Arrhythmia diagnosis system based on multi-sensor fusion

Abstract

The invention relates to the technical field of intelligent medical treatment, in particular to an arrhythmia diagnosis system based on multi-sensor fusion, and the system comprises an electrocardiosignal module which comprises an electrocardiosignal sensor and is used for collecting electrocardiosignals; the pulse wave signal module comprises a pulse wave sensor and is used for collecting pulse wave signals; the three-axis acceleration module comprises an acceleration sensor and is used for collecting acceleration signals; the main control chip module serves as a data processing module and is used for receiving and analyzing data to obtain an arrhythmia diagnosis result; the Bluetooth module serves as a communication module and is used for transmitting the electrocardiosignal, the pulse wave signal, the acceleration signal and the arrhythmia diagnosis result to a mobile terminal; and the power management module comprises a charging circuit, a power supply circuit and a switching circuit and is used for providing power supply for the system. The system provided by the invention is stable, portable and low in power consumption, and has important significance on prevention of cardiovascular diseases and electrocardiogram monitoring of the Internet of Things.

Description

Technical Field

[0001] The present invention relates to the field of intelligent medical technology, and in particular to an arrhythmia diagnosis system based on multi-sensor fusion. Background Art

[0002] Heart disease is a disease that poses a great threat to human health. Electrocardiogram, as a means of examining the electrical activity of the heart, can reflect the complete process of the heart. Clinically, a 12-lead system is mostly used. However, traditional 12-lead devices have portability problems. Therefore, wearable electrocardiogram devices have become a research hotspot. With the rapid development of wearable computing technology, it has had a significant impact on the fields of health and medicine. Wearable devices are small and light, suitable for long-term daily monitoring. However, wearable devices still need to be optimized in terms of accuracy and reliability, and sensor technology needs to be improved to meet broader application requirements. The present invention combines technical theories such as electronic information technology, communication technology, and biomedicine, and discloses a dynamic electrocardiogram monitoring terminal for arrhythmia diagnosis, which uses three different sensor modules, an electrocardiogram signal sensor, a pulse wave sensor, and an acceleration sensor, to perform dynamic electrocardiogram monitoring on subjects.

[0003] Currently, certain achievements have been made in the research on electrocardiogram signal filtering algorithms. However, during the use of wearable electrocardiogram monitoring devices, baseline drift and motion artifact interference are often generated due to human activities. In particular, the frequency of electrocardiogram signals during exercise is similar to the interference frequency, making it difficult to separate. Traditional adaptive filtering methods perform poorly in dealing with non-stationary noise and rapidly changing signals, and signal processing algorithms need to be improved. The present invention discloses an optimized filtering algorithm, which processes the collected information, effectively removes myoelectric signals and motion artifacts from the electrocardiogram signals, obtains relatively clean electrocardiogram signals, and improves the robustness and adaptability in a complex noise environment.

[0004] Traditional electrocardiogram analysis relies on manual annotation by experts, which is time-consuming, laborious, and easily affected by subjective factors. In recent years, deep learning-based methods can automatically extract electrocardiogram features and achieve efficient and accurate arrhythmia classification. However, the increase in model complexity also brings challenges to computing resources. Therefore, the present invention discloses an arrhythmia classification model, which introduces a dynamic convolution and multi-scale feature fusion strategy to reduce the model complexity, maintains efficient arrhythmia classification, reduces the demand for computing resources, allows users to perform real-time diagnosis, and synchronizes data to a mobile terminal, facilitating remote monitoring by professional medical staff and family members, and providing a convenient, low-cost, and comfortable technical solution for preventing heart health risks. Summary of the Invention

[0005] In view of the deficiencies in the prior art and in combination with the requirements of practical applications, the present invention provides an arrhythmia diagnosis system based on multi-sensor fusion. The system includes: an electrocardiogram (ECG) signal module, which includes an ECG signal sensor for collecting the ECG signal of a subject; a pulse wave signal module, which includes a pulse wave sensor for collecting the pulse wave signal of the subject; a three-axis acceleration module, which includes an acceleration sensor for collecting the acceleration signal of the subject; a main control chip module, which serves as the data processing module of the system for data reception and data analysis to obtain an arrhythmia diagnosis result; a Bluetooth module, which serves as the communication module of the system for transmitting the ECG signal, the pulse wave signal, the acceleration signal, and the arrhythmia diagnosis result to a mobile terminal; and a power management module, which includes a charging, power supply, and switching circuit for providing power supply to the system. By integrating an ECG signal sensor, a pulse wave sensor, and a three-axis acceleration sensor, the present invention realizes synchronous acquisition and fusion analysis of multi-source physiological signals, significantly improves the accuracy of arrhythmia diagnosis. The main control chip module intelligently processes the composite signal, effectively filters out interference, and makes the diagnosis result more reliable. The Bluetooth module supports real-time data transmission to the mobile terminal, facilitating remote monitoring of the heart health status by users and doctors, realizing all-weather continuous monitoring. The power management module optimizes energy consumption, ensures the long-term effective operation of the system, and takes into account portability and comfort, providing an efficient and convenient solution for the prevention and daily management of heart diseases.

