Self-supervised ventricular arrhythmia anomaly detection method and system based on double-domain knowledge distillation
By employing a self-supervised method of dual-domain knowledge distillation, a time-frequency domain dual-branch model is constructed. Self-supervised training is performed using pseudo-noise and pseudo-anomaly data to optimize the student model. This solves the problem of noise interference in ventricular arrhythmia detection, improves detection accuracy and noise discrimination ability, and reduces the false alarm rate.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHENGZHOU UNIV
- Filing Date
- 2026-02-04
- Publication Date
- 2026-05-19
AI Technical Summary
Existing machine learning-based methods for detecting ventricular arrhythmias struggle to effectively distinguish between ventricular arrhythmias and noise features in complex electrocardiogram signals. They rely on data augmentation or a large number of negative sample pairs, resulting in insufficient detection accuracy.
A self-supervised method based on dual-domain knowledge distillation is adopted to construct a time-frequency domain dual-branch processing model. Self-supervised training is performed using pseudo-noise and pseudo-anomaly data. The student model is optimized through a hypersphere structure to reduce noise interference and improve detection accuracy.
It effectively improves the accuracy of ventricular arrhythmia detection and noise discrimination ability, reduces the false alarm rate, and can assist in the diagnosis of cardiac abnormalities, especially performing excellently on datasets with severe noise interference.
Smart Images
Figure CN122065211A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of time-series data processing technology, and in particular to a self-supervised method and system for detecting ventricular arrhythmias based on dual-domain knowledge distillation. Background Technology
[0002] Heart disease, electrolyte imbalances, and genetic factors can cause "short circuits" in cardiac electrical activity, resulting in ventricular arrhythmias, which can lead to palpitations in severe cases. Current diagnostic methods primarily rely on electrocardiograms (ECGs), Holter monitoring, and echocardiography. Current techniques utilize end-to-end learning to identify abnormal rhythms such as premature ventricular contractions (PVCs) and ventricular tachycardia (VT) from raw ECG signals, overcoming the limitations of traditional manual feature extraction and directly mining deep discriminative patterns from time-series signals. However, ECG data often contains noise interference, and ventricular arrhythmia data can resemble noisy data. Machine learning-based ventricular arrhythmia detection methods typically rely on data augmentation or a large number of negative sample pairs, making it difficult to effectively distinguish ventricular arrhythmias from noise features in complex ECG signals. Summary of the Invention
[0003] To address the problems of existing ventricular arrhythmia detection methods that rely on samples and struggle to effectively distinguish complex ECG signals, this invention provides a self-supervised ventricular arrhythmia anomaly detection method and system based on dual-domain knowledge distillation. By utilizing self-supervision to process noise, the impact of noise on model discrimination is reduced, thereby improving the accuracy of ventricular arrhythmia detection.
[0004] According to the design scheme provided by this invention, on one hand, a self-supervised method for detecting ventricular arrhythmias based on dual-domain knowledge distillation is provided, comprising:
[0005] A three-classification task sample dataset is constructed based on normal ECG data, pseudo-noise data, and pseudo-abnormal ECG data. The pseudo-noise data is generated by simulating a combination of noise from normal ECG signals, and the pseudo-abnormal ECG data is generated based on normal ECG data using frequency modulation technology.
[0006] A time-frequency domain dual-branch processing model architecture is constructed, with each processing branch consisting of an encoder and a classification head. The time-frequency domain dual-branch processing architecture is used as the student model, and a teacher model with the same structure is constructed based on the time-frequency domain dual-branch processing architecture. The sample signals in the three-class task sample dataset are input into the teacher model, and the teacher model is pre-trained with the goal of minimizing the difference between the classification prediction result of the teacher model and the true class of the sample signal.
[0007] With fixed parameters of the trained teacher model, the sample signals are input into the teacher model and the student model respectively. The time-frequency domain features output by the encoder of the student model are mapped onto the hypersphere structure. The student model is trained with the goal of minimizing the difference between the classification prediction results of the teacher model and the student model and the radius of the hypersphere structure under the normal ECG category.
