Respiratory frequency and heart rate joint detection method and equipment based on acceleration and LSTM (Long Short Term Memory)

Abstract

The invention relates to a respiratory rate and heart rate joint detection method and equipment based on acceleration and LSTM (Long Short Term Memory), and the method comprises the steps: carrying out the preprocessing of an obtained human body acceleration time sequence signal, and retaining signals in a frequency band where the respiratory rate and the heart rate are located; extracting an acceleration feature, a time domain feature and a frequency domain feature from the preprocessed signal; inputting the three features into a trained LSTM network model, and outputting prediction results of the respiratory rate and the heart rate in real time; the acceleration characteristic is the amplitude of the acceleration. Compared with the prior art, the method has the advantages of high noise immunity, high prediction precision, high efficiency and the like.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly to a method and device for jointly detecting respiratory rate and heart rate based on acceleration and LSTM. Background Art

[0002] With the rapid development of medical health monitoring and portable devices, heart rate and respiratory rate, as important indicators of human physiological signals, accurate detection of which is of great significance for health management, disease diagnosis, and monitoring of exercise status. Currently, traditional detection methods mainly rely on the following two types of technologies:

[0003] (1) Photoplethysmography (PPG) technology based on optics: This method reflects heart rate and respiratory rate by measuring changes in skin light absorption, and has high precision. However, it is sensitive to lighting conditions and is easily affected by factors such as motion artifacts and poor skin contact, which limits its application range.

[0004] (2) Electrocardiogram (ECG) technology based on electrical signals: This method detects heart rate and respiratory rate by recording the electrical activity signals of the heart, with high precision and mature application. However, ECG technology usually requires patch electrodes to be in direct contact with the skin, resulting in poor user wearing experience, and may cause skin discomfort during long-term use.

[0005] In recent years, detection methods based on acceleration sensors have received increasing attention due to their advantages such as non-contact, low power consumption, and portability. Acceleration sensors can provide rich time series data by capturing human motion and micro-vibration signals, providing a new solution for the detection of heart rate and respiratory rate. However, existing traditional detection methods based on acceleration signals mainly rely on the following means:

[0006] (1) Signal processing techniques: including Fourier transform, wavelet transform, etc., for extracting the frequency domain features of signals. These methods rely on manually designed features and are difficult to adapt to complex signal changes in different application scenarios.

[0007] (2) Simple statistical models: Using features such as peaks and zero crossings in acceleration signals for statistical analysis. This method has poor robustness to noise and is prone to losing key information in non-stationary signals.

[0008] Although the above methods can achieve certain effects under specific conditions, they still have the following defects in practical applications:

[0009] (1) Insufficient anti-noise ability: Acceleration signals are easily interfered by external factors such as body movement and environmental vibration, and the signal processing capabilities of traditional methods are difficult to effectively suppress noise.

[0010] (2) Limitations in feature extraction: Traditional methods rely too much on manually designed features, with insufficient applicability and generalization ability, and are difficult to handle complex and non-linear data features.

[0011] (3) Trade-off between real-time performance and accuracy: On portable devices, traditional methods are difficult to balance the requirements of high accuracy and low computational complexity.

[0012] After retrieval, Chinese Patent Application Publication No. CN110507299B discloses a heart rate signal detection device, including: a sample extractor for evenly dividing continuous human heart rate signals into signal samples of a plurality of fixed time lengths; a sample processor for determining a feature map of the signal sample, the feature map including time domain features and frequency domain features; a probability calculator for determining the probability distribution of the signal sample belonging to each heart rate type according to the feature map, the heart rate type including normal heart rate type and arrhythmia type. This existing patent application has the problem of low detection accuracy.

[0013] How to achieve high-precision and efficient joint detection of respiratory rate and heart rate has become a technical problem to be solved. Summary of the Invention

[0014] The purpose of the present invention is to overcome the defects of the above-mentioned existing technologies and provide a method and device for joint detection of respiratory rate and heart rate based on acceleration and LSTM.

