and uniqueness that only meets the needs of “living body” collection, but also has one-dimensionality that can meet the needs of rapid corresponding signal processing operations. Therefore, the present invention proposes a biometric identity recognition technology with higher security and reliability based on the electrocardiogram signal.

[0003] In recent years, biometric identification algorithms based on ECG signals have developed rapidly. From feature-based recognition algorithms that require manual extraction of waveform temporal and morphological features from large, complex one-dimensional ECG signals, to non-feature-based recognition algorithms that extract wavelet transform coefficients, approximate entropy, and other methods that can reflect ECG waveform characteristics, while feature extraction methods can classify and identify different types of ECG signals, extracting numerous and complex features from ECG waveforms consumes considerable computational resources. When ECG waveforms change due to illness or emotional changes, feature extraction algorithms based on ECG waveform morphology can negatively impact individual identification results. With the rapid development of artificial intelligence (AI), the powerful self-learning capabilities of deep learning have eliminated the need for complex feature extraction methods. Furthermore, feature self-learning eliminates the need for manual extraction of ECG waveform morphological features and overcomes the impact of ECG signal variations caused by individual activity, pathological conditions, and emotional states on identification accuracy. Furthermore, while single-lead ECG signals offer a convenient and fast acquisition method, they can introduce noise interference into the original ECG signal, thus impacting the accuracy of identification.

[0004] In view of the above discussion, the present invention provides an identity recognition model integrated with ECG signal quality assessment and deep transfer learning for single-lead ECG signals with large noise interference. First, the quality level of the single-lead ECG signal is evaluated, and then deep transfer learning is performed on the ECG signals that pass the quality assessment, thereby realizing identity recognition and authentication between different individuals for single-lead ECG signals with large noise interference. Summary of the Invention

[0005] The first purpose of the present invention is to solve the above-mentioned difficulties in single-lead ECG signal identity recognition technology, and provide an ECG identity recognition method based on quality assessment and deep transfer learning, which integrates an ECG signal identity recognition and authentication model based on multi-parameter fusion quality assessment and deep transfer learning.

[0006] In the first aspect, an embodiment of the present invention provides an ECG signal quality assessment method, specifically a multi-parameter fusion quality assessment model based on SVM. First, based on the analysis of the waveform characteristics and noise content of the ECG signal, 8 signal quality indices (SQI) that can reflect the quality level of the ECG signal are extracted as evaluation criteria for the quality level of the ECG signal; then the extracted SQI features are used as the input of the SVM quality classification model to complete the classification of the three different quality levels of the single-lead ECG signal. The present invention divides the quality level of the single-lead ECG signal with greater noise interference, deletes the original ECG signal with poor quality performance in time, and effectively reduces the calculation loss of the model and the impact on the recognition accuracy.

[0007] In the second aspect, the present invention provides a signal denoising method, specifically a denoising algorithm based on a wavelet adaptive threshold transform. The wavelet adaptive threshold transform method determines a suitable threshold function by adjusting the size of the parameters in the threshold function, thereby selecting a threshold function that is harder than the traditional hard threshold function to denoise the decomposition coefficients on the detail size; and using a threshold function that is softer than the traditional soft threshold function to denoise the coefficients on the approximate size. The present invention controls the switching of the threshold function within the critical area of ​​the soft and hard thresholds, thereby achieving corresponding denoising processing of the original electrocardiogram signals at different decomposition scales.

[0008] In the third aspect, the present invention provides a method for constructing a network input layer, specifically, performing time-frequency domain transformation on the original one-dimensional time-series ECG signal based on the generalized S-transform (GST). After the generalized S-transform, the original ECG signal outputs a complex matrix containing real and imaginary parts. Each column in the matrix represents the instantaneous time-frequency domain characteristics. The real part of each complex number in the matrix is ​​taken as the horizontal coordinate and the imaginary part as the vertical coordinate to draw the ECG trajectory at each time point. The generated ECG trajectory is then captured as a movie frame by the getframe function in MATLAB, thereby realizing the conversion of the one-dimensional ECG signal into a two-dimensional frequency domain trajectory diagram as the input of the deep learning network layer.