[0008] The trained student model is used as a time-frequency domain anomaly detection model. The heart rhythm signal to be processed is input into the time-frequency domain anomaly detection model, and the time-frequency domain anomaly detection model is used to detect and identify anomalies in the heart rhythm signal to be processed.
[0009] As part of the self-supervised ventricular arrhythmia detection method based on dual-domain knowledge distillation of this invention, the method further includes generating pseudo-noise data based on a combination of normal electrocardiogram signal and noise, comprising:
[0010] Randomly set the window length for each lead of a normal electrocardiogram signal;
[0011] One or more noises are randomly selected from the ECG noise set and combined to generate pseudo-noise data that simulates the real noise of the ECG signal in the simulated window.
[0012] As a self-supervised ventricular arrhythmia abnormality detection method based on dual-domain knowledge distillation of the present invention, the pseudo-noise data simulation generation function is further expressed as: ,in, This represents a function that represents a random combination of noise types. For probability, This represents a function for random window selection. Normal ECG signal The characteristics of a single lead. Normal ECG signal The corresponding generated pseudo-noise data samples.
[0013] As a self-supervised ventricular arrhythmia detection method based on dual-domain knowledge distillation of the present invention, the process of generating pseudo-abnormal ECG data based on normal ECG data and using frequency modulation technology is further represented as follows: ,in, For time, This is the raw signal of normal electrocardiogram data. The fundamental frequency of the signal. The modulation index determines the degree of frequency variation. These are false abnormal ECG data.
[0014] As a self-supervised ventricular arrhythmia abnormality detection method based on dual-domain knowledge distillation of the present invention, further, during the training of the student model, the time-domain and frequency-domain features of normal electrocardiogram data are mapped onto the hyperspherical structure to construct a dual-sphere hyperspherical structure under the normal electrocardiogram category.
[0015] As a self-supervised ventricular arrhythmia detection method based on dual-domain knowledge distillation of this invention, further, in the student model training, the calculation process for minimizing the radius of the hyperspherical structure under the normal ECG category includes:
[0016] The distance between the feature vector and the center of the hypersphere in the time-frequency domain branch is obtained by using cosine similarity, where the initial value of the hypersphere center is determined by the mean of all feature vectors;
[0017] The hypersphere loss is obtained based on the distance and the proportion of the time-frequency branch loss, and the hypersphere radius during model training is updated based on the hypersphere loss.
[0018] As a self-supervised ventricular arrhythmia detection method based on dual-domain knowledge distillation of the present invention, the hypersphere loss calculation process is further expressed as follows: ,in, , These represent the time-domain feature vector and the frequency-domain feature vector of the electrocardiogram (ECG) signal, respectively. Represents the center of the time sphere. Represents the center of the sphere in the frequency domain. This indicates the proportion of time-frequency domain branching loss.
[0019] Furthermore, this invention also provides a self-supervised ventricular arrhythmia anomaly detection system based on dual-domain knowledge distillation, comprising: a sample construction module, a model pre-training module, a model training module, and an anomaly detection module, wherein...
[0020] The sample construction module is used to construct a three-classification task sample dataset based on normal ECG data, pseudo-noise data, and pseudo-abnormal ECG data. The pseudo-noise data is generated by simulating a combination of noise from normal ECG signals, and the pseudo-abnormal ECG data is generated based on normal ECG data and using frequency modulation technology.
[0021] The model pre-training module is used to construct a time-frequency domain dual-branch processing model architecture. Each processing branch consists of an encoder and a classification head. The time-frequency domain dual-branch processing architecture is used as the student model, and a teacher model with the same structure is constructed based on the time-frequency domain dual-branch processing architecture. The sample signals in the three-class task sample dataset are input into the teacher model, and the teacher model is pre-trained with the goal of minimizing the difference between the teacher model's classification prediction result and the true class of the sample signal.
[0022] The model training module is used to fix the parameters of the trained teacher model, input the sample signals into the teacher model and the student model respectively, and map the time-frequency domain features output by the encoder of the student model onto the hypersphere structure. The student model is trained with the goal of minimizing the difference between the classification prediction results of the teacher model and the student model and the radius of the hypersphere structure under the normal ECG category.