[0006] Optionally, the main control chip module includes: an ECG signal filtering module, which uses the ECG signal as an input signal and is used for preprocessing the input signal; and an arrhythmia diagnosis module, which is used for diagnosing arrhythmia of the subject based on the input signal to obtain the arrhythmia diagnosis result. By preprocessing the original ECG signal through the ECG signal filtering module, the present invention effectively filters out noises such as motion artifacts and electromyogram interference, significantly improves the signal-to-noise ratio, and lays a foundation for accurate diagnosis. The arrhythmia diagnosis module then uses an algorithm model to intelligently analyze the ECG signal, automatically extracts features and identifies abnormal heart rhythms through a deep learning model, greatly improving the diagnosis accuracy rate; it not only ensures the real-time processing efficiency but also improves the diagnosis reliability, enabling the system to stably output professional-level diagnosis results in a complex environment, providing efficient and accurate technical support for heart health monitoring.

[0007] Optionally, the electrocardiogram signal filtering module includes: a filtering algorithm unit for constructing an optimized filtering algorithm based on an adaptive hybrid step factor and a noise power estimation mechanism; and an algorithm experiment unit for filtering the input signal according to the optimized filtering algorithm. By introducing the adaptive hybrid step factor and the noise power estimation mechanism in the filtering algorithm unit, the present invention enables the filtering algorithm to dynamically match the signal characteristics, showing stronger adaptability and robustness in a complex noise environment. The algorithm experiment unit effectively filters out noises such as motion artifacts and electromyogram interference through repeated verification of the optimized filtering algorithm, significantly improving the signal-to-noise ratio of the electrocardiogram signal. The two units cooperate to ensure both the real-time nature of the filtering process and a substantial improvement in signal quality, providing a high-fidelity data basis for subsequent arrhythmia diagnosis, ensuring the stability of filtering in complex scenarios, and significantly enhancing the diagnostic reliability.

[0008] Optionally, constructing the optimized filtering algorithm based on the adaptive hybrid step factor and the noise power estimation mechanism includes: constructing a dynamic step factor based on the adaptive hybrid step factor according to the error signal and the input signal; dynamically adjusting the dynamic step factor according to the noise power estimation mechanism, and constructing the optimized filtering algorithm in combination with a hybrid strategy. By constructing the optimized filtering algorithm through the dynamic step factor and adjusting the parameters in real time according to the error signal and the input signal, the present invention significantly enhances the adaptability of the algorithm to signal changes. The noise power estimation mechanism dynamically optimizes the step factor, finely adjusts the filtering intensity according to the real-time noise level, enabling the algorithm to maintain high-efficiency noise reduction ability in a complex noise environment. Combining with the hybrid strategy, it not only ensures the fidelity of the filtered signal but also improves the processing efficiency, making the filtering process both flexible and accurate, effectively coping with diverse interferences, providing high-quality electrocardiogram signals for subsequent diagnosis, and greatly improving the accuracy of arrhythmia detection and the stability of the system.

[0009] Optionally, constructing the dynamic step factor based on the adaptive hybrid step factor according to the error signal and the input signal includes: wherein, is the dynamic step factor, is the initial step factor, is the energy of the error signal, is the energy of the input signal, is the noise power estimation, is a small positive number. Through the dynamic step factor, the filtering algorithm of the present invention can respond to the changes of signals and noises in real time, automatically reduce the step when the noise is high to avoid overshoot, and increase the step when the noise is low to accelerate convergence, significantly improving the stability and accuracy of filtering; effectively balancing noise suppression and signal fidelity, enabling the algorithm to exhibit stronger adaptability in complex physiological signal environments.

[0010] Optionally, filtering the input signal according to the optimized filtering algorithm includes: acquiring the electrocardiogram signal, the pulse wave signal, and the acceleration signal of the subject in a jogging state, and using the pulse wave signal and the acceleration signal as reference input signals; passing the reference input signals through a cross-interference cancellation system, and establishing an iterative expression of weight coefficients based on an adaptive algorithm to adjust the weight coefficients of the reference input signals; using the optimized filtering algorithm to perform cyclic iteration on the input signal based on the iterative expression of weight coefficients to achieve filtering of the input signal. By collecting multi-source physiological signals in a jogging state, using the pulse wave and acceleration signals as reference inputs, and combining the cross-interference cancellation system with an adaptive algorithm to dynamically adjust the signal weight coefficients, the present invention effectively eliminates motion interference; the optimized filtering algorithm performs cyclic iteration based on the weight coefficients to further purify the electrocardiogram signal, significantly improving the signal-to-noise ratio, not only retaining the key features of the electrocardiogram signal but also stably outputting high-quality signals in complex motion scenarios, providing a reliable basis for subsequent arrhythmia diagnosis, and making the monitoring system more robust and practical in actual applications.