[0015] The purpose of the present invention can be achieved through the following technical solutions:

[0016] According to one aspect of the present invention, a method for joint detection of respiratory rate and heart rate based on acceleration and LSTM is provided, and the method includes the following steps:

[0017] Step S1, preprocess the acquired acceleration time series signal, and retain the signal within the frequency band where the respiratory rate and heart rate are located;

[0018] Step S2, extract acceleration features, time domain features and frequency domain features from the preprocessed signal;

[0019] Step S3, input the three features in Step S2 into the trained LSTM network model, and output the prediction results of the respiratory rate and heart rate in real time;

[0020] The acceleration feature is the amplitude of the acceleration. The acceleration signals at time t are x(t), y(t) and z(t), and the amplitude of the acceleration is calculated by the following formula:

[0021]

[0022] Among them, x(t), y(t), and z(t) are the acceleration components in the x, y, and z directions at time t, respectively.

[0023] Preferably, the preprocessing includes signal denoising with a band-pass filter, signal smoothing, and signal normalization.

[0024] Preferably, the frequency-domain features of the preprocessed acceleration signal are extracted by fast Fourier transform (FFT).

[0025] Preferably, the time-domain features include the mean and variance related to the breathing frequency and the kurtosis and skewness related to the heart rate based on the preprocessed acceleration signal, which are used to reflect the periodic change characteristics of the breathing frequency and heart rate signals in the acceleration signal.

[0026] Preferably, the LSTM network model takes the error between the predicted heart rate and breathing frequency as the optimization objective, and uses the mean square error as the loss function to evaluate the deviation of the network model in predicting physiological signals.

[0027] More preferably, for the prediction errors of the breathing frequency and heart rate signals, the backpropagation algorithm is used to update the weights of the LSTM by the gradient descent method to optimize the loss function.

[0028] Preferably, the output gate of the LSTM network maps the hidden state of the last time step of the LSTM to the predictions of the heart rate and breathing frequency.

[0029] Preferably, the method further includes evaluating the prediction results, and the evaluation metrics include the mean square error and the mean absolute error.

[0030] Preferably, the acceleration signal is obtained by an acceleration sensor.

[0031] According to another aspect of the present invention, an electronic device is provided, including a memory and a processor, where a computer program is stored on the memory, and the processor implements the method when executing the program.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1) The present invention extracts acceleration features, time-domain features, and frequency-domain features from the time series data obtained by an acceleration sensor, and realizes real-time joint detection of the human heart rate and breathing frequency through training of the LSTM network model, having the advantages of strong anti-noise performance, high prediction accuracy, and high efficiency.

[0034] 2) The present invention performs band-pass filtering on the acceleration time series signal, retains the heart rate and breathing frequency, and removes high-frequency noise and low-frequency drift through signal smoothing, reducing the influence of data noise on model training and prediction, and effectively ensuring the quality of the input of the prediction model.

[0035] 3) Joint prediction enables the LSTM model to utilize the information of another task when learning one task, thereby improving the prediction accuracy of both metrics.

[0036] 4) Joint prediction shares some network parameters, provides more comprehensive physiological information for users, and reduces the training complexity of the model. It is applicable to the real-time monitoring of heart rate and respiratory rate in portable devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic flowchart of the detection method in the present invention;

[0038] Figure 2 It is a schematic diagram of the training process of the LSTM model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0040] This embodiment relates to a method for jointly detecting respiratory rate and heart rate based on acceleration and LSTM, as Figure 1 , and the method includes the following steps:

[0041] Step 1, data preprocessing

[0042] To ensure the accuracy and robustness of the model, a series of preprocessing is required for the original acceleration signal obtained from the acceleration sensor, where the acceleration sensor can be placed on parts of the human body such as the chest, abdomen, forehead, and jaw. The preprocessing steps include operations such as signal denoising, smoothing, and normalization. The specific steps are as follows:

[0043] Step 101, signal denoising

[0044] The original acceleration signal often contains various noises, especially high-frequency noise and low-frequency drift. To remove these noises, a band-pass filter can be used for signal denoising. The band-pass filter can retain the frequency bands of interest while filtering out irrelevant frequency components.

