A heart rate self-calibration monitoring system and method based on deep learning model

Abstract

The present invention discloses a heart rate self-calibration monitoring system and method based on a deep learning model. In the system, a data acquisition module acquires ECG signals, multi-band PPG signals and acceleration signals; a data processing module filters and amplifies the signals acquired by the data acquisition module; an analog-to-digital conversion module performs analog-to-digital conversion on the output signal of the data processing module; a state detection module classifies the user state based on the acceleration signal; a signal calibration module is used to dynamically adjust the luminous wave band weight of the PPG signal acquisition according to the user state, obtain a weighted optimized multi-band PPG signal, adjust the peak position of the PPG signal according to the R wave position in the ECG signal, and complete the calibration of the PPG signal; a heart rate calculation module predicts the heart rate based on the deep learning model and outputs the heart rate prediction result; and an interactive module transmits the heart rate prediction result. The present invention can provide stable and accurate heart rate monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of heart rate monitoring, and in particular to a heart rate self-calibration monitoring system and method based on a deep learning model. Background Art

[0002] Currently, wearable devices such as smart watches and bracelets have been widely equipped with heart rate monitoring functions. Most of these devices use photoplethysmography (PPG) for pulse rate detection. The PPG method emits green, red or infrared light and detects changes in the absorption and reflection of light by blood to infer the pulsation frequency of the artery. However, although infrared light with a longer wavelength can increase the penetration depth of light, the light intensity of the light source is relatively weak, making the signal more unstable and less obvious when it is acquired.

[0003] Normally, the pulse rate and heart rate have the same frequency, and the PPG method can provide a certain reference value. However, in some special cases, such as heart abnormalities or premature beats, there may be significant differences between the pulse rate and the heart rate. This difference may cause the device to falsely report or miss abnormal heart rate. Pulse changes are the external manifestation of heartbeats, but the pulse rate is not always consistent with the heart rate. For example, in atrial fibrillation, the heart rate may be as high as 120 beats per minute, while the pulse rate is only 90 beats per minute. Traditional PPG-based devices have difficulty detecting such complex heart problems, especially when the pulse is weak or the blood flow changes are not obvious. Coupled with factors such as ambient light interference, exercise status and skin characteristics, the accuracy of PPG heart rate measurement is often not as good as ECG signal detection equipment. Especially under complex physiological conditions, the measurement results are prone to large deviations.

[0004] Existing wearable sports devices can measure heart rate signals through the PPG method, or measure ECG signals by pinching the electrocardiogram of the watch with your hand. However, the PPG method does not directly monitor the heartbeat, and there is a time delay (pulse wave transmission time, PTT) between the R wave of the ECG signal and the pulse wave of the PPG signal, which results in a large error. In addition, the ECG method is more complex in measurement and has higher requirements for the measurement environment. It generally requires the object to be stationary, and continuous monitoring cannot be achieved. Therefore, for wearable devices, there is a certain contradiction between the accuracy and continuity of the heart rate monitoring function. Summary of the invention

[0005] In order to overcome the defects and shortcomings of the prior art, the present invention provides a heart rate self-calibration monitoring system and method based on a deep learning model. The present invention uses ECG signals to assist PPG signal measurement. At the beginning of monitoring, the ECG signal is collected once and combined with the deep learning model for self-calibration. After calibration, the continuous dependence on ECG signals can be eliminated. High-precision heart rate prediction can be achieved only through PPG signals, which ensures the accuracy of heart rate monitoring while taking into account continuity.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The present invention provides a heart rate self-calibration monitoring system based on a deep learning model, comprising: a data acquisition module, a data processing module, an analog-to-digital conversion module, a state detection module, a signal calibration module, a heart rate calculation module and an interaction module;

[0008] The data acquisition module is used to collect ECG signals, multi-band PPG signals and acceleration signals;

[0009] The data processing module is used to filter and amplify the ECG signal, multi-band PPG signal and acceleration signal collected by the data acquisition module;

[0010] The analog-to-digital conversion module is used to perform analog-to-digital conversion on the output signal of the data processing module;

[0011] The state detection module is used to classify the user state based on the acceleration signal;

[0012] The signal calibration module is used to dynamically adjust the luminous wave band weights of the PPG signal acquisition according to the user status, obtain a weighted optimized multi-band PPG signal, extract the peak and trough features in the amplified PPG signal, and the R wave features in the ECG signal, adjust the peak position of the PPG signal according to the R wave position in the ECG signal, and complete the calibration of the PPG signal;

[0013] The heart rate calculation module is used to predict the heart rate based on the deep learning model, input the weighted optimized multi-band PPG signal and the calibrated PPG signal into the deep learning model respectively, splice the feature vectors output by the deep learning model to obtain a feature data training set, obtain the importance score of each feature based on regression training, select features based on contribution and perform bagging integration training, and output the heart rate prediction result;

[0014] The interaction module is used to transmit the heart rate prediction result.

[0015] As a preferred technical solution, the data acquisition module includes an ECG signal acquisition module, a PPG signal acquisition light source, a photoelectric detection module and an acceleration acquisition module. The ECG signal acquisition module is used to acquire ECG signals, the PPG signal acquisition light source adopts a multi-band light source, the photoelectric detection module is used to acquire multi-band PPG signals, and the acceleration acquisition module is used to acquire acceleration signals;

[0016] And / or the photoelectric detection module adopts a multi-band narrowband organic photodetector, which comprises a substrate, a conductive anode, a hole transport layer, an organic photoactive layer, an electron transport layer and a conductive cathode from bottom to top.

[0017] As a preferred technical solution, the data processing module includes a transimpedance amplifier circuit, a buffer isolation module, a bandpass filter, and an AC signal negative feedback amplifier circuit;

[0018] The transimpedance amplifier circuit is used to preliminarily amplify the ECG signal, PPG signal and acceleration signal;

[0019] The buffer isolation module is provided with a source follower and an acceleration signal isolation buffer, wherein the source follower is used to isolate the bandpass filter and the AC signal negative feedback amplifier circuit, and the acceleration signal isolation buffer is used to isolate the acceleration acquisition module and the bandpass filter;

[0020] The bandpass filter is used to filter high-frequency noise and baseline drift of the PPG signal, ECG signal and ACC signal, and isolate the DC component and output the AC component;

[0021] The AC signal negative feedback amplifier circuit is used for secondary amplification of the output AC component signal of the bandpass filter.