[0023] The anomaly detection module is used to take the trained student model as the time-frequency domain anomaly detection model, input the heart rhythm signal to be processed into the time-frequency domain anomaly detection model, and use the time-frequency domain anomaly detection model to detect and identify anomalies in the heart rhythm signal to be processed.
[0024] The beneficial effects of this invention are:
[0025] This invention constructs two types of pseudo-abnormal data. Using normal data and these two types of pseudo-abnormal data, a classification task is built in the teacher model. An offline knowledge distillation method is used to enhance the encoder's ability to distinguish between noisy data and ventricular arrhythmia data during unsupervised training. Experiments on three real-world ECG datasets—VFDB, MIT-BIH, and CUDB—show that the proposed solution can effectively improve the accuracy of ventricular arrhythmia detection and the model's noise discrimination ability, effectively reduce the false alarm rate of ventricular arrhythmias, and effectively assist in the diagnosis and treatment of cardiac abnormalities in patients. Attached Figure Description
[0026] Figure 1 This is a schematic diagram of the self-supervised ventricular arrhythmia abnormality detection process based on dual-domain knowledge distillation in the embodiment.
[0027] Figure 2 This is a schematic diagram of the model training process in the embodiment;
[0028] Figure 3 This is a schematic diagram showing the comparison of experimental results in the examples. Detailed Implementation
[0029] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the invention will be further described in detail below with reference to the accompanying drawings and technical solutions.
[0030] To address the issue of ventricular arrhythmias, this invention classifies ventricular arrhythmias as an anomaly detection problem and uses pseudo-data to construct a self-supervised task. (See [link to relevant documentation]). Figure 1 As shown, a self-supervised method for detecting ventricular arrhythmias based on dual-domain knowledge distillation is provided, comprising:
[0031] S101. Construct a three-classification task sample dataset based on normal ECG data, pseudo-noise data, and pseudo-abnormal ECG data. The pseudo-noise data is generated by simulating a combination of normal ECG signal and noise, and the pseudo-abnormal ECG data is generated based on normal ECG data and using frequency modulation technology.
[0032] Specifically, the window length of each lead of a normal electrocardiogram (ECG) signal can be randomly set; one or more types of noise can be randomly selected from the ECG noise set and combined to generate pseudo-noise data that simulates the real noise of the ECG signal within the simulated window.
[0033] In this embodiment, the pseudo-random noise generation uses various noise combinations based on normal ECGs to simulate real ECG noise. These noises include Gaussian noise, Rayleigh noise, gamma noise, exponential noise, Poisson noise, and uniform noise. To simulate real ECG noise, a window length is first randomly selected for each lead of the ECG signal, and then the above noise types are randomly combined. Through this random combination, the complexity and diversity of real noise can be approximated as closely as possible. Specifically, the method defines the input... The function for generating simulated real noise data is expressed as:
[0034]
[0035] in This represents a random combination of the above noise types, with a probability of... , Indicates random window selection. for The characteristics of a single lead. This is a pseudo-sample.
[0036] Frequency modulation (FM) technology was used to generate pseudo-ventricular arrhythmia ECG data, mathematically expressed as follows:
[0037]
[0038] in For time, The original signal, The fundamental frequency of the signal. The modulation index determines the degree of frequency variation. Data on pseudoventricular arrhythmias.
[0039] S102. Construct a time-frequency domain dual-branch processing model architecture. Each processing branch consists of an encoder and a classification head. Use the time-frequency domain dual-branch processing architecture as the student model, and construct a teacher model with the same structure based on the time-frequency domain dual-branch processing architecture. Use the sample signals in the three-class task sample dataset as input to the teacher model, and use minimizing the difference between the teacher model's classification prediction result and the true category of the sample signal as the objective to pre-train the teacher model.
[0040] like Figure 2 As shown, a dual-branch structure is adopted, performing feature extraction separately in the time and frequency domains, and constructing a double-centered hypersphere structure. The time-frequency domain analysis model mainly includes a time-domain encoder. and frequency domain encoder First, wavelet transform is used to extract the frequency domain signal from the original electrocardiogram (ECG) signal. Then, the normal time-domain ECG signal is input into a time-domain encoder to extract and embed the normal time-domain features. Simultaneously, the normal frequency domain ECG signal is input into the frequency domain encoder to extract and embed the normal frequency domain features. Finally, the two sets of features are mapped into the hypersphere for processing.