[0045] The sampling frequency f of the signal is 1 Hz, and the design goal of the band-pass filter is to filter out frequency components below 0.1 Hz or above 1 Hz. The frequency response H(f) of the band-pass filter can be expressed as:

[0046]

[0047] In actual use, a Butterworth filter can be selected, and its transfer function is:

[0048]

[0049] where f c is the cut-off frequency and n is the order of the filter. By designing an appropriate filter, frequency components below 0.1 Hz or above 1 Hz can be removed, and the main information of the heart rate and respiratory rate can be retained.

[0050] Step 102: Signal smoothing

[0051] To further remove noise, the sliding average method is used to smooth the signal. The sliding average helps to eliminate high-frequency noise and maintain the trend of the signal. This smoothing operation is calculated for the acceleration signals of the x, y, and z axes respectively at each time step to filter out rapidly changing noise. Here, only the x-axis is taken as an example for smoothing. Assuming the sliding window size is N, the formula for calculating the sliding average value is:

[0052]

[0053] where S(t) is the smoothed acceleration signal at time t, a(i) is the value of the acceleration signal after signal denoising at time i, and N is the sliding window size.

[0054] Step 103: Signal normalization

[0055] Since the amplitude of the acceleration signal may be affected by factors such as the device position and posture changes, normalization can eliminate the influence of these external factors. A common normalization method is to subtract the mean and divide by the standard deviation. Signal normalization is calculated for the acceleration signals of the x, y, and z axes respectively at each time step. Here, only the x-axis is taken as an example for signal normalization. For the acceleration signal of the x-axis, the normalization formula is:

[0056]

[0057] where μ x is the mean of the acceleration signal x(t), and σ x is the standard deviation of the acceleration signal.

[0058] The present invention combines a band-pass filter and signal smoothing technology to remove high-frequency noise and low-frequency drift in the original acceleration signal, effectively ensuring the quality of the model input, specifically:

[0059] Strong denoising ability: The band-pass filter can effectively filter out irrelevant frequency components and retain the main frequency bands of the heart rate and respiratory rate, while the sliding average further smooths the signal and reduces the influence of high-frequency noise.

[0060] Signal quality improvement: Through these processing steps, the signal quality input into the LSTM model has been greatly improved, reducing the impact of data noise on model training and prediction.

[0061] Enhanced model generalization ability: The application of denoising and smoothing techniques enables the LSTM model to better generalize in different environments, thereby improving the stability and robustness of the system.

[0062] Step 2, Feature extraction

[0063] Extract the features for the LSTM model from the denoised and smoothed signals, including acceleration features, time-domain features, and frequency-domain features. The extraction of these three parts of features can be carried out in parallel.

[0064] Step 201: Acceleration feature: Acceleration amplitude calculation

[0065] The acceleration signal contains all the information of human movement. Directly using the data of the acceleration signal on each axis may not be intuitive enough. To obtain more representative features, calculate the amplitude of the acceleration signal. The acceleration signal at each moment is x(t), y(t), and z(t). The amplitude of the acceleration can be calculated by the following formula:

[0066]

[0067] where x(t), y(t), and z(t) are the acceleration components in the x, y, and z directions respectively. In this way, we obtain a feature representing the overall acceleration change.

[0068] Step 202: Frequency-domain feature (Periodic feature extraction (FFT))

[0069] Since the heart rate and respiratory rate are periodic signals, frequency-domain analysis is crucial for capturing the periodic features of the signals. At each time step, perform frequency-domain feature extraction on the acceleration signals of the x, y, and z axes respectively. Here, only the x-axis is taken as an example for frequency-domain feature extraction. Use the fast Fourier transform (FFT) to extract the frequency-domain features of the x-axis acceleration signal. The formula for FFT is as follows:

[0070]

[0071] Among them, X(f) is the Fourier transform of the acceleration signal at frequency f, x(t) is the acceleration signal on the x-axis in the time domain, and T is the total duration of the acceleration signal. The FFT calculation can obtain the spectrum of the signal. By selecting the typical frequency ranges of the heart rate and respiratory rate (for example, the heart rate is generally between 0.8 - 2 Hz, and the respiratory rate is generally between 0.1 - 0.5 Hz), the frequency components contributing to the heart rate and respiratory rate can be extracted.