[0011] Optionally, passing the reference input signals through a cross-interference cancellation system and establishing an iterative expression of weight coefficients based on an adaptive algorithm includes: wherein, is the weight coefficient at the th iteration, is the weight coefficient at the th iteration, is the dynamic step factor, is the input signal at the th iteration, is the error signal at the th iteration, is the energy of the input signal, is a small positive number. Through the dynamic adjustment of the weight coefficients, the filtering algorithm of the present invention can respond to the changes of signals and errors in real time, automatically adjust the weight when the error increases to strengthen filtering, and maintain the weight when the signal is stable to retain details, significantly improving the accuracy and stability of adaptive filtering, effectively balancing noise suppression and signal fidelity, enabling the algorithm to exhibit stronger anti-interference ability in complex motion scenarios, and improving the overall performance of the system.

[0012] Optionally, the arrhythmia diagnosis module includes: a model construction unit that constructs an arrhythmia classification model based on dynamic convolution and multi-scale feature fusion strategies; and a model application unit that trains the arrhythmia classification model according to overall sample data and obtains an arrhythmia classification result. The model construction unit of the present invention adopts dynamic convolution and multi-scale feature fusion strategies, enabling the classification model to adaptively adjust the convolution kernel parameters, accurately capture transient features in the electrocardiogram signal, and at the same time fuse physiological information of different scales, significantly improving the accuracy of arrhythmia classification; the model application unit is trained through overall sample data to ensure that the model covers a wide range of physiological scenarios and enhances the generalization ability. The diagnostic system can not only identify complex arrhythmia patterns but also adapt to individual differences, maintaining high performance in diverse clinical data and providing efficient and accurate intelligent assistance for cardiac disease screening.

[0013] Optionally, constructing the arrhythmia classification model based on dynamic convolution and multi-scale feature fusion strategies includes: replacing the one-dimensional convolutional layer with a dynamic one-dimensional convolutional layer, and dynamically adjusting the convolution kernel according to the input signal in combination with the dynamic one-dimensional convolutional layer; introducing a multi-scale convolutional layer in the convolutional neural network, and extracting multi-scale features based on the convolutional kernels of different sizes according to the multi-scale convolutional layer; performing feature fusion on the multi-scale features based on the multi-scale feature fusion strategy to construct fused features; and using the fused features as the feature input of the transformer module to construct the arrhythmia classification model. By adopting a dynamic one-dimensional convolutional layer, the present invention can adjust the convolution kernel parameters in real time according to the characteristics of the input signal, significantly enhancing the adaptability of the model to signal changes. The design of the multi-scale convolutional layer enables the model to capture local details and global features simultaneously, extracting multi-scale information through convolutional kernels of different sizes and improving the richness of feature expression. The feature fusion strategy further integrates multi-scale information to form more representative fused features, enabling the model to exhibit higher classification accuracy in the recognition of complex arrhythmia patterns.

[0014] Optionally, training the arrhythmia classification model based on the overall sample data and obtaining an arrhythmia classification result includes: dividing the overall sample data to construct a training set and a test set, and training the arrhythmia model; obtaining the arrhythmia classification result according to the trained arrhythmia classification model as the arrhythmia diagnosis result. By dividing the overall sample data into a training set and a test set, the present invention effectively avoids the risk of overfitting, ensures that the model has good generalization ability in diverse data, fully optimizes the arrhythmia classification model using the training set, makes the model parameters more conform to the actual data distribution, significantly improves the classification accuracy, objectively evaluates the model performance based on the independent verification of the test set, ensures the reliability of the diagnosis result, improves the model's recognition ability for different arrhythmia types, enhances the practicality of the system in clinical applications, and provides efficient and accurate technical support for arrhythmia diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 FIG. is a hardware framework diagram of an arrhythmia diagnosis system based on multi-sensor fusion according to an embodiment of the present invention; Figure 2 FIG. is an algorithm implementation block diagram of an improved NLMS algorithm according to an embodiment of the present invention; Figure 3 FIG. is a signal waveform diagram of a subject in a jogging state according to an embodiment of the present invention; Figure 4 FIG. is a mean square error comparison diagram of a filtering algorithm according to an embodiment of the present invention; Figure 5 FIG. is a filtering waveform comparison diagram of a filtering algorithm according to an embodiment of the present invention; Figure 6 FIG. is a model structure diagram of an arrhythmia classification model according to an embodiment of the present invention; Figure 7 FIG. is a curve graph of the change of a loss function during the training process of an arrhythmia model according to an embodiment of the present invention; Figure 8 FIG. is a curve graph of the change of accuracy during the training process of an arrhythmia model according to an embodiment of the present invention; Figure 9 FIG. is a confusion matrix diagram of an arrhythmia classification model according to an embodiment of the present invention; Figure 10 FIG. is an information display diagram of a mobile terminal according to an embodiment of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] Specific embodiments of the present invention will be described in detail below. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the present invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention. However, it will be apparent to those of ordinary skill in the art that the present invention may be practiced without these specific details. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the present invention.