[0009] Fourthly, the present invention provides an identity recognition and authentication method using deep transfer learning, specifically a deep transfer network model based on GoogleNet. By optimizing and improving the original GoogleNet network model, it uses it as input for ECG trajectory graphs, enabling self-learning of ECG signal features. This deep learning model overcomes the shortcomings of traditional manual extraction of signal waveform features, mitigating the impact of ECG signal differences on recognition accuracy caused by changes in individual activity status, pathological conditions, and psychological emotions.

[0010] A second object of the present invention is to provide an electrocardiogram identification system, comprising:

[0011] ECG signal quality assessment and classification module, which uses the ECG signal quality assessment and classification model based on SVM to classify the quality level of ECG signals;

[0012] The pre-processing module is used to process the "suspicious" signal obtained by the ECG signal quality assessment and classification module using the wavelet transform denoising algorithm to obtain the "qualified" signal after denoising;

[0013] An ECG trajectory diagram generation module is used to generate an ECG trajectory diagram by performing a generalized S transform on the "qualified" signal obtained by the ECG signal quality assessment and classification module and the "qualified" signal after noise removal by the preprocessing module;

[0014] The ECG signal identity recognition module uses the trained deep migration network model to perform ECG signal identity recognition on the data processed by the ECG trajectory generation module.

[0015] The embodiments of the present invention have the following beneficial effects:

[0016] 1. This paper constructs a quality assessment model based on multi-parameter fusion. By analyzing the waveform characteristics and noise content of single-lead ECG signals, it extracts eight characteristic indices that reflect the quality level of the ECG signals. These characteristic indices are then input into the model for training, enabling the quality level classification of the original ECG signals. For single-lead ECG signals with significant noise interference, this quality assessment algorithm not only promptly eliminates "poor" ECG signals that affect recognition results before signal preprocessing, but also reduces computing resource consumption and model training time.

[0017] 2. This paper constructs an identity recognition model based on deep transfer learning. Transfer learning overcomes the limitation of traditional machine learning, where training and test samples reside in the same feature space, enabling model training across different target domains. Furthermore, the GoogleNet model, a convolutional neural network with a deep network structure, is optimized for transfer learning-based identification and authentication training, improving recognition accuracy while also reducing authentication time. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0019] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] The efficient and accurate authentication of modern identity recognition technology based on ECG signals has great theoretical research and practical application value, and can provide new solution guidance for solving the security of biometric-based identity recognition at home and abroad. Figure 1 This is a system structure diagram of the ECG identity recognition method provided by an embodiment of the present invention. The present invention is further described below:

[0021] Step 1: Construct an ECG signal quality assessment classification model based on SVM to classify the quality level of ECG signals;

[0022] The electrocardiogram signal quality assessment classification model includes a quality assessment index calculation layer and an SVM classifier;

[0023] The quality assessment index calculation layer is used to analyze the waveform characteristics and noise content of the ECG signal and extract eight quality assessment indices of the ECG signal:

[0024] (1) QRS waveform clarity index SQI1

[0025] Two noise detectors with different noise sensitivities are used to detect the same QRS complex, and the number of heart beats Ne and Nw detected by the two detectors are obtained respectively. SQI1 is defined as the ratio of the number of heart beats detected by the two detectors:

[0026]

[0027] (2) Skewness SQI2

[0028] Skewness is generally used to indicate the asymmetric characteristics of data or signal distribution. When the ECG signal quality is qualified, its waveform distribution will be relatively uniform, and the corresponding skewness value will be small. Conversely, when the ECG signal contains a lot of noise and the waveform distribution is uneven, the skewness value will be large. The skewness value of the signal is defined as SQI2:

[0029]

[0030] in, represents the mean of the signal x, σ is the standard deviation of the signal x, and N represents the number of sampling points of the ECG signal.

[0031] (3)Kurtosis SQI3

[0032] Kurtosis is called the fourth-order moment of a signal and is often used to describe the degree to which data deviates from the mean of a data group and the symmetry of the signal. A larger peak indicates a greater distance between the data and the mean of the data group, and a greater probability of extreme values ​​occurring in the data group. A high-quality ECG signal has a more balanced waveform distribution and a lower probability of extreme values, resulting in a smaller kurtosis. Conversely, a signal with more noise has a more unstable waveform, a greater probability of extreme values ​​occurring, and a larger kurtosis. The formula for SQI3, defined by kurtosis, is:

[0033]

[0034] Among them, x i is the measured value of the i-th peak of signal x, represents the mean of the signal x, and σ is the standard deviation of the signal x.