[0041] S103. Fix the parameters of the trained teacher model, input the sample signals into the teacher model and the student model respectively, and map the time-frequency domain features output by the encoder of the student model onto the hypersphere structure. The student model is trained with the goal of minimizing the difference between the classification prediction results of the teacher model and the student model and the radius of the hypersphere structure under the normal ECG category.
[0042] During the training of the student model, the time-domain and frequency-domain features of normal ECG data are mapped onto a hyperspherical structure to construct a double-center hyperspherical structure for the normal ECG category. Cosine similarity can be used to obtain the distance between the feature vector and the center of the hypersphere in the time-frequency domain branch. The initial value of the hypersphere center is determined by the mean of all feature vectors. The hypersphere loss is obtained based on the distance and the weight of the time-frequency domain branch loss, and the hypersphere radius is updated during model training based on the hypersphere loss.
[0043] Pseudo-anomaly data also follows the same feature extraction procedure as normal data. The difference is that pseudo-anomaly data does not participate in the hypersphere construction process; instead, it is extracted from pseudo-temporal features. pseudo-frequency domain features Subsequently, these two sets of features will participate in the self-supervised task, simultaneously improving the encoder's ability to distinguish noise and anomalous signals during the training process.
[0044] The model training aims to learn a hypersphere with the smallest possible radius, which can contain all normal training data in the feature space. The model encloses the data within the hypersphere by minimizing its radius while bringing as many data points as possible close to its center. The distance from the hypersphere's center serves as a measure of anomalousness; data points farther from the center are more likely to be considered anomalous. First, normal and pseudo-anomalous data are simultaneously input into the time-domain and frequency-domain encoders for feature extraction. After feature extraction at each time-frequency branch, the model calculates the feature vector for each data point and its relationship to the branch's hypersphere center. The distance between the centers of each hypersphere. The initial value is determined by calculating the mean of the feature vectors of all training data extracted by the model, that is:
[0045]
[0046] Where N is the number of training data. Cosine similarity is used as the distance metric, and the distance formula is expressed as:
[0047]
[0048] in and Representing vectors respectively and No. One value, This represents the vector length. After calculating the distance... Then, the hypersphere loss function can be defined as:
[0049]
[0050] in Represents the center of the time sphere. Represents the center of the sphere in the frequency domain. This indicates the proportion of time-frequency domain branching loss.
[0051] ResNet-18 can be used as the teacher model for pre-training. In the pre-training stages of the time-frequency domain branch, a three-class classification task is constructed using two types of pseudo-data and normal data. The final teacher model is then obtained. Then, for each branch, the encoder is treated as the student model and the teacher model parameters are frozen. Simultaneously with anomaly detection, two types of pseudo-anomaly data are input into the teacher model for classification, resulting in... However, the classification results are not included in the teacher model update. For each student and the three data features output by the teacher model, they will be input into a system with temperature... Classification Head In the process, the prediction results are obtained. Finally, the loss function for self-supervised training of a single branch adopts the mean squared error loss function, which is expressed as:
[0052] in These represent the classification probability distributions of the output features of the model encoder and the teacher model, respectively.
[0053] S104. The trained student model is used as the time-frequency domain anomaly detection model. The heart rhythm signal to be processed is input into the time-frequency domain anomaly detection model, and the time-frequency domain anomaly detection model is used to detect and identify anomalies in the heart rhythm signal to be processed.
[0054] Furthermore, based on the above method, this embodiment of the invention also provides a self-supervised ventricular arrhythmia anomaly detection system based on dual-domain knowledge distillation, comprising: a sample construction module, a model pre-training module, a model training module, and an anomaly detection module, wherein:
[0055] The sample construction module is used to construct a three-classification task sample dataset based on normal ECG data, pseudo-noise data, and pseudo-abnormal ECG data. The pseudo-noise data is generated by simulating a combination of noise from normal ECG signals, and the pseudo-abnormal ECG data is generated based on normal ECG data and using frequency modulation technology.