[0017] Throughout the specification, references to "one embodiment", "an embodiment", "one example", or "an example" mean that a particular feature, structure, or characteristic described in connection with the embodiment or example is included in at least one embodiment of the present invention. Thus, the phrases "in one embodiment", "in an embodiment", "one example", or "an example" appearing throughout the specification are not necessarily all referring to the same embodiment or example. Additionally, the specific features, structures, or characteristics may be combined in any suitable combination and / or sub-combination in one or more embodiments or examples. Moreover, those of ordinary skill in the art will understand that the diagrams provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0018] Please refer to Figure 1 , according to the design requirements of the portable system, and in combination with signal acquisition, algorithms, and wireless transmission, a set of excellent performance hardware circuit system is constructed as the carrier for signal processing; an embodiment of the present invention provides an arrhythmia diagnosis system based on multi-sensor fusion, the system includes: an electrocardiogram signal module, the electrocardiogram signal module includes an electrocardiogram signal sensor for collecting the electrocardiogram signal of the subject; a pulse wave signal module, the pulse wave signal module includes a pulse wave sensor for collecting the pulse wave signal of the subject; a three-axis acceleration module, the three-axis acceleration module includes an acceleration sensor for collecting the acceleration signal of the subject; a main control chip module, the main control chip module serves as the data processing module of the system for data reception and data analysis to obtain an arrhythmia diagnosis result; a Bluetooth module, the Bluetooth module serves as the communication module of the system for transmitting the electrocardiogram signal, the pulse wave signal, the acceleration signal, and the arrhythmia diagnosis result to a mobile terminal; a power management module, the power management module includes charging, power supply, and switching circuits for providing power supply to the system.

[0019] In this embodiment, the electrocardiogram (ECG) signal module uses a high-precision ADS1292R chip from Texas Instruments and is connected to the main control chip module through the Serial Peripheral Interface (SPI); the pulse wave signal module is connected to the main control chip module through the Inter-Integrated Circuit (I2C); the three-axis acceleration module uses an ADXL345 chip and is connected to the main control chip module through I2C; the main control chip module uses an ultra-low-power STM32L452 series chip from STMicroelectronics; the Bluetooth module uses a V4.0 BLE CC2540F256 chip and is connected to the main control chip module through the Universal Asynchronous Receiver-Transmitter (UART); the power management module includes a charging, power supply, and switching circuit. Among them, the BQ21040 is selected for the charging circuit, the TLV70033DCK voltage regulator chip is selected for the power supply voltage regulator circuit, and the switching circuit uses a PMOS transistor to realize the switching of the system charging and power supply states, thereby reducing the system power consumption to increase the battery life of the ECG detection device.

[0020] In an arrhythmia diagnosis system based on multi-sensor fusion provided by the present invention, the main control chip module includes: A1. An ECG signal filtering module, taking the ECG signal as an input signal, and the ECG signal filtering module is used to preprocess the input signal.

[0021] The ECG signal filtering module includes: A11. A filtering algorithm unit, and the filtering algorithm unit is used to construct an optimized filtering algorithm according to the adaptive hybrid step factor and the noise power estimation mechanism.

[0022] In the prior art, certain achievements have been made in the research on ECG signal filtering algorithms; however, during the use of wearable ECG monitoring devices, baseline drift and motion artifact interference are often generated due to human activities. Especially, the frequency of the ECG signal during exercise is similar to the interference frequency, making it difficult to separate. Traditional adaptive filtering methods, such as the LMS algorithm and the NLMS algorithm, are widely used because of their simple calculation and good performance, but they perform poorly when dealing with non-stationary noise and rapidly changing signals. The traditional NLMS algorithm uses a fixed step factor, resulting in limitations on the convergence speed and stability performance of the filter in different noise environments.

[0023] In this embodiment, an improved NLMS algorithm is proposed, which introduces an adaptive hybrid step factor and a noise power estimation mechanism. By combining the dynamic step factor, noise power estimation, and hybrid strategy, the robustness and adaptability of the algorithm in complex noise environments are improved.

[0024] Specifically, the adaptive hybrid step factor is dynamically adjusted according to the historical information of the error signal and the energy of the input signal. By calculating the square of the current error and the square of the historical error in real time, the current step size is determined, enabling the algorithm to quickly respond to changes in the noise environment.

