Physiological signal monitoring method and wearable device

Abstract

The invention discloses a physiological signal monitoring method and wearable equipment, relates to the technical field of detection, and mainly aims to solve the problem of poor accuracy of heart state judgment for an electrocardiogram in the prior art. Comprising the steps that basic information of a user and physiological signals collected in real time are obtained, the physiological signals comprise electrocardiosignals, electromyographic signals and acceleration signals of the heart, the physiological signals are collected based on a multi-mode sensor on the wearable device, and the wearable device is fixed to the body of the user through a connecting assembly of a strip type structure; the physiological signals are subjected to multi-modal fusion based on a multi-modal fusion algorithm, physiological parameter data are obtained, a monitoring task is determined, and the multi-modal fusion algorithm comprises one of parallel fusion, serial fusion and attention fusion; and calling a monitoring prediction model which is matched with the monitoring task and has completed model training to perform monitoring processing on the physiological parameters and the user basic information to obtain a monitoring result, and performing feature comparison based on the monitoring result.

Description

Technical Field

[0001] This application relates to the field of detection technologies, and particularly to a method for monitoring physiological signals and a wearable device. Background Art

[0002] As a physiological signal, electrocardiogram data is an important basis for judging whether the human body is operating normally. Effective monitoring of electrocardiogram data has become the main means of medical diagnosis. Especially for critically ill patients, strict and accurate electrocardiogram data monitoring is required to ensure targeted treatment for patients.

[0003] Currently, existing electrocardiogram data shows the heart activity in the form of images, that is, the device briefly records the heart's electrical activity, and then the heart activity on the electrocardiogram is manually judged. For example, only judging whether the heart rate exceeds a threshold, etc. However, since critically ill patients in the intensive care unit need real-time electrocardiogram monitoring, a large amount of electrocardiogram data will be generated. Single manual judgment may result in a large number of errors and consume a large amount of human resources, thus affecting the accuracy of electrocardiogram data as a physiological signal. Summary of the Invention

[0004] In view of this, this application provides a method for monitoring physiological signals and a wearable device, mainly aiming to solve the problem of poor accuracy in judging the heart state based on existing electrocardiograms.

[0005] According to one aspect of this application, a method for monitoring physiological signals is provided, which is applied to a wearable device and includes:

[0006] Obtain the user's basic information and the physiological signals collected in real time. The physiological signals include the electrocardiogram signal, electromyogram signal, and acceleration signal of the heart. The physiological signals are collected based on multi-modal sensors on the wearable device, and the wearable device is fixed on the user through a connecting component with a strip structure;

[0007] Perform multi-modal fusion on the physiological signals based on a multi-modal fusion algorithm to obtain physiological parameter data, and determine a monitoring task. The multi-modal fusion algorithm includes one of parallel fusion, serial fusion, and attention fusion;

[0008] Retrieve a monitoring prediction model that matches the monitoring task and has completed model training to perform monitoring processing on the physiological parameters and the user's basic information, and obtain a monitoring result for feature comparison based on the monitoring result.

[0009] Further, before retrieving the monitoring prediction model that matches the monitoring task and has completed model training to perform monitoring processing on the physiological parameters and the user's basic information to obtain a monitoring result, the method further includes:

[0010] Obtain physiological parameter training samples matching different monitoring tasks;

[0011] Create a multi-modal dynamic hierarchical correlation prediction network matching the monitoring task, the multi-modal dynamic hierarchical correlation prediction network includes a modal feature encoding layer, a correlation capture layer, and a modal fusion layer, and the output layer of the multi-modal dynamic hierarchical correlation prediction network includes multiple branch networks matching the monitoring task;

[0012] Train the model of the multi-modal dynamic hierarchical correlation prediction network based on the physiological parameter training samples to obtain a monitoring prediction model;

[0013] Among them, the loss function of the multi-modal dynamic hierarchical correlation prediction network is adjusted based on the weight value of the monitoring task.