[0022] As a preferred technical solution, the state detection module is used to classify the user state based on the acceleration signal, specifically including:

[0023] The state detection module extracts the time domain features of the ACC signal, normalizes the time domain features of the ACC signal, inputs the normalized time domain features into the logistic regression classification model, and predicts the probability of the current motion state through the linear combination of feature weights, which is specifically expressed as:

[0024] ;

[0025] in, P Indicates the probability of the current motion state, which is used to determine the current motion state. represents the sigmoid activation function, represents the weight vector of the logistic regression classification model, is the bias, used to adjust the model output, is the normalized time domain feature.

[0026] As a preferred technical solution, the signal calibration module is used to dynamically adjust the luminous waveband weights of the PPG signal acquisition according to the user status to obtain a weighted optimized multi-band PPG signal, specifically including:

[0027] Set the infrared band weight, green band weight and blue band weight in motion state;

[0028] Set the infrared band weight, green band weight and blue band weight in the resting state;

[0029] Update the weight of the corresponding band according to the signal-to-noise ratio of the current band, expressed as:

[0030] ;

[0031] in, represents the updated weight, is the current weight, is the learning rate of the current band, is the signal-to-noise ratio of the current band, is the average value of the signal-to-noise ratio of all bands;

[0032] Normalize the weights:

[0033] ;

[0034] in, N Indicates the number of weights;

[0035] Based on the normalized weights, weighted optimization multi-band PPG signals are obtained, which can be specifically expressed as:

[0036] ;

[0037] in, represents the weighted optimized multi-band PPG signal, , , Respectively represent the normalized blue band weight, green band weight and infrared band weight, , , They represent the PPG signals corresponding to the blue band, green band and infrared band respectively.

[0038] As a preferred technical solution, the peak position of the PPG signal is adjusted according to the R wave position in the ECG signal to complete the calibration of the PPG signal, which specifically includes:

[0039] Get ECG signal R wave position sequence R z , based on the sliding window method, the peak position sequence of the PPG signal during the systolic period is obtained R x and PPG peak sequence R y ;

[0040] Based on the peak position sequence R x Calculate the average of the intervals between adjacent peak points to get the average pulse interval estimate PI g , estimated by the average pulse interval PI g As the step size to set the search interval, the peak position sequence R x and PPG peak sequence R y The starting point of the alignment is based on the peak position sequence R x and PPG peak sequence R y Find the base point based on the difference in the corresponding value of R p ;

[0041] Select the first R wave position of the ECG signal as the initial base point and find the PPG peak sequence R x Candidate points for valid peak positions in , the candidate points Subtract from the current R wave base point to obtain multiple delayed sequences ;

[0042] Calculate the time interval between adjacent R waves and generate a difference sequence , calculate the Pearson correlation coefficient between the difference series and the delayed series , select Pearson correlation coefficient Maximum delay sequence , and record the corresponding R wave to dictionary;

[0043] The delay sequence and the difference sequence are combined into a dictionary, and the R wave position sequence of the complete ECG signal and the complete PPG signal are gradually input. The optimal candidate point is selected in each iteration until the entire PPG signal is traversed to obtain the corresponding PPG peak estimation position sequence R e ;

[0044] Estimating position sequence based on PPG peak R e For the interpolated PPG peak sequence R p To perform the correction, set an empty sequence, and compare the first peak position values ​​in the two sequences from the starting point of the signal. If the two are close, take the average value of the two and round it up, and store it in the final corrected PPG peak sequence. Rc If the difference is far, the distance between the second point and the first point in the two sequences is determined respectively, and the point with the closer distance is used as the first base point to continue searching backwards, and finally the PPG peak sequence after ECG signal correction is obtained. R c .

[0045] As a preferred technical solution, the deep learning model includes an initial convolution module, a deep feature extraction module, a time series modeling module and a feature fusion and optimization module;

[0046] The initial convolution module performs feature extraction based on the convolution layer. The extracted feature map is subjected to the global maximum pooling operation to obtain the reduced-dimensional features, and then normalized to obtain the feature ;

[0047] The deep feature extraction module extracts the features Extracting deep features through grouped convolution ;

[0048] The time series modeling module captures the temporal dynamic characteristics of the signal through the projection layer, constructs long-term and short-term dependencies, processes each time step through the GRU unit, and outputs the time series features. ;

[0049] The feature fusion and optimization module realizes the feature mapping between channels through the projection layer 1×1 convolution to obtain the fusion feature after dimension reduction. , expressed as:

[0050] ;

[0051] Output through the dual-channel pooling layer:

[0052] ;

[0053] in, represents the maximum pooling layer, represents the average pooling layer, Indicates preset weight;

[0054] Generate heart rate prediction results through a fully connected layer.

[0055] As a preferred technical solution, the projection layer and the dual-channel pooling layer are respectively inserted into the ReLU activation layer, and an L2 norm normalization layer is added, and the fusion features are processed by the ReLU activation layer and the L2 norm normalization layer.

[0056] As a preferred technical solution, the feature vectors output by the deep learning model are spliced ​​to obtain a feature data training set, the importance score of each feature is obtained based on regression training, the features are screened based on contribution and bagging ensemble training is performed, and the heart rate prediction results are output, specifically including:

[0057] Filter features based on contribution and obtain filtered features , based on the corresponding labels Constructing a training set ;

[0058] For the training set The number of features and the number of data are randomly sampled according to the set ratio, and input into multiple deep learning models for bagging ensemble training. The outputs of multiple deep learning models are weighted averaged to obtain the final heart rate prediction result.

[0059] The present invention also provides a heart rate self-calibration monitoring method based on a deep learning model, which is provided with the above-mentioned heart rate self-calibration monitoring system based on a deep learning model, and comprises the following steps:

[0060] Collect ECG signals, multi-band PPG signals and acceleration signals, classify the user status based on the acceleration signal, re-collect the ECG signal when a change in the user status is detected, and calibrate the PPG signal;

[0061] The collected ECG signals, multi-band PPG signals and acceleration signals are filtered and amplified, and the signals are converted into analog-to-digital signals;

[0062] Dynamically adjust the light wave band weights of the PPG signal collection according to the user status to obtain a weighted optimized multi-band PPG signal, extract the peak and trough features in the amplified PPG signal, and the R wave features in the ECG signal, adjust the peak position of the PPG signal according to the R wave position in the ECG signal, and complete the calibration of the PPG signal;

[0063] Heart rate prediction is performed based on the deep learning model. The weighted optimized multi-band PPG signal and the calibrated PPG signal are respectively input into the deep learning model. The feature vectors output by the deep learning model are concatenated to obtain the feature data training set. The importance score of each feature is obtained based on regression training. The features are screened based on the contribution and bagging ensemble training is performed to output the heart rate prediction results.