[0035] (4) Power spectrum distribution index SQI4

[0036] The energy of the ECG signal waveform is mainly concentrated near the QRS waveform at around 10 Hz. The ratio of the power spectral density of a small frequency band (5-15 Hz) containing the energy center frequency to the power spectral density of the overall ECG signal concentrated frequency band (5-40 Hz) is defined as SQI4:

[0037]

[0038] Where f represents the frequency and p(f) represents the power of the signal at that frequency.

[0039] (5) Baseline shift index SQI5

[0040] The baseline causes the ECG signal waveform to rise and fall. Even a small amount of interference from the signal rise and fall can cause the ECG signal to vary greatly at low amplitudes, thus affecting the overall signal quality. The ratio of the power spectral density of the baseline frequency distribution between 0-1Hz and the power spectral density of the entire frequency band is defined by the SQI5 of the baseline offset:

[0041]

[0042] Where f represents the frequency and p(f) represents the power of the signal at that frequency.

[0043] (6) QRS wave energy index SQI6

[0044] The QRS waveform has the highest energy content in the ECG signal. By calculating the ratio of QRS waveform energy to the total ECG waveform energy, the QRS waveform energy index is used as an indicator to measure the ECG quality level, which is defined as index SQI6:

[0045]

[0046] Among them, Er i It represents the energy of a QRS complex, i represents the detected QRS complex, and Ea represents the total energy of the entire ECG segment.

[0047] (7) RR interval stability index SQI7

[0048] For electrocardiographic signals, the RR interval shows periodic changes in the ECG signal waveform. However, when the signal contains a large amount of noise, the RR interval is easily interfered by noise such as myoelectricity. Therefore, the RR interval stability index is defined to measure the stability of the ECG signal waveform within a cycle. The definition of SQI7 is:

[0049]

[0050] in, and They represent the standard deviation and mean of the ECG signal within the RR interval, respectively. The worse the signal quality, the more unstable the RR interval and the larger the SQI7 value. Conversely, the better the signal quality.

[0051] (8) Purity SQI8

[0052] The purity of the ECG signal can be used to reflect the noise content of the signal. The lower the signal purity value, the greater the signal noise and the worse the signal quality. On the contrary, the larger the signal purity value, the smaller the signal noise and the better the signal quality. SQI8 is defined as:

[0053]

[0054] Among them, w n represents the nth-order spectral moment, p(e jw ) represents the power spectrum of the entire ECG segment.

[0055] The SVM classifier is used to classify the eight quality assessment indices output by the quality assessment index calculation layer into signal quality levels according to thresholds; the signal quality level classification includes three categories: "qualified" signal, "suspicious" signal, and "poor" signal; and then eliminate the "poor" signal;

[0056] Step 2: Process the "suspicious" signal based on the wavelet transform denoising algorithm to obtain the "qualified" signal after denoising; specifically:

[0057] The "suspicious" signal is decomposed by wavelet transform by selecting appropriate wavelet basis function and decomposition scale; then threshold denoising is performed on each decomposition layer; finally, the denoised signal of each decomposition layer is reconstructed to obtain a clean ECG signal.

[0058] Step 3: Perform a generalized S-transform on the “qualified” signal from step 1 and the “qualified” signal after noise elimination from step 2 to generate an ECG trajectory graph; specifically:

[0059] While preserving the signal's phase information, the generalized S-transform introduces a Gaussian window function whose width varies with frequency, overcoming the traditional Fourier transform's problem of constant resolution in the time and frequency domains. This paper proposes a time-frequency domain analysis algorithm based on the generalized S-transform, converting one-dimensional ECG signals into two-dimensional frequency domain traces as input for deep transfer learning network models.

[0060] ① Time-frequency domain transformation. Two parameters λ and ρ are introduced to adjust the window width and amplitude of the window function g(t), thereby changing the time resolution and frequency resolution of the signal and realizing multi-resolution analysis. The generalized S transform of the signal x(t) is expressed as follows:

[0061]

[0062] Where τ is the time shift factor. The window function and wavelet basis function of the generalized S transform are as follows:

[0063]

[0064]

[0065] Setting parameters A suitable window function is used to ensure optimal accuracy. The signal x(t) is processed by the GST of equation (9) above to obtain a complex matrix S, which contains real and imaginary parts, as shown in equation (12). Here, N represents the frequency magnitude and M represents the time point distribution.