[0014] Further, the multi-modal fusion of the physiological signals based on the multi-modal fusion algorithm to obtain physiological parameter data includes:

[0015] When performing multi-modal fusion on the physiological signals based on the attention fusion, perform modal feature partitioning on the physiological signals to determine the feature types, and the feature types include low-level modal features, global modal features, and cross-modal features;

[0016] If it is the low-level modal feature, calculate the local time series weights based on the electrocardiogram signal, the electromyogram signal, and the acceleration signal of the heart respectively;

[0017] If it is the global modal feature, calculate the global weight based on the feature mean value in the electrocardiogram signal, the electromyogram signal, and the acceleration signal of the heart;

[0018] If it is the cross-modal feature, calculate the interaction weight based on the interaction dot product between the electrocardiogram signal, the electromyogram signal, and the acceleration signal of the heart;

[0019] Calculate based on the local time series weight, the global weight, and the interaction weight and the weight coefficient to determine the total attention weight value;

[0020] Determine the physiological parameter data of feature fusion based on the total attention weight value and the dynamic weight coefficient.

[0021] Further, before obtaining the user's basic information and the real-time collected physiological signals, the method further includes:

[0022] Synchronize the timestamps of the electromyogram sensor, heart rate sensor, and acceleration sensor in the multimodal sensor according to a dynamically adjusted sampling frequency, and collect electrocardiogram signals and electromyogram signals through the synchronized electromyogram sensor and heart rate sensor, and collect acceleration signals based on the synchronized acceleration sensor.

[0023] Further, after retrieving the monitoring prediction model that matches the monitoring task and has completed model training to perform monitoring processing on the physiological parameters and the user's basic information, and obtaining a monitoring result, the method further includes:

[0024] Map the electrocardiogram signal, electromyogram signal, and acceleration signal of the heart to wavelet scales, and extract dynamic features at the wavelet scales based on the correlation in the time domain and frequency domain;

[0025] Calculate the correlation and minimum correlation deviation of the dynamic features, and use sparse regularization constraints to constrain the attention network;

[0026] Extract the dynamic features through the constrained attention network, the correlation, and the minimum correlation deviation to obtain analysis features, and perform comparison based on the analysis features to obtain classification features.

[0027] Further, the method further includes:

[0028] Generate dynamic labels according to the user's basic data and the analysis features, and store the monitoring results, the electrocardiogram signals, the electromyogram signals, and the acceleration signals based on the dynamic labels and signal modalities;

[0029] Wherein, when storing, at least one of a time window, a dynamic sorting priority, a multimodal sorting, and data editing is used for storage.

[0030] According to another aspect of the present application, a wearable device is provided, including: a wearable body, a processor, a multimodal sensor, and a strip-shaped connection component,

[0031] The connection component includes a plurality of magic tapes. A collar is provided at the top of the wearable body, and a plurality of openings are provided on the wearable body. The magic tapes are connected through the openings to fix the user's body;

[0032] The multimodal sensor is disposed on the wearable body for collecting multimodal physiological data;

[0033] The processor is disposed on the wearable garment body and is used to obtain the user's basic information and the physiological signals collected in real time. The physiological signals include the electrocardiogram signal, electromyogram signal, and acceleration signal of the heart; perform multimodal fusion on the physiological signals based on a multimodal fusion algorithm to obtain physiological parameter data, and determine a monitoring task. The multimodal fusion algorithm includes one of parallel fusion, serial fusion, and attention fusion; retrieve a monitoring prediction model that matches the monitoring task and has completed model training to perform monitoring processing on the physiological parameters and the user's basic information to obtain a monitoring result, so as to perform feature comparison based on the monitoring result.

[0034] Further, the connection component includes a first magic tape, a second magic tape, a third magic tape, and a fourth magic tape. A protective piece is also provided on the wearable device. The first magic tape is provided on one side of the protective piece close to the second magic tape;

[0035] Two symmetrical cuffs are formed on the wearable garment body. A first opening is formed on the wearable garment body, and the interior of the first opening communicates with the interior of the collar and the cuffs; an elastic band is stitched to the front side of the wearable garment body, and two symmetrical notches are formed on the front side of the wearable garment body.