[0064] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0065] (1) The present invention adopts a photoelectric detection module to realize the acquisition of multi-channel PPG signals. Combining the PPG multi-channel features, the dynamic features after ECG correction, and the motion state features detected by the state detection module, it provides a rich and accurate data basis for the construction of the deep learning model. The use of the photoelectric detection module enhances the sensitivity and specificity of signals of different wavelengths, making the extraction of heart rate signals more accurate. At the same time, the integration of dynamic features and the motion features provided by the state detection module help the model adapt to the user's changeable physiological state, thereby further improving the accuracy and stability of heart rate monitoring in training and prediction, and significantly reducing the interference of irrelevant signals and motion artifacts on the measurement results.

[0066] (2) The photoelectric detection module of the present invention effectively reduces the influence of external ambient light interference. By using different wavelength bands (380nm, 520nm and 880nm) in coordination, it increases the sensitivity of the device to blood absorption signals of specific wavelengths, improves the signal-to-noise ratio, and effectively solves the contradiction between signal penetration depth and light intensity, thereby achieving high-quality heart rate monitoring. Even under complex conditions, such as strong light, low light or when the user is exercising, it can still provide stable and accurate heart rate monitoring.

[0067] (3) The state detection module of the present invention can monitor the user's current physiological state (such as exercise or rest) in real time, automatically re-collect the ECG signal for correction when the user's state changes, maintain the timing consistency of the PPG and ECG signals, and select corresponding data for analysis under different physiological activities, thereby improving the accuracy and applicability of heart rate monitoring.

[0068] (4) The present invention adopts projection layer and dual-channel pooling layer to replace the traditional fully connected layer, and increases the diversity of features and the robustness of the model by combining weighted pooling with ReLU activation layer. At the same time, the anti-interference ability of input changes is enhanced by L2 normalization processing. In terms of feature fusion and regression model integration, the feature vectors and dynamic features generated by deep neural networks are combined, and deep learning models are used for feature fusion and model integration. The whole process integrates multiple deep learning models for bagging training, uses feature importance indicators to screen out the most influential feature vectors, and achieves high-precision monitoring of heart rate changes through multi-stage training. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Picture 1 It is a schematic diagram of the overall architecture of the heart rate self-calibration monitoring system based on the deep learning model of the present invention;

[0070] Picture 2 This is a schematic diagram of the external quantum efficiency data of the photoelectric detection module of the present invention;

[0071] Picture 3It is a schematic diagram of the device structure of the multi-band narrowband organic photodetector of the present invention. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0073] Example 1

[0074] like Picture 1 As shown, this embodiment provides a heart rate self-calibration monitoring system based on a deep learning model, including: a data acquisition module, a data processing module, an analog-to-digital conversion module, a state detection module, a signal calibration module, a heart rate calculation module, and an interaction module;

[0075] In this embodiment, the data acquisition module includes: an ECG signal acquisition module, a PPG signal acquisition light source, a photoelectric detection module, and an acceleration acquisition module;

[0076] Among them, the ECG signal acquisition module is used to collect ECG signals, which can be integrated into a wrist device or collected through additional electrodes. The ECG signal is attached to the user's wrist or chest through the electrode, which can accurately record the electrical activity of the heart. The ECG signal acquisition module can specifically use an ECG single-lead electrode; the PPG signal acquisition light source includes an LED light source or LED matrix that can emit 380nm (blue light), 520nm (green light) and 880nm (near-infrared light), which is modulated by a microprocessor to flash alternately; Picture 2 As shown in the figure, the response band of the photoelectric detection module corresponds to the luminous wave band of the PPG signal acquisition light source one by one. It is a multi-band narrow-band organic photodetector with a response spectrum peak of about 380nm, 520nm, and 880nm and a half-peak width of about 50nm. It can receive the PPG light signal emitted by the PPG signal acquisition light source and transmitted through the human finger through multiple channels, and output and transmit the pulse blood flow information of the human body in the form of electrical signals; the acceleration acquisition module is equipped with a three-axis acceleration sensor for capturing the user's dynamic data, which can obtain the acceleration ACC signal in the x-axis, y-axis and z-axis directions, and can collect data at the configured sampling frequency (50Hz), thereby ensuring the temporal resolution of the data. By monitoring the acceleration of each axis, the user's motion state can be accurately identified, thereby assisting in state recognition and artifact elimination during heart rate measurement, and serving as an important input for algorithm analysis and scene recognition in the system.

[0077] In this embodiment, the multi-band narrowband organic photodetector only needs an external filter to achieve three-band selective detection with response peaks at 380nm, 520nm, and 880nm to filter out ambient light noise interference and further improve the accuracy of heart rate detection results. Its multi-channel acquisition enables the system to process multi-wavelength signals in parallel, which not only improves the stability of the signal under different physiological conditions, but also effectively balances the penetration depth and intensity of the light source, enabling the system to obtain information from multiple wavelengths, significantly improving the accuracy and robustness of heart rate measurement, and overcoming the limitations of single wavelength acquisition. Therefore, by using blue light, green light, and infrared light at the same time, more comprehensive pulse blood flow information can be provided.

[0078] like Picture 3 As shown, the structure of the multi-band narrowband organic photodetector is specifically as follows: from bottom to top, it is a substrate, a conductive anode, a hole transport layer, an organic photoactive layer, an electron transport layer and a conductive cathode;

[0079] Wherein, the substrate is a transparent or translucent polymer material flexible substrate or a quartz glass rigid substrate;

[0080] The conductive anode is any one of materials with similar functions such as fluorine-doped tin dioxide, graphene, carbon nanotubes, metal nanowires, indium tin oxide, etc., which serves as an incident light window and has a semi-transparent property. Preferably, the conductive anode is indium tin oxide (ITO);

[0081] The hole transport layer is any one of a small molecule material, an organic P-type polymer material, a metal oxide material and a material with similar functions. The small molecule material is 6,13-bis(triisopropylsilylethynyl)pentacene, polyvinyldimethylphenazine, etc. The organic P-type polymer material is poly(4-butyltriphenylamine) (poly-TPD), 3,4-ethylenedioxythiophene mixed polystyrene sulfonate, etc. The metal oxide material is molybdenum oxide, nickel oxide, etc. Preferably, the hole transport layer is PEDOT:PSS;

[0082] The organic photoactive layer is formed by mixing N-type materials and P-type materials, and is distributed in a vertical gradient, wherein the P-type material is formed by layering or mixing one or more organic materials with electron-donating properties, and the N-type layer is composed of organic materials with electron-withdrawing properties and is distributed in the middle or above the P-type layer. The total thickness of the organic material mixture is 500~1500nm. The P-type material is an organic conjugated polymer or conjugated small molecule material with an electron donating unit, one or more of PTB7 (PCE9), PTB7-Th (PCE10), PBDB-T (PCE12), PBDB-T-2F (PM6), PBDB-T-2Cl (PM7), D18 or NT812, and the N-type material is a non-fullerene electron acceptor material, such as one or more of BTP-4F (Y6), Y6-BO, BTP-4Cl (Y7), ITIC, ITIC-Th, IEICO, IEICO-4F, IEICO-4Cl, COTIC-4F, COTIC-4Cl, etc.