[0072] Step 203: Time-domain features

[0073] In addition to frequency-domain features, some common time-domain features can also be extracted, such as the mean, variance, kurtosis, and skewness of the signal. These features help to further capture the dynamic characteristics of the signal and are used to effectively capture the distribution and deviation of the signal. The time-domain feature extraction is performed on the acceleration signals of the x, y, and z axes respectively. Here, only the acceleration signal on the x-axis is taken as an example for processing.

[0074] Mean:

[0075]

[0076] Variance:

[0077]

[0078] Kurtosis:

[0079]

[0080] Skewness:

[0081]

[0082] Among them, μ x is the mean of the acceleration signal x(t) on the x-axis in the time domain, σ x is the standard deviation of the acceleration signal on the x-axis, and T is the total duration of the acceleration signal.

[0083] The present invention enhances the input features of the LSTM model by calculating the amplitude of the acceleration signal and extracting time-domain features and frequency-domain features (such as the FFT spectrum). These additional features can provide more physiological information to help the LSTM model better understand and predict the heart rate and respiratory rate. The introduction of the amplitude and frequency-domain features effectively increases the diversity of the input information, enabling the model to more comprehensively understand the change trend and periodic characteristics of the signal.

[0084] Through the extraction of frequency-domain features, the periodic components related to the heart rate and respiratory rate can be accurately identified, thereby improving the accuracy and precision of the model prediction.

[0085] In a noisy environment, frequency-domain features can help the model better identify and eliminate the influence of noise, thereby reducing prediction errors.

[0086] Step 3, LSTM model training

[0087] The LSTM model is used to learn the time-series patterns of heart rate and respiratory rate from the processed acceleration signals. LSTM is a recurrent neural network that can capture long-term dependencies.

[0088] Step 301, Structure of the LSTM network model

[0089] The features of the x, y, and z axes are processed separately through the LSTM network model. Here, only the acceleration features of the x-axis are taken as an example for processing.

[0090] The LSTM network model consists of the following parts:

[0091] Input gate: Controls the importance of the current input information.

[0092] i t = σ(W i · [h t-1 , x t + b i )

[0093] Forget gate: Controls the degree of forgetting of the memory cell.

[0094] f t = σ(W f · [h t-1 , x t + b f )

[0095] Memory cell: Updates the content of the memory cell.

[0096] c t = f t * c t-1 + i t * tanh(W c · [h t-1 , x t + b c )

[0097] Output gate: Determines the output at the current moment.

[0098] h t = o t * tanh(c t )

[0099] Among them, h t and h t-1Outputs at time t and t-1 respectively; c t Memory unit at time t; c t-1 Memory unit at time t-1; x t Input x-axis features, including acceleration features, time-domain features, and frequency-domain features; o t Output gate; σ and tanh are both activation functions; W i ,W f, W c , are all weight matrices; b i ,b f ,b c are all bias terms.

[0100] The training process of the LSTM network model is as Figure 2 shown. When dealing with large amounts of data, the present invention adopts GPU acceleration technology to improve the computing speed. Especially when dealing with long time series data, using GPU can significantly reduce the training and inference time and improve the model computing efficiency. Substantial improvement in computing speed: Through GPU acceleration, the computing time can be greatly shortened, enabling the system to respond within milliseconds in real-time monitoring and prediction. Efficient processing of large data: The advantage of GPU parallel computing enables the system to maintain a high processing speed even in the face of massive data and adapt to different scales of data sets. Optimization of model training: Through GPU acceleration, a larger data set can be used for training, improving the accuracy and generalization ability of the model, thereby optimizing the system performance.