[0036] The second magic tape is provided on the front side of the wearable device. A second opening is formed on the wearable device, the fourth magic tape is fixedly connected to the wearable device, and the third magic tape is fixedly connected to one side of the wearable device close to the fourth magic tape;

[0037] A transmission and elastic device is provided on the wearable device. A second connecting piece is stitched to the front side of the wearable device. A male buckle is provided on the front side of the wearable device, and a female buckle is provided on one side of the second connecting piece close to the male buckle. The male buckle and the female buckle are buckled with each other.

[0038] Further, the transmission and elastic device includes a fixed block.

[0039] One side of the fixed block close to the wearable device is fixedly connected to the surface of the wearable device. A fixed piece is fixedly connected to the front side of the wearable device, and a power piece is fixedly connected to one side of the fixed piece away from the wearable device; a chute is formed on one side of the fixed block close to the power piece, and the surface of the power piece is slidably connected to the interior of the chute. A power slot is formed in the interior of the fixed block, and a mounting frame is slidably connected to the interior of the power slot. A mounting groove is formed on one side of the power piece close to the mounting frame, and one side of the mounting frame close to the mounting groove slidably penetrates the interior of the power slot and extends into the interior of the mounting groove. The surface of the mounting frame is slidably connected to the interior of the mounting groove;

[0040] A spring is fixedly connected to the top of the mounting frame, and the top of the spring is fixedly connected to the inner wall of the power slot;

[0041] One side of the mounting bracket is fixedly connected to a transmission bracket that is slidably connected to the inside of the power slot. The side of the transmission bracket away from the mounting bracket slidably penetrates the inside of the power slot and extends to the outside of the fixed block. A handle is fixedly connected to the side of the transmission bracket away from the mounting bracket.

[0042] Further, heat dissipation holes extending to the outside of the wearable device are provided inside the wearable device; a fold line is provided on the front side of the wearable device.

[0043] By means of the above technical solutions, the technical solutions provided in the embodiments of the present application have at least the following advantages:

[0044] The present application provides a method for monitoring physiological signals and a wearable device. Compared with the prior art, in the embodiments of the present application, by obtaining user basic information and physiological signals collected in real time, the physiological signals include electrocardiogram signals, electromyogram signals, and acceleration signals of the heart, and the physiological signals are collected based on multi-modal sensors on the wearable device; the wearable device is fixed on the user's body through a connection component with a strip structure; based on a multi-modal fusion algorithm, multi-modal fusion is performed on the physiological signals to obtain physiological parameter data, and a monitoring task is determined, and the multi-modal fusion algorithm includes one of parallel fusion, serial fusion, and attention fusion; a monitoring prediction model that matches the monitoring task and has completed model training is called to perform monitoring processing on the physiological parameters and the user basic information to obtain a monitoring result, so as to perform feature comparison based on the monitoring result. After stably fixing the wearable device on the user's body, accurate physiological signals are collected to realize the monitoring of the user's physiological condition, greatly improving the monitoring accuracy of physiological signals, so as to provide personalized suggestions with different monitoring features for the user.

[0045] The above description is only an overview of the technical solutions of the present application. In order to be able to understand the technical means of the present application more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features, and advantages of the present application more obvious and understandable, the specific embodiments of the present application are specifically exemplified below. Description of the Drawings

[0046] By reading the detailed description of the preferred embodiments below, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only for the purpose of showing the preferred embodiments and are not considered to be a limitation of the present application. And throughout the drawings, the same reference numerals are used to represent the same components. In the drawings:

[0047] Figure 1 A flowchart of a method for monitoring physiological signals provided in an embodiment of the present application is shown;

[0048] Figure 2Shows a schematic diagram of multimodal signal fusion provided by an embodiment of the present application;

[0049] Figure 3 Shows a schematic diagram of a physiological signal monitoring process using a general deep neural network provided by an embodiment of the present application;

[0050] Figure 4 Shows a schematic diagram of the hardware structure of a wearable device provided by an embodiment of the present application;

[0051] Figure 5 Shows a schematic diagram of the first wearable body provided by an embodiment of the present application;

[0052] Figure 6 Shows a schematic diagram of the second wearable body provided by an embodiment of the present application;