[0083] The electron transport layer is composed of a material having an electron transport function, which is any one of a low work function metal material, a water-alcohol soluble material, and a material having similar functions. The water-alcohol soluble material is brominated-[9,9-dioctylfluorene-9,9-bis(N,N-dimethylaminopropyl)fluorene] (PFN-Br), polyethoxyethyleneimine (PEIE), derivatives of naphthalene diimide (NDI), and perylene diimide (PDI), etc. Preferably, the electron transport layer is brominated-[9,9-dioctylfluorene-9,9-bis(N,N-dimethylaminopropyl)fluorene] (PFN-Br).

[0084] The conductive cathode is a metal electrode selected from one or more of lithium (Li), magnesium (Mg), calcium (Ca), strontium (Sr), barium (Ba), aluminum (Al), copper (Cu), gold (Au), silver (Ag) or indium (In), or materials with similar functions, with a thickness of 50 to 150 nm. Preferably, the conductive cathode is aluminum (Al) with a thickness of about 80 nm.

[0085] In this embodiment, the data processing module includes: a transimpedance amplifier circuit, a buffer isolation module, a bandpass filter, and an alternating current (AC) signal negative feedback amplifier circuit;

[0086] Among them, the transimpedance amplifier circuit is used to convert the current signal output by the photoelectric detection module, the electrocardiogram signal acquisition module, and the acceleration acquisition module into an analog voltage signal and amplify it, so as to restore the PPG signal, the ECG signal, and the ACC signal; the buffer isolation module is provided with a source follower, which is used to isolate the bandpass filter and the alternating current (AC) signal negative feedback amplifier circuit. The buffer isolation module is also provided with an acceleration signal isolation buffer, which is specifically used to isolate the acceleration sensor and the bandpass filter, so as to ensure that the signal output by the acceleration sensor is not affected by the interference and load of other signal processing units during the transmission process, so as to ensure the stability of signal transmission and avoid distortion and attenuation; the bandpass filter is used to filter the high-frequency noise and baseline drift of the PPG signal, the ECG signal, and the ACC signal, and at the same time isolate the DC component and output the AC component. The AC signal negative feedback amplifier circuit performs secondary amplification on the output AC component signal of the bandpass filter;

[0087] In this embodiment, a Butterworth bandpass filter with a frequency range of 0.5-10 Hz is used to remove high-frequency noise and low-frequency baseline drift in the PPG signal, a Butterworth bandpass filter with a frequency range of 0.5-35 Hz is used to remove high-frequency electromyographic noise and baseline drift in the ECG signal, and a bandpass filter with a frequency range of 0.5 Hz-20 Hz is used to filter the ACC signal.

[0088] In this embodiment, the analog-to-digital conversion module converts the PPG signal, ECG signal and ACC signal processed by the data processing module from analog voltage signal form to digital signal form.

[0089] In this embodiment, the state detection module is used to monitor the user's dynamic state in real time. By classifying the user's current state (such as resting, exercising, etc.), the consistency of subsequent data analysis and the user's actual state is ensured. The binary classification of the user's dynamic state is mainly completed based on the time domain features extracted from the acceleration ACC signal, and the classification results are fed back to the signal calibration module for adjusting the weighted ratio of the multi-channel PPG signal.

[0090] The state detection module first extracts the time domain features of the ACC signal. The ACC signal feature extraction includes mean, standard deviation, maximum and minimum values, peak interval, amplitude and frequency; these features reflect the intensity, fluctuation range, frequency and rhythm of the movement respectively;

[0091] This embodiment uses zero crossing rate (ZCR) to extract frequency features. The zero crossing rate refers to the frequency at which the acceleration signal changes from positive to negative or from negative to positive. ZCR can be used to determine the activity intensity of the user. Frequent zero crossings usually correspond to more intense exercise, which is specifically expressed as:

[0092] ;

[0093] in, is each sampling point of the acceleration signal, is a sign function, which indicates the sign of the signal.

[0094] Then the time domain features of the ACC signal are normalized and the feature values ​​are mapped to the interval [0, 1] to ensure that features of different dimensions can be uniformly processed in the classification model. The normalized time domain features are input into the logistic regression classification model, which predicts the probability of the current motion state through the linear combination of feature weights. The model uses the logistic regression algorithm, and its output formula is:

[0095] ;

[0096] in, is the sigmoid activation function, which is used to map the result of linear combination into probability value; is the weight vector of the logistic regression classification model, determined by the training process; is the bias, used to adjust the model output; is the normalized time domain feature.

[0097] The probability value output by the logistic regression model is used for state determination. When the probability value is greater than or equal to 0.5, the state is determined to be "movement"; when the probability value is less than 0.5, the state is determined to be "resting". The classification result is used as the output of the state detection module and fed back to the signal calibration module.

[0098] Through this process, the state detection module not only provides key dynamic feedback for the weighting of the PPG signal, but also ensures that the system processes the heart rate measurement signal more accurately and robustly under different user states.

[0099] In dynamic monitoring systems, it is particularly important to adopt a multi-band design, which can enhance the capture of physiological characteristics through signals of different wavelengths, offset the contradictions of each band in terms of penetration depth and signal strength, and improve the stability and accuracy of the overall measurement. However, due to different physiological states and environmental conditions, the quality of PPG signals in different bands for heart rate monitoring is different, that is, the user's exercise state will affect the signal-to-noise ratio (SNR) of each band. Therefore, the signal calibration module dynamically adjusts the weight according to the user's state. The blue band (380nm) has a relatively unstable signal performance in a dynamic environment, so its weight is set lower. The green band (520nm) is stable in a resting state and is very suitable for heart rate monitoring, so it is given a higher weight. The infrared band (880nm) is more effective in capturing pulse changes in a moving state, so its weight is increased accordingly in a moving state. This dynamic weight setting strategy helps to optimize the signal processing effect under different physiological states and environmental conditions, thereby improving the accuracy and reliability of monitoring data.