[0025] Furthermore, to improve the robustness of the algorithm, a noise power estimation mechanism is introduced. The noise power estimation mechanism reduces the step size when the noise is high to avoid excessive adjustment; when the noise is low, it increases the step size to accelerate the convergence speed, thereby ensuring the stability and efficiency of the filter.

[0026] The dynamic step size factor is calculated based on the energy of the current error signal and the energy of the input signal, satisfying the following relationship: where, is the dynamic step size factor, is the initial step size factor, is the energy of the error signal, is the energy of the input signal, is the noise power estimation, is a small positive number.

[0027] The energy of the input signal satisfies the following relationship: where, is the energy of the input signal, is the th iteration of the input signal.

[0028] It should be noted that is a small positive number used to avoid division by zero.

[0029] In this embodiment, combining the advantages of the dynamic step size factor and the noise power estimation, a hybrid strategy is proposed. In each iteration, according to the current error signal and the energy of the input signal, the direction and amplitude of the weight update are adaptively adjusted. The dynamic step size factor is dynamically adjusted based on the noise power estimation mechanism, and an optimized filtering algorithm is constructed by combining the hybrid strategy. The improved NLMS algorithm uses an adaptive step size factor and the computational complexity does not increase significantly.

[0030] Please refer to Figure 2 , which shows the algorithm implementation block diagram of the improved NLMS algorithm. The pulse wave signal and the acceleration signal are used as the reference input signals, and the electrocardiogram signal is used as the original input signal.

[0031] Specifically, the filter output satisfies the following relationship: where, is the filter output, is the sub-filter output, is the pulse wave signal at the The weight coefficient at the is the input signal, and is the weight coefficient of the acceleration signal at the is the error signal, and

[0032] A12, an algorithm experiment unit, is used to filter the input signal according to the optimized filtering algorithm.

[0033] In this embodiment, in order to verify the effectiveness of the improved NLMS algorithm, tests and comparative analyses were carried out through experiments.

[0034] The algorithm experiment unit performs the following steps, including: S1. Obtain the electrocardiogram signal, the pulse wave signal, and the acceleration signal of the subject in the jogging state, and the pulse wave signal and the acceleration signal are used as reference input signals.

[0035] Please refer to Figure 3 for the signal waveform diagram of the subject in the jogging state, including Figure 3 (a) is the electrocardiogram signal waveform diagram, Figure 3 (b) is the pulse wave signal waveform diagram, and Figure 3 (c) is the acceleration signal waveform diagram; they are collected by using an electrocardiogram signal sensor, a pulse wave sensor, and an acceleration sensor respectively, and the sampling frequency is 500 Hz.

[0036] S2. Pass the reference input signal through the interference cancellation system, and establish an iterative expression of the weight coefficient based on the adaptive algorithm to adjust the weight coefficient of the reference input signal.

[0037] Specifically, based on the interference cancellation system, combined with the adaptive algorithm, an iterative expression of the weight coefficient is constructed according to the dynamic step factor, satisfying the following relationship: where is the weight coefficient at the iteration, is the weight coefficient at the iteration, is the dynamic step factor, is the input signal at the iteration, is the error signal at the iteration, is the energy of the input signal, is a small positive number.

[0038] Based on the improved NLMS algorithm and combined with the weight coefficient iteration expression, the weight coefficients of the pulse wave signal and the acceleration signal are adjusted to satisfy the following relationship: Wherein, is the weight coefficient of the pulse wave signal at the -th iteration, is the weight coefficient of the pulse wave signal at the -th iteration, is the dynamic step size factor at the -th iteration, is the error signal, is the input signal, represents transpose, is a small positive number, is the weight coefficient of the acceleration signal at the -th iteration, is the weight coefficient of the acceleration signal at the -th iteration.

[0039] In an optional embodiment, the computational complexities of the weight coefficient update processes of the LMS algorithm, the NLMS algorithm, and the improved NLMS algorithm are compared to obtain the filtering effects of each algorithm, as shown in Table 1: Table 1 Furthermore, please refer to Figure 4 , which shows the mean square error comparison diagram of the filtering algorithms, including the mean square error learning curves of the LMS algorithm, the NLMS algorithm, and the improved NLMS algorithm.

[0040] Combined with Table 1 and Figure 4 for comprehensive analysis, the experimental results show that compared with the NLMS algorithm, the improved NLMS algorithm shows higher robustness and adaptability in noise removal. The signal-to-noise ratio SNR of the improved NLMS algorithm reaches 18.853 dB, which is 16% higher than that of the NLMS algorithm and 56.6% higher than that of the LMS algorithm. The improved NLMS algorithm disclosed in the present invention is significantly superior to the LMS and NLMS algorithms. In a non-stationary signal and a rapidly changing noise environment, the improved algorithm can converge to the steady state more quickly and has a good denoising effect.

[0041] S3. Use the optimized filtering algorithm to perform cyclic iteration on the input signal based on the weight coefficient iteration expression to implement filtering of the input signal.