[0056] The model pre-training module is used to construct a time-frequency domain dual-branch processing model architecture. Each processing branch consists of an encoder and a classification head. The time-frequency domain dual-branch processing architecture is used as the student model, and a teacher model with the same structure is constructed based on the time-frequency domain dual-branch processing architecture. The sample signals in the three-class task sample dataset are input into the teacher model, and the teacher model is pre-trained with the goal of minimizing the difference between the teacher model's classification prediction result and the true class of the sample signal.
[0057] The model training module is used to fix the parameters of the trained teacher model, input the sample signals into the teacher model and the student model respectively, and map the time-frequency domain features output by the encoder of the student model onto the hypersphere structure. The student model is trained with the goal of minimizing the difference between the classification prediction results of the teacher model and the student model and the radius of the hypersphere structure under the normal ECG category.
[0058] The anomaly detection module is used to take the trained student model as the time-frequency domain anomaly detection model, input the heart rhythm signal to be processed into the time-frequency domain anomaly detection model, and use the time-frequency domain anomaly detection model to detect and identify anomalies in the heart rhythm signal to be processed.
[0059] To verify the effectiveness of this solution, the following explanation is based on experimental data:
[0060] All methods were evaluated using six random seeds and the average experimental results were taken. Figure 3 The experimental results of this proposed solution are compared with existing anomaly detection methods. On the VFDB dataset, compared with other anomaly detection methods with fixed network parameters, the proposed method achieves the highest AUC and AP scores, exceeding the best comparison methods by 15.73% and 25.28%, respectively. On the MITBIH dataset, the scores are 6.36% and 14.5% higher, respectively. On the CUDB dataset, the scores are 13.27% and 18.48% higher, respectively.
[0061] In summary, our proposed solution significantly outperforms existing methods on all three datasets, particularly achieving the greatest performance advantage on the VFDB dataset, which has the most abundant noisy data. This demonstrates that our solution can effectively perform high-quality learning on noisy data. Furthermore, the superior performance on all three datasets indicates that our proposed dual-branch time-frequency domain analysis structure can effectively extract time-frequency domain features from ECG data and effectively combine self-supervised tasks to optimize the encoder, thereby learning more discriminative embedded features. In addition, noise recognition accuracy tests were conducted, and our method achieved a noise recognition accuracy 14.36% higher than the baseline, indicating superior noise discrimination capabilities.
[0062] Unless otherwise specifically stated, the relative steps, numerical expressions, and values of the components and steps described in these embodiments do not limit the scope of the invention.
[0063] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0064] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.
[0065] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.
[0066] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A self-supervised method for detecting ventricular arrhythmias based on dual-domain knowledge distillation, characterized in that, Include: A three-classification task sample dataset is constructed based on normal ECG data, pseudo-noise data, and pseudo-abnormal ECG data. The pseudo-noise data is generated by simulating a combination of noise from normal ECG signals, and the pseudo-abnormal ECG data is generated based on normal ECG data using frequency modulation technology. A time-frequency domain dual-branch processing model architecture is constructed, with each processing branch consisting of an encoder and a classification head. The time-frequency domain dual-branch processing architecture is used as the student model, and a teacher model with the same structure is constructed based on the time-frequency domain dual-branch processing architecture. The sample signals in the three-class task sample dataset are input into the teacher model, and the teacher model is pre-trained with the goal of minimizing the difference between the classification prediction result of the teacher model and the true class of the sample signal. With fixed parameters of the trained teacher model, the sample signals are input into the teacher model and the student model respectively. The time-frequency domain features output by the encoder of the student model are mapped onto the hypersphere structure. The student model is trained with the goal of minimizing the difference between the classification prediction results of the teacher model and the student model and the radius of the hypersphere structure under the normal ECG category. The trained student model is used as a time-frequency domain anomaly detection model. The heart rhythm signal to be processed is input into the time-frequency domain anomaly detection model, and the time-frequency domain anomaly detection model is used to detect and identify anomalies in the heart rhythm signal to be processed.