[0100] In motion, the infrared band weight is set to:

[0101] ;

[0102] When the SNR is greater than 5 dB, the signal quality is good, so the focus is retained on the infrared signal; when the SNR is between 2dB and 5 dB, the signal quality is medium and the weight is increased to 0.15; when the SNR is less than 2 dB, the signal quality is poor and the weight is set to 0.2 to ensure that the heart rate features can still be captured in low-quality signals.

[0103] In motion, the green band weights are set to:

[0104] ;

[0105] When the SNR is greater than 4 dB, a higher weight of 0.3 is given to ensure its stability; when the SNR is between 2dB and 4 dB, a weight of 0.25 is given; and when the signal becomes unreliable, the weight is set to 0.2 to maintain a certain sensitivity. The blue band weight is set to , due to its generally inferior quality to the green and red bands and its purpose being to complement them, it is given a fixed weight.

[0106] In the resting state, the weights of the green band are set to:

[0107] ;

[0108] If it is below 2 dB, the weight is set to 0.2 to account for potential degradation in signal quality. The weight for the infrared band is set to , does not change with SNR, ensuring that a certain heart rate signal can still be obtained when there is less exercise. The blue band also maintains a fixed weight of 0.1.

[0109] The weight update formula is:

[0110] ;

[0111] in, represents the updated weight, is the current weight, is the learning rate of the current band, is the signal-to-noise ratio of the current band, is the average of the signal-to-noise ratios of all bands. The formula fine-tunes the weights according to the SNR of the current signal. The higher the signal-to-noise ratio, the more the weight increases, aiming to compensate for bands with insufficient signal quality.

[0112] The weight normalization formula is:

[0113] ;

[0114] Keep the updated weights normalized to ensure that the sum is 1 to avoid excessive influence of a certain band on the result. The signal fusion formula is:

[0115] ;

[0116] in, represents the weighted optimized multi-band PPG signal, , , Respectively represent the normalized blue band weight, green band weight and infrared band weight, , , They represent the PPG signals corresponding to the blue band, green band and infrared band respectively.

[0117] By weighting the denoised signals of different bands, all information is combined to generate a comprehensive signal, improving the accuracy and stability of heart rate monitoring.

[0118] Through these dynamic weight adjustments and appropriate denoising, the system is able to optimize its heart rate monitoring capabilities under different user states to cope with complex physiological conditions. These measures work together to improve the accuracy and stability of heart rate measurement.

[0119] In this embodiment, the signal calibration module first adjusts the weighted ratio of the multi-channel PPG signal according to the classification result of the motion scene by the state detection module, and then extracts the peak and trough features in the multi-band PPG signal after secondary amplification, as well as the R wave features in the ECG signal, and completes the calibration of the PPG signal in combination with the time synchronization information based on the R wave position, and adjusts the peak position of the PPG signal according to the R wave position in the ECG signal to achieve time alignment, completes the calibration of the PPG signal, and uses the calibrated PPG signal as a dynamic feature data set;

[0120] Specifically, the signal-to-noise ratio (SNR) indicator is used to evaluate the ECG signal quality, and signal segments with good quality and a length of not less than 10 seconds are selected for correction; if multiple R wave templates cannot be matched continuously, the missing R waves will be filled in by the QRS detector to obtain a complete 10-second ECG signal R wave position sequence. R z ;

[0121] At the same time, the corresponding PPG signal is processed and the sliding window method is used to obtain the peak position sequence of the PPG systolic period. Rx and PPG peak sequence R y ; In this process, the missing peaks may appear in the PPG signal, so for the PPG peak sequence R y Interpolation is performed to supplement missing values ​​and ensure the consistency of the PPG signal and the ECG signal on the time axis. By identifying the R wave position in the ECG signal and the contraction peak position of the PPG signal, the signal alignment and interpolation strategy is used to correct the position of the PPG peak. This interpolation operation will rely on the obtained ECG R wave position information sequence. R z , will use R z The position information of the R wave is used to determine the interpolation point, that is, to insert the R y The missing peak in .

[0122] In this embodiment, the peak position sequence of the systolic period is used. R x Calculate the average value of the intervals between adjacent peak points and obtain the estimated value of the average pulse interval, recorded as PI g When the interpolation operation starts, the detection starting point will be selected and the peak position sequence will be R x and PPG peak sequence R y Align the starting point of the signal and search for the base point forward. If the peak position sequence R x and PPG peak sequence R y The corresponding values ​​in the two sequences (such as x i and y i ) differ by less than 10 sampling points, the subscript is The point is regarded as the base point and stored in the interpolation peak position sequence being constructed. R p Based on this, an estimate of the pulse interval is used PI g As the step size, set the interval to 0.8 PI g To 1.2 PI g , continue to search backward along the signal. In this process, if or If it falls within this interval, the point is considered to be the peak position and continues to be stored in the interpolation peak position sequence being constructed. R p, if both are in the interval, the average of the two is taken as the peak position. Finally, after the above process, the interpolated peak position sequence is obtained. R p .

[0123] Considering the large correlation between the systolic peak in the PPG signal and the R wave in the ECG signal, the fixed delay relationship based on the propagation velocity can be further calculated:

[0124] Condition 1: Select the first R wave position of ECG as the initial base point. If it is within the calculated range (0.9 PI g To 1.1 PI g ) has PPG peak sequence R x If a point falls into this interval, it is considered to be a valid peak position. .

[0125] Condition 2: When there are multiple points that meet the conditions, the system will calculate the time interval between adjacent R waves and generate a difference sequence By respectively dividing the PPG peak sequence R x Subtract the candidate points in the R wave from the current base point to obtain multiple delay sequences , and calculate the Pearson correlation coefficient between the difference series and the delayed series .choose The largest And record the corresponding R wave In the dictionary.

[0126] Furthermore, after finding the delay sequence that leads to the maximum correlation coefficient, the corresponding PPG peak is stored in the final position, the PPG peak candidate point is repeatedly searched, and the correlation between the candidate sequence and the difference sequence is analyzed to determine the final stored sequence. Finally, the delay sequence and the difference sequence are combined into a dictionary, and the R wave position sequence of the complete ECG signal and the complete PPG signal are gradually input, that is, the new base point is continuously used as the starting point, and the judgment conditions 1 and 2 are repeatedly judged. The optimal candidate point selected in each iteration is obtained until the entire PPG signal is traversed to obtain the corresponding PPG peak estimated position sequence. R e .