[0042] In this embodiment, the LMS algorithm, the NLMS algorithm, and the improved NLMS algorithm are designed using Python. In the jogging state, for the electrocardiogram signals containing motion artifact interference and baseline drift noise, the weight coefficient iteration expression is used for cyclic iteration, and the signals are respectively filtered by the LMS algorithm, the NLMS algorithm, and the improved NLMS algorithm to obtain the filtered waveforms.

[0043] Please refer to Figure 5 , the comparison diagram of the filtered waveforms of the filtering algorithms, including the filtered waveforms of the LMS algorithm, the NLMS algorithm, and the improved NLMS algorithm.

[0044] Furthermore, a comparative analysis of the weight coefficient updates of the LMS algorithm, the NLMS algorithm, and the improved NLMS algorithm is shown in Table 2 as follows: Table 2 Combined with Figure 5 and Table 2 for comprehensive analysis, it can be seen that the LMS algorithm has the smallest computational complexity, but is only suitable for environments with limited computing resources. The NLMS algorithm adds additional multiplication and addition operations compared to the LMS algorithm, improving the stability and convergence speed of the algorithm. The improved NLMS algorithm further adds a small number of multiplication and addition operations, and its characteristic of dynamically adjusting the step size factor significantly improves the performance of the algorithm, especially enhancing the adaptability and stability in different noise environments.

[0045] A2. The arrhythmia diagnosis module is used to diagnose arrhythmia of the subject based on the input signal to obtain the arrhythmia diagnosis result.

[0046] The arrhythmia diagnosis module includes: A21. The model construction unit constructs an arrhythmia classification model based on the dynamic convolution and multi-scale feature fusion strategy; In the prior art, traditional electrocardiogram analysis relies on manual annotation by experts, which is time-consuming, laborious, and easily affected by subjective factors. In recent years, deep learning-based methods can automatically extract electrocardiogram features to achieve efficient and accurate arrhythmia classification, but the increase in model complexity also brings challenges to computing resources.

[0047] In this embodiment, the dynamic convolution and multi-scale feature fusion strategy are introduced to construct an arrhythmia classification model to reduce the model complexity. The device computer uses a 4070ti graphics card and an i7-13700kf processor.

[0048] Please refer to Figure 6, which is the model structure diagram of the arrhythmia classification model; an improved CNN-Transformer is obtained based on CNN-Transformer and used as the arrhythmia classification model. First, the input layer is used for the input of signals. The input signals pass through two one-dimensional convolutional layers and are batch-normalized. Secondly, the one-dimensional convolutional layer is replaced by a dynamic one-dimensional convolutional layer, and the convolutional kernel is dynamically adjusted in combination with the dynamic one-dimensional convolutional layer to reduce the number of parameters and the amount of calculation. Subsequently, through the one-dimensional max pooling layer, a multi-scale convolutional layer is introduced in the convolutional neural network. Based on convolutional kernels of different sizes, multi-scale features are extracted according to the multi-scale convolutional layer to enhance the model's representation ability, and the multi-scale features are fused based on the multi-scale feature fusion strategy to construct fused features. Then, the fused features are used as the feature input of the transformer module to capture the temporal dependence relationship in the signals. Finally, through the flattening layer, the dropout layer and two fully-connected layers, and the output is performed through the output layer.

[0049] A22. The model application unit is used to train the arrhythmia classification model according to the overall sample data and obtain the arrhythmia classification result.

[0050] In this embodiment, when training the arrhythmia classification model, the overall sample data is divided. 80% of the overall sample data is used as the training set, and 20% of the overall sample data is used as the test set.

[0051] Please refer to Figure 7 , which is the curve graph of the change of the loss function during the training process of the arrhythmia model; it shows the change of the loss function of the arrhythmia model with the increase of the number of iterations, including the training loss curve and the test loss curve.

[0052] Please refer to Figure 8 , which is the curve graph of the change of the accuracy rate during the training process of the arrhythmia model; it shows the change of the accuracy rate of the arrhythmia model with the increase of the number of iterations, including the training accuracy rate and the test accuracy rate.

[0053] Combined with Figure 7 and Figure 8 Through comprehensive analysis, it can be observed that with the increase of the number of iterations, the accuracy rate during the training process is continuously rising, but the rising speed is getting slower and slower and finally tends to be stable; on the contrary, the change of the loss function during the training process is opposite, and finally it also tends to a stable value; finally, the accuracy rate of the arrhythmia model on the training set reaches 99.2%, and the accuracy rate on the validation set reaches 98%.

[0054] Furthermore, the arrhythmia classification result is obtained based on the trained arrhythmia classification model and used as the arrhythmia diagnosis result.

[0055] In an optional embodiment, after the training process of the arrhythmia classification model is completed, the performance of the arrhythmia classification model is verified using a test set.