[0053] Figure 7 Shows a schematic diagram of a heat dissipation hole structure provided by an embodiment of the present application;

[0054] Figure 8 Shows a schematic diagram of the installation structure of the first wearable body provided by an embodiment of the present application;

[0055] Figure 9 Shows a schematic diagram of the installation structure of the second wearable body provided by an embodiment of the present application;

[0056] Among them, wearable body - 1; collar - 2; first opening - 3; cuffs - 4; protective piece - 5; notch - 6; first magic tape - 7; second magic tape - 8; elastic band - 9; fold line - 10; first connecting piece - 11; third magic tape - 12; fourth magic tape - 13; second connecting piece - 14; female buckle - 15; male buckle - 16; fixing piece - 17; power piece - 18; fixing block - 19; heat dissipation hole - 20; power groove - 21; spring - 22; mounting bracket - 23; mounting groove - 24; sliding groove - 25; transmission bracket - 26; handle - 27; second opening - 28; sensor - 29; heart rate sensor - 30; processor - 31. Detailed implementation manners

[0057] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be completely conveyed to those skilled in the art.

[0058] An embodiment of the present application provides a method for monitoring physiological signals, as Figure 1 shown, the method includes:

[0059] 101. Obtain the user's basic information and the physiological signals collected in real time.

[0060] In the embodiments of the present application, the current execution entity as the monitoring execution end of the physiological signals can be a processor installed on a wearable device or a server that can perform remote processing. The embodiments of the present application do not make specific limitations. Among them, the wearable device is a device worn on the user's body for monitoring physiological signals to collect physiological signals for a long time in real time. At this time, the physiological signals include electrocardiogram signals, electromyogram signals, and acceleration signals of the heart. Moreover, the physiological signals are collected based on multi-modal sensors on the wearable device. The wearable device is fixed on the user's body through a connection component with a strip structure so that the user can stably wear it on the body for physiological signal collection. The embodiments of the present application do not make specific limitations.

[0061] 102. Perform multi-modal fusion on the physiological signals based on a multi-modal fusion algorithm to obtain physiological parameter data and determine a monitoring task.

[0062] In the embodiments of the present application, since the physiological signals are collected based on multi-modal sensors, in order to perform effective feature recognition, the current execution end performs multi-modal fusion on the physiological signals based on a multi-modal fusion algorithm to obtain physiological parameter data for monitoring and prediction. At this time, the multi-modal fusion algorithm includes one of parallel fusion, serial fusion, and attention fusion. In addition, for different monitoring and prediction models, in order to meet the monitoring tasks of different physiological signals, the current execution end first determines the monitoring task. At this time, the monitoring tasks include motion mode monitoring, fatigue degree monitoring, abnormal situation monitoring, etc. The embodiments of the present application do not make specific limitations. At this time, the fused physiological parameter data not only includes cardiac activity data, but also includes electromyogram data, motion state data, and characteristic data of other related signals to comprehensively reflect the user's health and motion state. In a specific embodiment, as Figure 2 shown, when performing feature fusion, the physiological signals can be subjected to feature segmentation, feature alignment, and feature calculation, and then the selected filtered features, deep features, and artificial features are fused. The embodiments of the present application do not make specific limitations.