[0127] Finally, the position sequence is estimated based on the PPG peak R e For the interpolated PPG peak sequence R pSet up an empty sequence, and compare the first peak position values ​​in the two sequences from the start point of the signal. If the two are close, take the average of the two and round them up. , round it and store it in the final corrected PPG peak sequence R c middle.

[0128] If the difference is far, the distance between the second point and the first point in the two sequences is determined respectively. The point with the closer distance is used as the first base point, and the search starts backwards. This point is stored in the empty sequence as the peak position. The interval size is set to 0.8. PI g ~1.2 PI g If a point in the two sequences falls within the interval in this area, the point is considered to be the next peak position and stored in the sequence. If the points in the two sequences are within the interval, the value of the two points closest to the first base point is the next peak. Continue to search backwards and repeat the above operation. Finally, a complete PPG peak sequence after ECG signal correction will be obtained. R c , effectively reducing the impact of noise.

[0129] In this embodiment, the heart rate calculation module constructs a deep learning model for high-precision heart rate monitoring based on the weighted optimized multi-band PPG signal (original PPG multi-channel data set) and the corrected PPG signal (dynamic feature data set);

[0130] Specifically, the ResNeXt module based on the grouped convolution technology constitutes the main architecture of the deep learning model. It can ensure the regression accuracy while improving the calculation speed and the real-time performance of heart rate monitoring under the condition of limited data training. The ResNeXt module is designed as three units (Unit1, Unit2, Unit3), which output 128, 256 and 512 feature channels respectively. Compared with traditional convolution operations, grouped convolution can reduce the amount of calculation and the number of parameters. Assume that the number of channels of the input data is C, the number of channels of the output data is M, and the size of the convolution kernel is K × K , the grouped convolution divides the input data and convolution kernel into G groups, so the number of parameters required for grouped convolution is: , the first item is the number of parameters required for each group, and the second item is the number of bias parameters required for each output channel, while the number of parameters required for the traditional convolution operation is: ,This significant reduction not only reduces the memory consumption but also improves the ,computational efficiency, which meets the high requirements of real-time ,and accuracy of heart rate monitoring.

[0131] In the optimized design of the fully connected layer in the feature fusion part, the deep learning model decomposes the traditional fully connected layer into a projection layer and a dual-channel pooling layer to achieve the fusion and enhancement of high-dimensional features.

[0132] Furthermore, in order to solve the potential nonlinear problem, ReLU activation layers are inserted in the projection layer and the dual-channel pooling layer, and an L2 norm normalization layer is added at the end to ensure that the feature vectors obtained by the two pooling layers are normalized before being transmitted downward. This process can improve the model's robustness to input changes and reduce the risk of overfitting. After using this transformation, the number of parameters is reduced from Reduce to , which helps to improve the training speed and real-time inference capability of the model. The high-dimensional features after feature fusion are further processed by the ReLU activation layer and the L2 normalization layer, which improves the robustness of the model to input changes and reduces the risk of overfitting.

[0133] In this embodiment, the PPG signal before weighted correction is input. , weighted and corrected PPG signal , each signal is a one-dimensional time series data , each time step corresponds to a signal value.

[0134] The following operations are consistent for both signals:

[0135] (1) Initial convolution module

[0136] 1. The input signal is subjected to feature extraction using a 3×3 convolution kernel, and the feature map is obtained. , preserving the local characteristics of the signal.

[0137] 2. Reduce the dimension of the feature map and extract global key features. The dimension reduction is performed through the global maximum pooling operation to obtain the reduced dimension features:

[0138] ;

[0139] 3. Improve training stability and accelerate convergence through normalization, and normalize the features of each batch:

[0140] ;

[0141] (2) Deep feature extraction module

[0142] 1. Extract deep features through group convolution and reduce the number of parameters. After processing by 3 ResNeXt blocks, the output feature dimensions of each block are 128, 256, and 512 respectively:

[0143] ;

[0144] (3) Timing Modeling Module

[0145] 1. The projection layer captures the temporal dynamic characteristics of the signal, builds long-term and short-term dependencies, and processes each time step through the GRU unit:

[0146] ;

[0147] in, is the input of the current time step, is the hidden state of the previous time step, outputting the time series features ;

[0148] (4) Feature fusion and optimization module

[0149] 1. The projection layer uses 1×1 convolution to achieve feature mapping between channels and obtain fusion features after dimensionality reduction. :

[0150] ;

[0151] 2. The dual-channel layer combines maximum pooling and weighted average pooling to dynamically balance information density and dimensionality reduction efficiency:

[0152] ;

[0153] In the output stage, the model generates the final heart rate prediction result through the fully connected layer (FC1). Combining the output of the ResNeXt module, GRU layer and feature fusion layer, the model can make full use of the multi-dimensional information and temporal dynamic characteristics of the PPG signal to provide high-precision heart rate prediction.

[0154] In terms of regression model integration based on XGBoost, the models trained in the previous step for the original PPG multi-channel dataset and the dynamic feature dataset (a periodic cut dynamic signal set generated by accurately aligning the PPG signal and the ECG signal) are used for feature fusion to obtain the final prediction result. Specifically, after the original PPG multi-channel dataset is input into the deep learning model, the high-dimensional feature vector output by the last fully connected layer is obtained. ; After the dynamic feature data set is input into the deep learning model, the high-dimensional feature vector output by the last fully connected layer is obtained ;

[0155] Concatenate these two sets of feature vectors , the regression model feature representation of the data can be formed. Then, the feature representation obtained from all training data is constructed into a feature data training set, and regression training is performed through XGBoost to obtain the initial prediction model.

[0156] Considering that the model extracts a large number of feature vectors, the feature importance index provided by XGBoost is used to screen all feature vectors. After the initial prediction model training, the importance of each feature vector can be obtained. The feature vectors ranked from high to low are deleted, and only the top 80% (denoted as ). This example uses the initial XGBoost model training and the corresponding labels , get the importance score of each feature, and measure it by the contribution of the feature to the final formation of the complete tree, that is, the relative importance of each feature to the improvement of model performance; after feature screening, these feature vectors with higher contribution are input into the XGBoost model, and the filtered features Used to reconstruct the optimized training set Thus, the final prediction model is obtained.

[0157] The next step is model fusion, which is to input the above filtered feature vectors into 5 XGBoost models for bagging ensemble training. Five XGBoost models are trained separately by the above method. These five models are independent of each other. When training each model, the number of features and the number of data are randomly sampled in a certain ratio. Train 5 independent XGBoost models separately. After training, input the test data ( and ) The weighted average of the heart rate prediction results of the five models is used to obtain the final output result.