[0056] See also Figure 9 , the figure is the confusion matrix diagram of the arrhythmia classification model, which shows the comparison between the predicted labels and the real labels of the five ECG signals N, L, R, A and V by the arrhythmia classification model, and compares the arrhythmia classification results of the arrhythmia classification model with the labels of the MIT-BIH arrhythmia database data, and expresses them in the form of a confusion matrix. The values ​​on the main diagonal represent the number of samples whose real categories are correctly classified as real categories, and the remaining values ​​are the number of samples that are incorrectly classified.

[0057] Furthermore, the accuracy, precision, recall and F1 value of the arrhythmia classification model for the five ECG signals N, L, R, A and V are calculated based on the data in the confusion matrix, and the evaluation index values ​​are obtained, as shown in Table 3: Table 3 As shown in Table 3, the classification accuracy reached more than 98.6%, the classification precision of each label reached more than 95.3%, the recall rate of each label reached more than 95%, and the F1 value of each label reached more than 96.3%. It can be seen that the classification performance of the improved CNN-Transformer arrhythmia classification model is relatively excellent, and each indicator is relatively good, which basically meets the requirements of arrhythmia classification prediction.

[0058] Furthermore, in order to verify the advancement and accuracy of the arrhythmia classification model disclosed in the present invention and to better explore the improvability of the model, comparative experiments were carried out on the DNN model, CNN model, and CNN-Transformer model using the same data set, and the accuracy performance of the above three models and the improved CNN-Transformer arrhythmia classification model of the present invention under the same data set was compared. The accuracy comparison results are shown in Table 4: Table 4 It can be seen from Table 4 that the improved CNN-Transformer arrhythmia classification model adopted in the present invention has excellent classification performance, indicating that the arrhythmia classification model has a good effect in electrocardiogram classification work.

[0059] In another alternative embodiment, the arrhythmia diagnosis system (multi-sensor diagnosis terminal) based on multi-sensor fusion proposed by the present invention is actually tested and the results are analyzed. In the actual test, a total of 10 groups of experimental data of subjects aged 18 - 30 are collected, including 7 groups of males and 3 groups of females. None of the participants have movement disorders or abnormal heart rates.

[0060] First, connect the multi-sensor diagnosis terminal to the subject; wear a pulse wave sensor on the left index finger; since the waist position is close to the center of gravity of the human body and can better reflect the overall movement state, the acceleration sensor is fixed on the waist of the human body; and the electrocardiogram signal sensor connection electrode patches are worn on the left and right chests and the right abdomen respectively; by wearing the pulse wave sensor, electrocardiogram signal sensor and acceleration sensor, synchronously collect the electrocardiogram signal, pulse wave signal and acceleration signal for 1 minute and save them as a CSV file; after denoising through multi-sensor information fusion, use the trained improved CNN-Transformer arrhythmia classification model to classify and predict the data. The classification probability results of the electrocardiogram signal are shown in Table 5: Table 5 As can be seen from Table 5, the improved CNN-Transformer arrhythmia classification model proposed by the present invention has good classification and prediction performance.

[0061] Please refer to Figure 10 , which is an information display diagram of the mobile terminal; in order to display the vital signs signal in real time, an Android platform APP is developed, which can display the electrocardiogram signal, pulse wave signal and acceleration signal through the Bluetooth module, and monitor the electrocardiogram state in real time, with functions such as signal display, data saving and historical data review.

[0062] In summary, the present invention provides an arrhythmia diagnosis system based on multi-sensor fusion. Against the background of the continuous increase in the global incidence of cardiovascular diseases, a wearable electrocardiogram monitoring system with multi-sensor information fusion is designed; a hardware acquisition system based on the STM32L452CEU6 chip is designed, and the ADS1292R and ADXL345 chips are used to collect the electrocardiogram signal, pulse wave signal and acceleration signal, and transmit them to the mobile terminal through the Bluetooth module; an improved NLMS algorithm is proposed, which adopts an adaptive step factor and a noise power estimation mechanism to fuse the acceleration and pulse wave signals for filtering, and the signal-to-noise ratio is increased by 16% to reach 18.853 dB; the CNN-Transformer algorithm is improved to construct an improved CNN-Transformer arrhythmia classification model, introducing a dynamic convolution and multi-scale feature fusion strategy to achieve a lightweight model, and the classification accuracy rate is above 98.6%; the test results show that the system is stable, portable and low-power consumption, and is of great significance for the prevention of cardiovascular diseases and the Internet of Things electrocardiogram monitoring.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and the specification of the present invention.