[0063] It should be noted that physiological signals include: cardiac activity data, muscle activity data, motion state data, and other relevant data. Specifically, cardiac activity data may include, but is not limited to, heart rate (the number of heartbeats per unit time, reflecting the heart's pumping ability), which is a core indicator of cardiac activity; R-R interval (the interval between adjacent R waves in an electrocardiogram, characterizing heart rate variability (HRV)); QRS wave characteristics (the amplitude and duration of QRS waves in an electrocardiogram, used to analyze the electrical activity of the heart); heart rate recovery speed (HRR) (how quickly the heart rate recovers after exercise, reflecting the heart's adaptability and health status). Muscle activity data includes, but is not limited to, core indicators of muscle activity, electromyogram signal amplitude (RMS value) (reflecting the intensity of muscle contraction), spectral characteristics (mean frequency (Mean Frequency, MF), median frequency (Median Frequency, MDF), used to evaluate muscle fatigue status (for example, a decrease in the frequency of electromyogram signals usually indicates muscle fatigue)), muscle activation time (the time when the muscle starts and ends contracting, used to analyze the coordination of action execution). Motion state data includes, but is not limited to, core indicators of motion state, acceleration amplitude representing the speed and intensity of the user's motion, acceleration direction change for analyzing the motion trajectory and direction through three-axis acceleration signals, and the user's motion rhythm (step frequency, step amplitude) calculated through the periodic characteristics of acceleration signals. Other relevant data includes, but is not limited to, temperature signals (skin surface temperature), which are core indicators for evaluating the body's heat consumption and metabolic status, and posture data (such as sitting, standing, running) for analyzing body posture changes through acceleration and gyroscope signals. The embodiments of the present application do not make specific limitations. In addition, the above-mentioned physiological signal sources are sensors such as heart rate sensors (HR), electromyogram sensors (EMG), and acceleration sensors (ACC). At the same time, for multi-modal physiological signals, they can also be collected based on a combination of multi-modal sensors. The embodiments of the present application do not make specific limitations.

[0064] 103. Retrieve the monitoring prediction model that matches the monitoring task and has completed model training, and perform monitoring processing on the physiological parameters and the user's basic information to obtain a monitoring result for feature comparison based on the monitoring result.

[0065] In the embodiments of the present application, the current execution end retrieves the monitoring prediction model that has been pre-trained based on the monitoring task, and thus performs monitoring processing on the physiological parameters and the user's basic information based on this monitoring prediction model to obtain a monitoring result. Among them, the user's basic information includes, but is not limited to, user age, user gender, etc., which can be obtained through the input method. The embodiments of the present application do not make specific limitations. In addition, in the current execution end, a monitoring prediction model that matches the different monitoring tasks can be pre-trained to perform feature comparison for different monitoring tasks based on the obtained monitoring results.

[0066] It should be noted that in a specific implementation scenario, the current execution end can be trained using a general deep neural network model. For example, Figure 3 as shown, a monitoring prediction model is obtained to simultaneously complete the monitoring results of multiple target predictions (such as motion pattern recognition, fatigue assessment, anomaly detection, etc.). Among them, the input layer: combines the fused physiological parameters with the user's basic information (age, height, weight) to construct a multi-modal input matrix, denoted as X = [X fusion , X static . The feature encoding layer encodes the features of different modal data through the multi-head attention mechanism, denoted as: H i = Attention(Q i , K i , V i ), i ∈ {ECG, EMG, ACC}, where Q i , K i , V i are the query, key, and value matrices of different modal signals respectively, and H i is the feature vector of the corresponding modality. The concatenation layer: concatenates the multi-modal features with the user's basic information to construct a global feature vector: H global = Concat([H ECG , H EMG , H ACC , X statlc , ]). The multi-task output layer: uses different fully connected layers to output the multi-task prediction results respectively: y mode = σ(W mode H global + b mode ); y fatigue = σ(W fatigue H global + b fatigue ); y abnormal = σ(W abnormal H global + b abnormal ). In addition, the multi-task joint loss function in the embodiments of this application is denoted as: L global = λ 1 L mode + λ 2 L fatigue + λ 3 L abnormal , where L mode , L fatigue , L abnormal are the task losses of motion pattern, fatigue level, and anomaly detection respectively, and λ 1 , λ 2 , λ 3is the loss weight, which is used to balance the importance of each monitoring task. Moreover, the loss function can be constrained by the true labels to complete the update of the model parameters. During the training process, the data is input into the network in batches for forward propagation to calculate the predicted values, and the model parameters are updated through backpropagation until the loss function converges. At this time, the general model is corrected by the individual historical data to generate a personalized deep neural network model suitable for the individual. Furthermore, the model can also be optimized through correction. The correction parameter update formula is expressed as θ personalized = θ universal + Δθ, where θ universal is the general model parameter, and Δθ is the personalized adjustment term, which is obtained through fine-tuning with a small sample. At this time, during the real-time monitoring process, the personalized adjustment can update the parameters of the key network layer according to the feedback signal of the current user. where L personalized is the personalized loss function, and η is the learning rate.