[0158] In this embodiment, projection layers and dual-channel pooling layers are used to replace traditional fully connected layers, and weighted pooling is combined with ReLU activation layers to increase feature diversity and model robustness. At the same time, L2 normalization is used to enhance the ability to resist input changes. In terms of feature fusion and regression model integration, feature vectors and dynamic features generated by deep neural networks are combined, and XGBoost is used for feature fusion and model integration. The entire process integrates multiple XGBoost models for bagging training, uses feature importance indicators to screen out the most influential feature vectors, and achieves high-precision monitoring of heart rate changes through multi-stage training.

[0159] In this embodiment, the interactive module is used to monitor the system and interact with the user, and to display or transmit the heart rate prediction results. The interactive module can select a Bluetooth transmitting module, a Bluetooth receiving module or a WiFi module, and a display module, and can upload the heart rate prediction results to the cloud to establish a personal database or share data.

[0160] The present invention collects and screens multi-channel PPG signals based on a multi-band detector. The multi-band narrowband detector uses light sources of different wavelengths to effectively utilize the relationship between the penetration depth and intensity of light, avoiding unstable signals that may be generated under a single wavelength. By optimizing the combined application of light sources, it can provide more stable and accurate heart rate monitoring results under various physiological conditions. The state detection module can accurately identify the user's motion state, and the extracted related motion features are input into the deep learning model to perform deep learning model training, so that the heart rate measurement results based on PPG gradually approach the accuracy of standard equipment. In particular, when the state changes, the ECG signal will be collected again for correction, thereby ensuring the continuity and accuracy of heart rate monitoring. In addition, by referring to the R wave position of the ECG signal, the unrecognized contraction peak in the PPG signal is corrected, thereby achieving more accurate and continuous heart rate monitoring, overcoming the defect of low measurement accuracy of existing photoelectric plethysmography equipment under complex physiological conditions. The present invention can not only improve the measurement accuracy through multi-band optical signals, but also perform self-calibration by referring to the data of medical equipment, thereby significantly reducing the impact of environmental interference and changes in physiological conditions on the measurement results, and improving the heart rate monitoring ability of the equipment under complex conditions.

[0161] Example 2

[0162] This embodiment provides a heart rate self-calibration monitoring method based on a deep learning model, which is provided with a heart rate self-calibration monitoring system based on a deep learning model of Embodiment 1, and includes the following steps:

[0163] Collect ECG signals, multi-band PPG signals and acceleration signals, classify the user status based on the acceleration signal, re-collect the ECG signal when a change in the user status is detected, and calibrate the PPG signal;

[0164] The collected ECG signals, multi-band PPG signals and acceleration signals are filtered and amplified, and the signals are converted into analog-to-digital signals;

[0165] Dynamically adjust the light wave band weights of the PPG signal collection according to the user status to obtain a weighted optimized multi-band PPG signal, extract the peak and trough features in the amplified PPG signal, and the R wave features in the ECG signal, adjust the peak position of the PPG signal according to the R wave position in the ECG signal, and complete the calibration of the PPG signal;

[0166] Heart rate prediction is performed based on the deep learning model. The weighted optimized multi-band PPG signal and the calibrated PPG signal are respectively input into the deep learning model. The feature vectors output by the deep learning model are concatenated to obtain the feature data training set. The importance score of each feature is obtained based on regression training. The features are screened based on the contribution and bagging ensemble training is performed to output the heart rate prediction results.

[0167] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A heart rate self-calibration monitoring system based on a deep learning model, characterized in that: include: Data acquisition module, data processing module, analog-to-digital conversion module, state detection module, signal calibration module, heart rate calculation module and interaction module; The data acquisition module is used to collect ECG signals, multi-band PPG signals and acceleration signals; The data processing module is used to filter and amplify the ECG signal, multi-band PPG signal and acceleration signal collected by the data acquisition module; The analog-to-digital conversion module is used to perform analog-to-digital conversion on the output signal of the data processing module; The state detection module is used to classify the user state based on the acceleration signal; The signal calibration module is used to dynamically adjust the luminous wave band weights of the PPG signal acquisition according to the user status, obtain a weighted optimized multi-band PPG signal, extract the peak and trough features in the amplified PPG signal, and the R wave features in the ECG signal, adjust the peak position of the PPG signal according to the R wave position in the ECG signal, and complete the calibration of the PPG signal; The signal calibration module is used to dynamically adjust the luminous waveband weights of the PPG signal acquisition according to the user status to obtain a weighted optimized multi-band PPG signal, specifically including: Set the infrared band weight, green band weight and blue band weight in motion state; Set the infrared band weight, green band weight and blue band weight in the resting state; Update the weight of the corresponding band according to the signal-to-noise ratio of the current band, expressed as: ； in, represents the updated weight, is the current weight, is the learning rate of the current band, is the signal-to-noise ratio of the current band, is the average value of the signal-to-noise ratio of all bands; Normalize the weights: ； in, N Indicates the number of weights; Based on the normalized weights, weighted optimization multi-band PPG signals are obtained, which can be specifically expressed as: ； in, represents the weighted optimized multi-band PPG signal, , , Respectively represent the normalized blue band weight, green band weight and infrared band weight, , , Respectively represent the PPG signals corresponding to the blue band, green band and infrared band; The heart rate calculation module is used to predict the heart rate based on the deep learning model, input the weighted optimized multi-band PPG signal and the calibrated PPG signal into the deep learning model respectively, splice the feature vectors output by the deep learning model to obtain a feature data training set, obtain the importance score of each feature based on regression training, select features based on contribution and perform bagging integration training, and output the heart rate prediction result; The interaction module is used to transmit the heart rate prediction result.

2. The heart rate self-calibration monitoring system based on deep learning model according to claim 1, characterized in that: The data acquisition module includes an ECG signal acquisition module, a PPG signal acquisition light source, a photoelectric detection module and an acceleration acquisition module. The ECG signal acquisition module is used to acquire ECG signals. The PPG signal acquisition light source adopts a multi-band light source. The photoelectric detection module is used to acquire multi-band PPG signals. The acceleration acquisition module is used to acquire acceleration signals. And / or the photoelectric detection module adopts a multi-band narrowband organic photodetector, which comprises a substrate, a conductive anode, a hole transport layer, an organic photoactive layer, an electron transport layer and a conductive cathode from bottom to top.