[0101] Step 302: Loss function and optimization

[0102] The LSTM network model uses the mean squared error (MSE) of the respiratory rate and heart rate as the loss function:

[0103]

[0104] where p i is the label of the true respiratory rate and heart rate, is the predicted value of the respiratory rate and heart rate output by the model, and N is the number of samples.

[0105] Use the backpropagation algorithm to update the weights of the LSTM through gradient descent to optimize the loss function.

[0106] Step 4, Heart rate and respiratory rate prediction

[0107] Input the acceleration features, time-domain features, and frequency-domain features extracted in Step 2 into the trained LSTM network model for predictive output. The output gate of the LSTM network maps the hidden state of the last time step of the LSTM to the estimated values of the heart rate and respiratory rate.

[0108] The present invention not only predicts heart rate alone, but also jointly predicts respiratory rate. By taking the prediction tasks of these two physiological indicators as the output of the LSTM simultaneously, it can provide more comprehensive physiological monitoring results.

[0109] Multi-task learning: Joint prediction enables the LSTM model to utilize the information of another task when learning one task, thereby improving the prediction accuracy of the two indicators.

[0110] High model efficiency: Joint training can share some network parameters, reduce the training complexity of the model, and maintain high-precision output at the same time.

[0111] Improve physiological monitoring: Simultaneously predicting heart rate and respiratory rate can provide users with more comprehensive physiological information and is applicable to various application scenarios such as health monitoring and sleep monitoring.

[0112] Step 5, Real-time detection and evaluation

[0113] Step 501: Real-time prediction

[0114] The LSTM network model of the present invention can continuously process acceleration signals in real-time applications. Each time the acceleration data of one moment is input, the model will output the predicted heart rate and respiratory rate.

[0115] Step 502: Accuracy evaluation

[0116] By comparing with the real heart rate and respiratory rate data, the prediction accuracy of the model is evaluated. The main evaluation indicators include mean square error (MSE) and mean absolute error (MAE), which are used to quantify the prediction performance of the model:

[0117] The present invention proposes to use an LSTM (Long Short-Term Memory network) model to predict heart rate and respiratory rate based on acceleration signals. Compared with traditional time series data analysis methods (such as linear regression, ARIMA model, etc.), LSTM can effectively capture the long-term dependence relationships in the signals and overcome the limitations of traditional methods in dealing with complex time series signals.

[0118] LSTM can utilize the input information of the previous period and make full use of the time dependence in historical data during prediction, thereby improving the prediction accuracy. Compared with traditional methods, LSTM has a higher tolerance for signal noise, can extract features related to heart rate and respiratory rate from acceleration signals, avoid noise interference, and has strong robustness. The LSTM network has stronger adaptability and can adjust internal parameters according to the changes in input data to adapt to complex and dynamic physiological signals, so it has good dynamic adaptability.

[0119] The LSTM network model adopted by the present invention optimizes the network structure and calculation process while ensuring high prediction accuracy, taking into account the requirements of real-time calculation, so that it still maintains high calculation efficiency in the case of a large amount of data. It can adapt to various device environments, support real-time heart rate and respiratory rate monitoring, and is suitable for a variety of scenario applications such as portable health monitoring devices and smart wearable devices.

[0120] The present invention uses a sliding window technique to predict the real-time heart rate and respiratory rate. The sliding window technique can achieve dynamic monitoring of the acceleration signal, and each window can independently predict the result, which enables the entire system to perform real-time feedback in practical applications. Compared with the traditional batch processing method, the sliding window method enables the system to perform fast calculations every time new data is received, improving the system's response speed and calculation efficiency. This method can adapt to different time scales, and users can adjust the size and step length of the window according to their needs, so as to flexibly adjust the system performance according to actual requirements. By inputting acceleration data of a certain duration each time, the model can dynamically update the prediction result, which makes this method applicable to real-time monitoring and feedback.

[0121] Compared with the prior art, the present invention has made innovations in many aspects, providing a more accurate, robust, and real-time heart rate and respiratory rate detection scheme. Through the combination of technologies such as the LSTM network model, signal processing methods, GPU acceleration, and feature extraction, the present invention not only improves the prediction accuracy, but also greatly improves the real-time performance and calculation efficiency of the system, and has broad application prospects.