[0067] In another embodiment of the present application, for further limitation and explanation, before the step of retrieving the monitoring prediction model that matches the monitoring task and has completed model training to monitor the physiological parameters and the user basic information to obtain the monitoring result, the method further includes:

[0068] Obtain physiological parameter training samples that match different monitoring tasks;

[0069] Create a multi-modal dynamic hierarchical correlation prediction network that matches the monitoring task;

[0070] Train the model of the multi-modal dynamic hierarchical correlation prediction network based on the physiological parameter training samples to obtain a monitoring prediction model.

[0071] To achieve accurate prediction based on machine learning algorithms and improve the monitoring accuracy of physiological signals, the current execution end pre-learns and trains the monitoring prediction model. That is, first, physiological parameter training samples matching different monitoring tasks are obtained. At the same time, a multi-modal dynamic hierarchical correlation prediction network matching different monitoring tasks is constructed. Specifically, in one embodiment, to better understand the complex relationship between the fused physiological parameter data and the health status (such as exercise mode, fatigue level, abnormal conditions), a multi-modal dynamic hierarchical correlation prediction model (Hierarchical Dynamic Multimodal Correlation Model, HD-MCM) matching the monitoring task is constructed. The multi-modal dynamic hierarchical correlation prediction model can capture the spatio-temporal correlation between multi-modal data and provide high-precision prediction in combination with user basic information (such as age, gender, etc.). Among them, the multi-modal dynamic hierarchical correlation prediction network includes a modal feature encoding layer, a correlation capture layer, and a modal fusion layer. The goal of the modal feature encoding layer is to extract temporal features and local features from data of different modalities. The model input parameters include the fused physiological parameter data: heart rate (HR), electromyogram (EGR), acceleration (ACC), and the user basic information includes age and gender, which are encoded respectively. The heart rate signal (HR) can be encoded by a convolutional neural network (CNN) to extract local features, denoted as h HR = CNN HR (x HR ); The electromyogram (EGR) encoding can use the short-time Fourier transform (STFT) to convert the electromyogram signal into a spectrogram and then input it into a two-dimensional convolutional neural network (2D-CNN), denoted as h EMG = 2D-CNN HR (STFT(x EMG )); The acceleration (ACC) encoding can capture the dynamic characteristics of the time series based on a recurrent neural network or a long short-term memory network (LSTM), denoted as h EMG = 2D-CNN HR (STFT(x EMG )) to improve the flexibility of data processing. The goal of the correlation capture layer is to capture the complex correlation between different modal data and in the time dimension. At this time, a multi-head dynamic attention mechanism is adopted to capture the relationship between modalities and in the time dimension, which can be expressed as:

[0072]

[0073] Among them, Q, K, V are the query, key, and value matrices from different modal features, d kis the dimension of the key vector for scaling. At the same time, modal correlation capture and temporal correlation capture are also introduced in the correlation capture layer. Among them, modal correlation capture is used to identify the information complementary relationship between different modalities, and temporal correlation capture is used for data changes of the same modality at different times, so as to enhance the modeling ability of the collaborative relationship between modalities. The goal of the modal fusion layer is to structure the basic information of the user, such as age and gender. At this time, the fusion method adopted can encode the personalized features through a multi-layer perceptron (MLP), expressed as: h user = MLP(x user ), and the personalized feature vector can also be concatenated with the multi-modal spatio-temporal feature vector, expressed as h final = [h fusion ; h user , or a fully connected layer can be used to perform a non-linear mapping on the fused vector to extract the final features, so as to add user personalized information on the basis of modal fusion, making the prediction results more accurate and applicable to individual differences.

[0074] In the embodiment of the present application, when performing model training, the output layer of the multi-modal dynamic hierarchical correlation prediction network includes multiple branch networks matching the monitoring tasks, that is, targeted prediction results are generated for different monitoring tasks (such as motion pattern recognition, fatigue degree evaluation, anomaly detection). For example, when the corresponding branch network is the motion pattern recognition network branch, it is expressed as: y motion = Softmax(W motion ·h final + b motion ); when it is the fatigue degree evaluation network branch, it is expressed as: y fatigue = σ(W fatigue ·h final + b fatigue ), and when it is the anomaly detection network branch, it is expressed as: y abnormal = σ(W abnormal ·h final + b abnormal ), and the embodiment of the present application does not make specific limitations.