3. The heart rate self-calibration monitoring system based on deep learning model according to claim 2, characterized in that: The data processing module includes a transimpedance amplifier circuit, a buffer isolation module, a bandpass filter, and an AC signal negative feedback amplifier circuit; The transimpedance amplifier circuit is used to preliminarily amplify the ECG signal, PPG signal and acceleration signal; The buffer isolation module is provided with a source follower and an acceleration signal isolation buffer, wherein the source follower is used to isolate the bandpass filter and the AC signal negative feedback amplifier circuit, and the acceleration signal isolation buffer is used to isolate the acceleration acquisition module and the bandpass filter; The bandpass filter is used to filter high-frequency noise and baseline drift of the PPG signal, ECG signal and ACC signal, and isolate the DC component and output the AC component; The AC signal negative feedback amplifier circuit is used for secondary amplification of the output AC component signal of the bandpass filter.

4. The heart rate self-calibration monitoring system based on deep learning model according to claim 1, characterized in that: The state detection module is used to classify the user state based on the acceleration signal, specifically including: The state detection module extracts the time domain features of the ACC signal, normalizes the time domain features of the ACC signal, inputs the normalized time domain features into the logistic regression classification model, and predicts the probability of the current motion state through the linear combination of feature weights, which is specifically expressed as: ； in, P Indicates the probability of the current motion state, which is used to determine the current motion state. represents the sigmoid activation function, represents the weight vector of the logistic regression classification model, is the bias, used to adjust the model output, is the normalized time domain feature.

5. The heart rate self-calibration monitoring system based on deep learning model according to claim 1, characterized in that: The step of adjusting the peak position of the PPG signal according to the R wave position in the ECG signal to complete the calibration of the PPG signal specifically includes: Get ECG signal R wave position sequence R z , based on the sliding window method, the peak position sequence of the PPG signal during the systolic period is obtained R x and PPG peak sequence R y ; Based on the peak position sequence R x Calculate the average of the intervals between adjacent peak points to get the average pulse interval estimate PI g , estimated by the average pulse interval PI g As the step size to set the search interval, the peak position sequence R x and PPG peak sequence R y The starting point of the alignment is based on the peak position sequence R x and PPG peak sequence R y Find the base point based on the difference in the corresponding value of R p ; Select the first R wave position of the ECG signal as the initial base point and find the PPG peak sequence R x Candidate points for valid peak positions in , the candidate points Subtract from the current R wave base point to obtain multiple delayed sequences ; Calculate the time interval between adjacent R waves and generate a difference sequence , calculate the Pearson correlation coefficient between the difference series and the delayed series , select Pearson correlation coefficient Maximum delay sequence , and record the corresponding R wave to dictionary; The delay sequence and the difference sequence are combined into a dictionary, and the R wave position sequence of the complete ECG signal and the complete PPG signal are gradually input. The optimal candidate point is selected in each iteration until the entire PPG signal is traversed to obtain the corresponding PPG peak estimation position sequence R e ; Estimating position sequence based on PPG peak R e For the interpolated PPG peak sequence R p To perform the correction, set an empty sequence, and compare the first peak position values ​​in the two sequences from the starting point of the signal. If the two are close, take the average value of the two and round it up, and store it in the final corrected PPG peak sequence. R c If the difference is far, the distance between the second point and the first point in the two sequences is determined respectively, and the point with the closer distance is used as the first base point to continue searching backwards, and finally the PPG peak sequence after ECG signal correction is obtained. R c .

6. The heart rate self-calibration monitoring system based on deep learning model according to claim 1, characterized in that: The deep learning model includes an initial convolution module, a deep feature extraction module, a time series modeling module and a feature fusion and optimization module; The initial convolution module performs feature extraction based on the convolution layer. The extracted feature map is subjected to the global maximum pooling operation to obtain the reduced-dimensional features, and then normalized to obtain the feature ; The deep feature extraction module extracts the features Extracting deep features through grouped convolution ; The time series modeling module captures the temporal dynamic characteristics of the signal through the projection layer, constructs long-term and short-term dependencies, processes each time step through the GRU unit, and outputs the time series features. ; The feature fusion and optimization module realizes the feature mapping between channels through the projection layer 1×1 convolution to obtain the fusion feature after dimension reduction. , expressed as: ； Output through the dual-channel pooling layer: ； in, represents the maximum pooling layer, represents the average pooling layer, Indicates preset weight; Generate heart rate prediction results through a fully connected layer.

7. The heart rate self-calibration monitoring system based on deep learning model according to claim 6, characterized in that: The projection layer and the dual-channel pooling layer are respectively inserted into the ReLU activation layer, and the L2 norm normalization layer is added, and the fusion features are processed by the ReLU activation layer and the L2 norm normalization layer.

8. The heart rate self-calibration monitoring system based on deep learning model according to claim 6, characterized in that: The feature vectors output by the deep learning model are concatenated to obtain the feature data training set. The importance score of each feature is obtained based on regression training. The features are screened based on contribution and bagging ensemble training is performed to output the heart rate prediction results, including: Filter features based on contribution and obtain filtered features , based on the corresponding labels Constructing a training set ; For the training set The number of features and the number of data are randomly sampled according to the set ratio, and input into multiple deep learning models for bagging ensemble training. The outputs of multiple deep learning models are weighted averaged to obtain the final heart rate prediction result.

9. A heart rate self-calibration monitoring method based on a deep learning model, characterized in that: A heart rate self-calibration monitoring system based on a deep learning model according to any one of claims 1 to 8 is provided, comprising the following steps: Collect ECG signals, multi-band PPG signals and acceleration signals, classify the user status based on the acceleration signal, re-collect the ECG signal when a change in the user status is detected, and calibrate the PPG signal; The collected ECG signals, multi-band PPG signals and acceleration signals are filtered and amplified, and the signals are converted into analog-to-digital signals; Dynamically adjust the light wave band weights of the PPG signal collection according to the user status to obtain a weighted optimized multi-band PPG signal, extract the peak and trough features in the amplified PPG signal, and the R wave features in the ECG signal, adjust the peak position of the PPG signal according to the R wave position in the ECG signal, and complete the calibration of the PPG signal; Heart rate prediction is performed based on the deep learning model. The weighted optimized multi-band PPG signal and the calibrated PPG signal are respectively input into the deep learning model. The feature vectors output by the deep learning model are concatenated to obtain the feature data training set. The importance score of each feature is obtained based on regression training. The features are screened based on the contribution and bagging ensemble training is performed to output the heart rate prediction results.

Images