[0122] This embodiment also relates to a method for jointly detecting respiratory rate and heart rate based on acceleration and LSTM. This embodiment conducts experiments based on two respiratory and heart rate data sets, which are: the PhysioNet respiratory database and the hospital self-test data set. These two data sets are obtained based on the sleeping posture, and actually data obtained based on other postures are also valid. In the experiment, a comparison of the time series analysis method based on LSTM was conducted, including experimental comparisons of the following three methods respectively:

[0123] 1) Traditional signal processing methods (such as peak detection, FFT frequency domain analysis);

[0124] 2) Traditional machine learning models (such as support vector machine SVM, random forest RF);

[0125] 3) The detection method based on acceleration and LSTM proposed by the present invention.

[0126] All experiments adopt the standard data set segmentation method, and the data set is divided according to the ratio of 70% training, 15% validation, and 15% testing. Each group of experiments is repeated 5 times, and the average result is taken to ensure the stability of the result.

[0127] Experimental setup:

[0128] Dataset 1: The PhysioNet respiration database contains respiration frequency signals of 30 subjects, with a sampling rate of 1 Hz and a total duration of 24 hours.

[0129] Dataset 2: The hospital self - measured dataset contains respiration frequency and heart rate signals of 50 patients, with a sampling rate of 1 Hz and a duration of 8 hours.

[0130] LSTM model parameters:

[0131] Hidden layer size: 64

[0132] Time step: 10 seconds (corresponding to 10 sampling points)

[0133] Optimizer: Adam

[0134] Loss function: Mean Squared Error (MSE)

[0135] Learning rate: 0.001

[0136] Traditional signal processing method: The peak detection method calculates the respiration frequency by detecting the periodic peaks in the signal.

[0137] Traditional machine learning models: Support Vector Machine (SVM) and Random Forest (RF) methods are based on FFT feature extraction, and the signal features are input into the classifier for regression prediction.

[0138] Table 1 and Table 2 show the experimental results using different methods based on Dataset 1 and Dataset 2 respectively, comparing the training error and the test error.

[0139] Table 1

[0140] Method Training Error (RMSE) Testing Error (RMSE) Traditional Signal Processing Method (Peak Detection) 1.62 2.45 Traditional Machine Learning Method (RF) 1.21 1.88 LSTM Method Proposed by the Present Invention 0.93 1.25

[0141] Table 2

[0142] Method Training Error (RMSE) Testing Error (RMSE) Traditional Signal Processing Method (FFT) 1.25 2.01 Traditional Machine Learning Method (SVM) 1.15 1.89 LSTM Method Proposed by the Present Invention 0.82 1.23

[0143] It can be seen from Table 1 and Table 2 that the proposed joint detection method based on acceleration and LSTM has significantly lower training error and test error than the traditional signal processing method and the traditional machine learning method on both datasets. Comparing the non - joint detection results of the same method for the respiration frequency signal in Dataset 1 and the joint detection results for the respiration frequency and heart rate signals in Dataset 2 in the two tables, the joint detection results have higher accuracy than the non - joint detection.

[0144] It has the following advantages:

[0145] Lower training error: The LSTM model can capture long-term dependencies in time series through a recurrent neural network, thereby improving the training effect of the model.

[0146] Lower test error: The LSTM model has strong generalization ability, avoids the strong dependence on feature engineering of traditional signal processing methods, and performs better on test data.

[0147] Strong anti-noise ability: The LSTM model of the present invention can automatically suppress noise in acceleration signals through a deep learning architecture, and is more robust than traditional peak detection methods.

[0148] In addition, the LSTM model of the present invention can be adjusted according to different respiration and heart rate data sets and is applicable to various medical monitoring scenarios.

[0149] The electronic device of the present invention includes a central processing unit (CPU), which can execute various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or computer program instructions loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM, and RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.