[0066]

[0067] ② Take the real part of each data point in the complex matrix S as the horizontal coordinate and the imaginary part as the vertical coordinate, and plot the ECG trajectory at each time point. The generated ECG trajectory is then captured as a movie frame using the getframe function in MATLAB. This yields a frequency domain trajectory of the ECG signal at that time point, which is used as input to the deep learning network layer.

[0068] Step 4: Perform a secondary deep migration on the improved GoogleNet network to build a deep migration network model, and use it to realize the identity recognition of ECG signals; specifically:

[0069] The deep migration network model adopts the GoogleNet network as the basic framework and performs secondary improvements and optimizations on the original GoogleNet network model to construct an identity recognition model based on electrocardiogram signals.

[0070] The GoogleNet network includes a GoogLeNet model with a total of 22 convolutional layers, of which 9 Inception modules are used to improve the depth and width of the network structure. Specifically: the first convolution layer uses a 7×7 convolution kernel, and then passes through a 3×3 maximum pooling layer and a 1×1 convolution layer for dimensionality reduction output; the second convolution layer uses a 3×3 convolution kernel, and then passes through a 3×3 convolution maximum pooling layer for output; the third layer is the inception (3a-3b) layer, and each inception layer is divided into 2 convolution layers and 4 branches. Different branches use convolution kernels of different scales for processing, which are 64 1×1 convolution kernels, 96 1×1 convolution kernels as input dimensionality reduction of 3×3 convolution kernels, 16 1×1 convolution kernels as input dimensionality reduction of 5×5 convolution kernels, and finally the 3×3 convolution kernel pooling layer output and pass through 32 1×1 convolution kernels for dimensionality reduction. These four output results are connected in parallel to form an inception. The output result of the ion module; the fourth layer is inception (4a-4e), and the output of inception (3a-3b) is used as the input of the fourth layer. The implementation process of the inception module is consistent with the above operation; the fifth layer is inception (5a-5b), and the output of inception (4a-4e) is also used as the input of the two inception modules connected in series in the fifth layer to complete the convolution operation; the sixth layer is the model output layer, which consists of a 7×7 average convolution layer, a fully connected layer and a Softmax classifier. The output of inception (5a-5b) is used as the input of the 7×7 average convolution layer for data dimensionality reduction, and finally a full connection is performed to complete feature classification.

[0071] During the model training process, the original GoogleNet network model is first optimized to complete the first transfer learning. Specifically, a 5×5 large convolution layer in the inception module of the original GoogleNet network model is changed to two series-connected 3×3 small convolution layers. The optimized network model increases the depth of the original network model while reducing the amount of calculation of the model parameters, thereby achieving a highly robust and accurate recognition effect. The frequency domain trajectory of the normal heart rhythm ECG signal processed by steps 1-3 is used as the input of the first transfer learning model for training, and the model trained by the first transfer learning is obtained and named GoogleNet-T1; then, the parameters of some network layers of the GoogleNet-T1 model after the first transfer learning are updated using the freeze plus fine-tuning technology to construct a second transfer learning model and named GoogleNet-T2. The frequency spectrum trajectory of the abnormal heart rhythm ECG signal processed by steps 1-3 is used as the input of the second transfer learning model for abnormal heart rhythm ECG signal recognition training, and finally a trained deep transfer network model is obtained;

[0072] The freezing network layer is a technology that fixes the first three layers of the first transfer learning model GoogleNet-T1, namely the 7×7 convolution kernel and 3×3 pooling layer of the first layer, the 3×3 convolution kernel and 3×3 pooling layer of the second layer, and the inception (3a-3b) module of the third layer. The parameters of the frozen network layer will not be updated during the model training process. It is mainly used to extract the same features of the electrocardiogram signal, thereby reducing the model recognition training time.

[0073] The fine-tuning layer updates the parameters of the fourth and fifth layers of the first transfer learning model, GoogleNet-T1, primarily the convolutional layers of the seven Inception modules (Inception (4a-4e) and Inception (5a-5b). This layer is used to extract detailed features of ECG signals. The fine-tuning layer design enables detailed feature extraction for ECG signals with abnormal heart rhythms, thereby enhancing the generalization and robustness of the recognition model.