[0075] In addition, when performing model training, the sample label as the output sample is used as the target value of the actual sample to guide the model learning. The real label can be specifically divided into the task label of motion pattern recognition, that is, the real category of the marked motion pattern (such as classification labels like "static", "jogging", "running fast", etc.), the task label of fatigue degree evaluation, that is, the real level of the marked fatigue state (such as "no fatigue", "mild fatigue", "severe fatigue"), and the task label of anomaly detection, that is, the abnormal state of the marked physiological signal (such as "normal", "abnormal"). At this time, it can be directly marked based on the real label. At this time, the source of the real label can be manually marked by medical experts (such as the definition of abnormal heart rate), and can also be manually marked through user feedback (such as the user manually recording the fatigue state), or can be inferred by rules (such as inferring fatigue based on the heart rate recovery time and the change of electromyogram signal). The embodiments of the present application do not make specific limitations.

[0076] It should be noted that, in order to improve the dynamic learning effect of the model, the loss function of the multi-modal dynamic hierarchical correlation prediction network is adjusted based on the weight value of the monitoring task, that is, for the learning model of multiple tasks, the current execution end pre-configures the total loss function, expressed as: L total = λ 1 L motion + λ 2 L fatigue + λ 3 L abnormal , where λ 1 , λ 2 , λ 3 are the weights of each monitoring task to achieve the purpose of dynamically adjusting and balancing the task priorities.

[0077] In another embodiment of the present application, for further limitation and explanation, the steps of performing multi-modal fusion on the physiological signals based on the multi-modal fusion algorithm to obtain physiological parameter data include:

[0078] When performing multi-modal fusion on the physiological signals based on the attention fusion, the modal features of the physiological signals are divided to determine the feature types;

[0079] If it is the low-level modal feature, the local temporal weights are calculated respectively based on the electrocardiogram signal, the electromyogram signal, and the acceleration signal of the heart;

[0080] If it is the global modal feature, the global weight is calculated based on the feature mean of the electrocardiogram signal, the electromyogram signal, and the acceleration signal of the heart;

[0081] If it is the cross-modal feature, the interaction weight is calculated based on the interaction dot product between the electrocardiogram signal, the electromyogram signal, and the acceleration signal of the heart;

[0082] Calculate based on the local temporal weight, the global weight, and the interaction weight and the weight coefficients to determine the total attention weight value;

[0083] Determine the physiological parameter data for feature fusion based on the total attention weight value and the dynamic weight coefficient.

[0084] For the effective prediction of multimodal physiological signals, the current execution end adopts a fusion method to perform physiological signal fusion once, and then uses it as the input parameter of the model for prediction. At this time, in order to perform effective fusion for different monitoring tasks, when performing multimodal fusion of physiological signals based on attention fusion, the current execution end determines to perform modal feature partitioning on the physiological signals to determine the feature types. At this time, the signal features are partitioned according to the feature types, including low-level modal features, global modal features, and cross-modal features, and single-modal attention, global attention, and cross-modal interaction attention are respectively applied for fusion to capture the spatio-temporal local characteristics, global relationships, and inter-modal dependencies of the signals. In addition, weight adjustment can be performed for different monitoring tasks, and the weights of the dynamic attention mechanism (Multi-Level Attention Mechanism) can be adjusted dynamically. The embodiments of the present application do not make specific limitations.

[0085] In a specific embodiment, if it is the low-level modal feature, calculate the local temporal weights based on the electrocardiogram signal, the electromyogram signal, and the acceleration signal of the heart respectively. Specifically, for the enhancement of single-modal features in the low-level modal features, the low-level feature attention method is adopted, that is, calculate the layout temporal weights within the signals for the electrocardiogram signal, electromyogram signal, and acceleration signal of the heart respectively to enhance the features at key time points, expressed as Among them, is the weight of signal i at time point t, is the feature of signal i at time point t, and T is the number of time steps.