[0150] Multiple components in the device are connected to the I / O interface, including: an input unit, such as a keyboard, a mouse, etc.; an output unit, such as various types of displays, speakers, etc.; a storage unit, such as a magnetic disk, an optical disc, etc.; and a communication unit, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit allows the device to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0151] The processing unit executes the various methods and processes described above. For example, in some embodiments, the method can be implemented as a computer software program that is tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device via the ROM and / or the communication unit. When the computer program is loaded into the RAM and executed by the CPU, one or more steps of the method described above can be executed. Alternatively, in other embodiments, the CPU can be configured to execute the method in any other suitable manner (e.g., by means of firmware).

[0152] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that can be used include: Field Programmable Gate Arrays (FPGA), Application Specific Integrated Circuits (ASIC), Application Specific Standard Products (ASSP), System on a Chip (SOC), Complex Programmable Logic Devices (CPLD), and so on.

[0153] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing devices, such that when the program codes are executed by the processor or controller, the functions / operations specified in the flowchart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, executed partially on the machine and partially on a remote machine as an independent software package, or executed entirely on a remote machine or server.

[0154] In the context of the present invention, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a Random Access Memory (RAM), a Read Only Memory (ROM), an Erasable Programmable Read Only Memory (EPROM or Flash Memory), an optical fiber, a portable Compact Disc Read Only Memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0155] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for joint detection of respiratory rate and heart rate based on acceleration and LSTM, characterized in that: The method comprises the following steps: Step S1, preprocessing the acquired acceleration time series signal to retain the signal within the frequency band where the respiratory frequency and heart rate are located; Step S2, extracting acceleration features, time domain features and frequency domain features from the preprocessed signal; Step S3, inputting the three features of step S2 into the trained LSTM network model, and outputting the prediction results of respiratory rate and heart rate in real time; The acceleration characteristic is the amplitude of acceleration. The acceleration signals at time t are x(t), y(t) and z(t). The amplitude of acceleration is calculated by the following formula: Among them, x(t), y(t) and z(t) are the acceleration components in the x, y and z directions at time t respectively.

2. According to claim 1, a respiratory rate and heart rate joint detection method based on acceleration and LSTM is characterized in that: The preprocessing includes signal denoising, signal smoothing and signal standardization processing by a bandpass filter.

3. The method for joint detection of respiratory rate and heart rate based on acceleration and LSTM according to claim 1, characterized in that: The frequency domain features of the pre-processed acceleration signal are extracted by fast Fourier transform (FFT).

4. The method for joint detection of respiratory rate and heart rate based on acceleration and LSTM according to claim 1, characterized in that: The time domain features include the mean and variance related to the respiratory frequency and the kurtosis and skewness related to the heart rate based on the preprocessed acceleration signal, and are used to reflect the periodic variation characteristics of the respiratory frequency and heart rate signals in the acceleration signal.

5. The method for joint detection of respiratory rate and heart rate based on acceleration and LSTM according to claim 1, characterized in that: The LSTM network model uses the error of predicted heart rate and respiratory rate as the optimization target, and uses mean square error as the loss function to evaluate the deviation of the network model in predicting physiological signals.

6. The method for joint detection of respiratory rate and heart rate based on acceleration and LSTM according to claim 5, characterized in that: For the prediction errors of respiratory rate and heart rate signals, the back propagation algorithm is used to optimize the loss function by updating the weights of LSTM through the gradient descent method.

7. The method for joint detection of respiratory rate and heart rate based on acceleration and LSTM according to claim 1, characterized in that: The output gate of the LSTM network maps the hidden state of the last time step of the LSTM into predictions of heart rate and breathing rate.

8. The method for joint detection of respiratory rate and heart rate based on acceleration and LSTM according to claim 1, characterized in that: The method also includes evaluating the prediction results, and the evaluation indicators include mean square error and mean absolute error.

9. The method for joint detection of respiratory rate and heart rate based on acceleration and LSTM according to claim 1, characterized in that: The acceleration signal is obtained through an acceleration sensor.

10. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the program, the method according to any one of claims 1 to 9 is implemented.

Images