Claims

1. An ECG identity recognition method based on quality assessment and deep transfer learning, characterized by The following steps are involved: Step 1: Construct an ECG signal quality assessment classification model based on SVM to classify the quality level of ECG signals; The signal quality level classification includes three categories: "qualified" signal, "questionable" signal, and "poor" signal; Step 2: Process the "suspicious" signal based on the wavelet transform denoising algorithm to obtain the "qualified" signal after denoising; specifically: Select appropriate wavelet basis functions and decomposition scales to perform wavelet transform decomposition on the "suspicious" signal; then perform threshold denoising on each decomposition layer; finally, reconstruct the denoised signal of each decomposition layer to obtain a clean ECG signal; Step 3: Perform generalized S-transformation on the qualified signal in step 1 and the qualified signal after noise elimination in step 2 to generate an ECG trajectory diagram; Step 4: Use the improved GoogleNet network for secondary transfer learning to build a deep transfer recognition network model, and use it to realize individual identity recognition based on ECG signals; Specifically: The deep migration network model adopts the GoogleNet network as the basic framework and performs secondary improvements and optimizations on the original GoogleNet network model to construct an identity recognition model based on ECG signals; The GoogleNet network has a total of 22 convolutional layers, of which 9 inception modules are used to improve the depth and width of the network structure. Specifically, the first convolution layer uses a 7×7 convolution kernel, and then passes through a 3×3 maximum pooling layer and a 1×1 convolution layer for dimensionality reduction output; the second convolution layer uses a 3×3 convolution kernel, and then passes through a 3×3 maximum pooling layer for output; the third layer consists of two inception modules 3a-3b; the fourth layer consists of five inception modules 4a-4e, and the outputs of the two inception modules 3a-3b in the third layer are used as the input of the fourth layer; the fifth layer consists of two inception modules 5a-5b, and the outputs of the five inception modules 4a-4e in the fourth layer are used as the input of the fifth layer to complete the convolution operation; the sixth layer is the model output layer; During the model training process, the original GoogleNet network model is first optimized to complete the first transfer learning. Specifically, a 5×5 large convolution layer in the inception module of the original GoogleNet network model is changed to two serially connected 3×3 small convolution layers. The optimized network model increases the depth of the original network model while reducing the calculation amount of the model parameters. The frequency domain trajectory map of the normal heart rhythm ECG signal after steps 1-3 is used as the input of the first transfer learning model for training, and the model GoogleNet-T1 after the first transfer learning training is obtained. Then, the parameters of some network layers of the GoogleNet-T1 model after the first transfer learning are updated using the freeze and fine-tuning technology to construct the second transfer learning model GoogleNet-T2. The frequency spectrum trajectory map of the abnormal heart rhythm ECG signal after steps 1-3 is used as the input of the second transfer learning model for abnormal heart rhythm ECG signal recognition training, and finally a trained deep transfer network model is obtained. The freezing method specifically involves freezing the first three layers of the first transfer learning model GoogleNet-T1, namely the 7×7 convolution kernel and 3×3 pooling layer of the first layer, the 3×3 convolution kernel and 3×3 pooling layer of the second layer, and the Inception modules 3a-3b of the third layer. The frozen network layers will not have their parameters updated during the model training process and are mainly used to extract the same features of the ECG signal, thereby reducing the model recognition training time. The fine-tuning specifically involves updating the parameters of the fourth and fifth layers in the first transfer learning model GoogleNet-T1, mainly including the convolutional layers of the inception modules 4a-4e and the inception modules 5a-5b, to extract detailed features of the electrocardiogram signal.