Claims

1. A cardiac arrhythmia diagnosis system based on multi-sensor fusion, characterized in that: include: An ECG signal module, the ECG signal module comprising an ECG signal sensor for collecting an ECG signal of a subject; A pulse wave signal module, the pulse wave signal module comprising a pulse wave sensor for collecting a pulse wave signal of the subject; A three-axis acceleration module, wherein the three-axis acceleration module includes an acceleration sensor for collecting an acceleration signal of the subject; A main control chip module, which serves as a data processing module of the system and is used for data reception and data analysis to obtain arrhythmia diagnosis results; A Bluetooth module, which serves as a communication module of the system and is used to transmit the electrocardiogram signal, the pulse wave signal, the acceleration signal and the arrhythmia diagnosis result to a mobile terminal; A power management module, the power management module includes charging, power supply and switching circuits, and is used to provide power supply for the system.

2. The arrhythmia diagnosis system based on multi-sensor fusion according to claim 1, characterized in that: The main control chip module includes: An ECG signal filtering module, wherein the ECG signal is used as an input signal, and the ECG signal filtering module is used to pre-process the input signal; An arrhythmia diagnosis module is used to perform arrhythmia diagnosis on the subject according to the input signal to obtain the arrhythmia diagnosis result.

3. The arrhythmia diagnosis system based on multi-sensor fusion according to claim 2, characterized in that: The electrocardiogram signal filtering module comprises: A filtering algorithm unit, wherein the filtering algorithm unit is used to construct an optimized filtering algorithm according to an adaptive hybrid step size factor and a noise power estimation mechanism; An algorithm experiment unit, wherein the algorithm experiment unit is used to filter the input signal according to the optimized filtering algorithm.

4. The arrhythmia diagnosis system based on multi-sensor fusion according to claim 3 is characterized in that: The optimization filtering algorithm is constructed according to the adaptive hybrid step size factor and the noise power estimation mechanism, including: Based on the adaptive hybrid step size factor, construct a dynamic step size factor according to the error signal and the input signal; The dynamic step size factor is dynamically adjusted according to the noise power estimation mechanism, and the optimized filtering algorithm is constructed in combination with a hybrid strategy.

5. The arrhythmia diagnosis system based on multi-sensor fusion according to claim 4 is characterized in that: The step of constructing a dynamic step factor based on the adaptive hybrid step factor according to the error signal and the input signal comprises: in, is the dynamic step size factor, is the initial step size factor, is the energy of the error signal, is the energy of the input signal, is the noise power estimate, A small positive number.

6. The arrhythmia diagnosis system based on multi-sensor fusion according to claim 3, characterized in that: The filtering the input signal according to the optimized filtering algorithm comprises: Acquiring the electrocardiogram signal, the pulse wave signal and the acceleration signal of the subject in a jogging state, wherein the pulse wave signal and the acceleration signal are used as reference input signals; Passing the reference input signal through a phase interference cancellation system, and establishing a weight coefficient iterative expression based on an adaptive algorithm to adjust the weight coefficient of the reference input signal; The optimized filtering algorithm is utilized to perform cyclic iteration on the input signal based on the weight coefficient iterative expression to achieve filtering of the input signal.

7. The arrhythmia diagnosis system based on multi-sensor fusion according to claim 6, characterized in that: The step of passing the reference input signal through a phase interference cancellation system and establishing an iterative expression of weight coefficients based on an adaptive algorithm includes: in, For the The weight coefficient at the iteration, For the The weight coefficient at the iteration, is the dynamic step size factor, For the The input signal at the iteration, For the The error signal at the iteration, is the energy of the input signal, A small positive number.

8. The arrhythmia diagnosis system based on multi-sensor fusion according to claim 2, characterized in that: The arrhythmia diagnosis module comprises: A model building unit, wherein the model building unit builds an arrhythmia classification model based on dynamic convolution and multi-scale feature fusion strategy; A model application unit, wherein the model application unit is used to train the arrhythmia classification model according to the overall sample data and obtain an arrhythmia classification result.

9. The arrhythmia diagnosis system based on multi-sensor fusion according to claim 8, characterized in that: The arrhythmia classification model is constructed based on dynamic convolution and multi-scale feature fusion strategy, including: The one-dimensional convolution layer is replaced by a dynamic one-dimensional convolution layer, and the convolution kernel is dynamically adjusted according to the input signal in combination with the dynamic one-dimensional convolution layer; Introducing a multi-scale convolutional layer into a convolutional neural network, and extracting multi-scale features based on the convolutional kernels of different sizes according to the multi-scale convolutional layer; Based on the multi-scale feature fusion strategy, the multi-scale features are fused to construct fused features; The fused features are used as feature inputs of a converter module to construct the arrhythmia classification model.

10. The arrhythmia diagnosis system based on multi-sensor fusion according to claim 8, characterized in that: The arrhythmia classification model is trained according to the overall sample data to obtain an arrhythmia classification result, including: Dividing the overall sample data to construct a training set and a test set, and training the arrhythmia model; The arrhythmia classification result is obtained according to the trained arrhythmia classification model as the arrhythmia diagnosis result.

Images