2. The self-supervised ventricular arrhythmia detection method based on dual-domain knowledge distillation according to claim 1, characterized in that, Pseudo-noise data is generated based on the simulation of normal ECG signal noise combination, including: Randomly set the window length for each lead of a normal electrocardiogram signal; One or more noises are randomly selected from the ECG noise set and combined to generate pseudo-noise data that simulates the real noise of the ECG signal in the simulated window.
3. The self-supervised ventricular arrhythmia detection method based on dual-domain knowledge distillation according to claim 1 or 2, characterized in that, The pseudo-noise data simulation generation function is expressed as: ,in, This represents a function that represents a random combination of noise types. For probability, This represents a function for random window selection. Normal ECG signal The characteristics of a single lead. Normal ECG signal The corresponding generated pseudo-noise data samples.
4. The self-supervised ventricular arrhythmia detection method based on dual-domain knowledge distillation according to claim 1, characterized in that, The process of generating pseudo-abnormal ECG data based on normal ECG data and using frequency modulation technology is represented as follows: ,in, For time, This is the raw signal of normal electrocardiogram data. The fundamental frequency of the signal. The modulation index determines the degree of frequency variation. These are false abnormal ECG data.
5. The self-supervised ventricular arrhythmia detection method based on dual-domain knowledge distillation according to claim 1, characterized in that, During the training of the student model, the time-domain and frequency-domain features of normal electrocardiogram (ECG) data are mapped onto the hyperspherical structure to construct a double-sphere hyperspherical structure for the normal ECG category.
6. The self-supervised ventricular arrhythmia detection method based on dual-domain knowledge distillation according to claim 1 or 5, characterized in that, In student model training, the calculation process for minimizing the radius of the hyperspherical structure under the normal ECG category includes: The distance between the feature vector and the center of the hypersphere in the time-frequency domain branch is obtained by using cosine similarity, where the initial value of the hypersphere center is determined by the mean of all feature vectors; The hypersphere loss is obtained based on the distance and the proportion of the time-frequency branch loss, and the hypersphere radius during model training is updated based on the hypersphere loss.
7. The self-supervised ventricular arrhythmia detection method based on dual-domain knowledge distillation according to claim 6, characterized in that, The calculation process for hypersphere loss is expressed as follows: ,in, , These represent the time-domain feature vector and the frequency-domain feature vector of the electrocardiogram (ECG) signal, respectively. Represents the center of the time sphere. Represents the center of the sphere in the frequency domain. This indicates the proportion of time-frequency domain branching loss.
8. A self-supervised ventricular arrhythmia detection system based on dual-domain knowledge distillation, characterized in that, It includes: a sample construction module, a model pre-training module, a model training module, and an anomaly detection module. The sample construction module is used to construct a three-classification task sample dataset based on normal ECG data, pseudo-noise data, and pseudo-abnormal ECG data. The pseudo-noise data is generated by simulating a combination of noise from normal ECG signals, and the pseudo-abnormal ECG data is generated based on normal ECG data and using frequency modulation technology. The model pre-training module is used to construct a time-frequency domain dual-branch processing model architecture. Each processing branch consists of an encoder and a classification head. The time-frequency domain dual-branch processing architecture is used as the student model, and a teacher model with the same structure is constructed based on the time-frequency domain dual-branch processing architecture. The sample signals in the three-class task sample dataset are input into the teacher model, and the teacher model is pre-trained with the goal of minimizing the difference between the teacher model's classification prediction result and the true class of the sample signal. The model training module is used to fix the parameters of the trained teacher model, input the sample signals into the teacher model and the student model respectively, and map the time-frequency domain features output by the encoder of the student model onto the hypersphere structure. The student model is trained with the goal of minimizing the difference between the classification prediction results of the teacher model and the student model and the radius of the hypersphere structure under the normal ECG category. The anomaly detection module is used to take the trained student model as the time-frequency domain anomaly detection model, input the heart rhythm signal to be processed into the time-frequency domain anomaly detection model, and use the time-frequency domain anomaly detection model to detect and identify anomalies in the heart rhythm signal to be processed.
9. An electronic device, characterized in that, include: At least one processor, and a memory coupled to said at least one processor; The memory stores a computer program that can be executed by the at least one processor to implement the method as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, enables the implementation of the method as described in any one of claims 1 to 7.