2. The ECG identity recognition method based on quality assessment and deep transfer learning as described in claim 1 is characterized in that The ECG signal quality assessment classification model in step 1 includes a quality assessment index calculation layer and an SVM classifier; The quality assessment index calculation layer extracts eight quality assessment indices of the ECG signal by analyzing the waveform characteristics and noise content of the ECG signal: (1) QRS waveform clarity index SQI1 Two noise detectors with different noise sensitivities are used to detect the same QRS complex, and the number of heart beats Ne and Nw detected by the two detectors are obtained respectively. SQI1 is defined as the ratio of the number of heart beats detected by the two detectors: (2) Skewness SQI2 in, represents the mean of signal x, σ is the standard deviation of signal x, and N represents the number of sampling points of the ECG signal; (3)Kurtosis SQI3 Among them, x i is the measured value of the i-th peak of signal x, represents the mean of signal x, and σ is the standard deviation of signal x; (4) Power spectrum distribution index SQI4 Where f represents the frequency and p(f) represents the power of the signal at that frequency; (5) Baseline shift index SQI5 Where f represents the frequency and p(f) represents the power of the signal at that frequency; (6) QRS wave energy index SQI6 Among them, Er i Represents the energy of the QRS complex, i represents the detected QRS complex, and Ea represents the total energy of the entire ECG segment; (7) RR interval stability index SQI7 in, and Represents the standard deviation and mean of the ECG signal in the RR interval; (8) Purity SQI8 Among them, w n represents the nth-order spectral moment, p(e jw ) represents the power spectrum of the entire ECG segment; The SVM classifier completes the quality level classification of the original electrocardiogram signal by training the 8 electrocardiogram signal quality evaluation index features output by the calculation layer.

3. The ECG identity recognition method based on quality assessment and deep transfer learning as described in claim 1 is characterized in that Step 3: Based on the time-frequency domain analysis algorithm of generalized S-transform, the one-dimensional ECG signal is converted into a two-dimensional ECG signal frequency domain trajectory image as the input of the deep transfer learning network model. Specifically: ① Time-frequency domain transformation: Two parameters λ and ρ are introduced to adjust the window width and amplitude of the window function g(t), thereby changing the time resolution and frequency resolution of the signal and realizing multi-resolution analysis. The generalized S transform of the signal x(t) is expressed as follows: Among them, τ is the time shift factor; the window function and wavelet basis function of the generalized S transform are as follows: Setting parameters The appropriate window function is used to ensure the best accuracy. The signal x(t) is processed by the GST of formula (9) above to obtain a complex matrix S, which contains real and imaginary parts, as shown in formula (12). Where N represents the frequency and M represents the time point. ② Take the real part of each data in the complex matrix S as the horizontal coordinate and the imaginary part as the vertical coordinate to draw the ECG trajectory at each time point; the generated ECG trajectory is then captured as a movie frame through the getframe function in MATLAB to obtain the frequency domain trajectory of the ECG signal at a time point, which is used as the input of the deep learning network layer.

4. The ECG identity recognition method based on quality assessment and deep transfer learning as claimed in claim 1, characterized in that The inception module in the deep migration network model includes 2 convolutional layers and 4 branches. Different branches use convolution kernels of different scales for processing, which are 64 1×1 convolution kernels, 96 1×1 convolution kernels as the input dimensionality reduction of 3×3 convolution kernels, 16 1×1 convolution kernels as the input dimensionality reduction of 5×5 convolution kernels, and finally the pooling layer output of the 3×3 convolution kernel and dimensionality reduction through 32 1×1 convolution kernels. These four output results are processed in parallel as the output result.

5. The ECG identity recognition method based on quality assessment and deep transfer learning according to claim 1 or 4, characterized in that The output layer of the deep migration network model consists of a 7×7 average convolution layer, a fully connected layer, and a Softmax classifier. The output of the inception module 5a-5b is used as the input of the 7×7 average convolution layer for data dimensionality reduction, and finally a full connection is performed to complete feature classification.

6. An electrocardiogram identification system implementing the method according to any one of claims 1 to 5, characterized in that include: ECG signal quality assessment and classification module, which uses the ECG signal quality assessment and classification model based on SVM to classify the quality level of ECG signals; The preprocessing module is used to process the "suspicious" signal obtained by the ECG signal quality assessment and classification module using the wavelet transform denoising algorithm to obtain the "qualified" signal after denoising; An ECG trajectory diagram generation module is used to generate an ECG trajectory diagram by performing a generalized S-transform on the "qualified" signal obtained by the ECG signal quality assessment and classification module and the "qualified" signal after noise removal by the preprocessing module; The ECG signal identity recognition module uses the trained deep migration network model to perform ECG signal identity recognition on the data processed by the ECG trajectory generation module.

Images