[0086] In a specific embodiment, if it is the global modal feature, calculate the global weight based on the feature means of the electrocardiogram signal, the electromyogram signal, and the acceleration signal of the heart respectively. Specifically, for the fusion of modal features in the global modal features, use global attention to calculate the global weights of different modal features to capture the correlations between modal features, expressed as β i is the global modal weight of signal i, mean(X i ) is the feature mean (global information) of signal i, and n is the number of modalities (ECG, EMG, ACC).

[0087] In a specific embodiment, for the cross-modal features, the interaction weights are calculated based on the dot product of the interaction between the electrocardiogram signal, the electromyogram signal, and the acceleration signal of the heart. For cross-modal features, cross-modal interaction attention is adopted. At this time, a bidirectional attention mechanism is used to enhance the feature interaction between modalities, and the fused features are generated, expressed as γ ij is the interaction weight between modality i and modality j, is the bilinear interaction (dot product) of the features of modalities i and j.

[0088] It should be noted that the current execution end can dynamically adjust the attention weights at different levels according to the type and importance of the monitoring task. At this time, the total attention weight is calculated based on the local temporal weight, the global weight, and the interaction weight and the weight coefficient, expressed as where λ 1 , λ 2 , λ 3 are weight coefficients, which can be dynamically optimized through task learning. At the same time, a temporal convolutional network (TCN) is used to extract the temporal context relationship of the signal to guide the weight assignment in the time dimension. δ t = softmax(W t ·TCN(X t-1；t+1 ))), δ t = softmax(W t ·TCN(X t-1；t+1 ))), TCN is a temporal convolutional module that captures the temporal dependence of the signal. Finally, the physiological parameter data of feature fusion is determined based on the total attention weight value and the dynamic weight coefficient, that is, the fused features are obtained through the multi-layer attention mechanism and the dynamic weight weighted summation method, expressed as Furthermore, the fused high-dimensional features are input into a fully connected network and a Softmax layer to complete the task output:

[0089] In another embodiment of the present application, for further limitation and explanation, before the steps of obtaining the user's basic information and the physiological signals collected in real time, the method further includes:

[0090] Synchronize the timestamps of the electromyogram sensor, the heart rate sensor, and the acceleration sensor in the multi-modal sensor according to the dynamically adjusted sampling frequency, and collect the electrocardiogram signal and the electromyogram signal through the synchronized electromyogram sensor and heart rate sensor, and collect the acceleration signal based on the synchronized acceleration sensor.

[0091] To improve the acquisition effectiveness of each signal, the current execution end first dynamically adjusts the sampling frequency. At this time, during the acquisition process, the sampling frequency, gain, or sensor sensitivity is dynamically adjusted according to the user's state (such as exercise intensity, physiological signal fluctuation). For example, through a real-time feedback mechanism, the sampling frequency of irrelevant signals is reduced to save power consumption, and the sampling resolution of key signals is improved. The dynamic sampling frequency can be expressed as: where f s (t) is the dynamic sampling frequency, f base is the basic sampling frequency, is the change speed of the signal, and α is the adjustment coefficient. In addition, for the synchronous acquisition of multi-modal data, the signal sampling frequencies of different sensors can also be different. The current execution end uses a synchronous correction algorithm (such as timestamp alignment, multi-modal cross-correlation analysis) to ensure the timing consistency of multi-sensor signals, expressed as t sync = t sensorl + Δt calibration , t sync is the timestamp after synchronization, t sensorl is the time of the reference sensor, and Δt calibration is the synchronous correction offset.

[0092] It should be noted that the acquisition of each signal in the embodiments of the present application can be based on a composite sensor, that is, integrating multiple sensor functions into a single sensor module, reducing the device volume and power consumption, and improving the consistency of data acquisition. Specifically, an ECG electrode (heart rate sensor) and an sEMG electrode (electromyogram sensor) are integrated onto the same electrode surface to uniformly acquire electrocardiogram and electromyogram signals. A pressure sensing film is attached to the surface of the acceleration sensor to monitor the wearing fit pressure and ensure the signal acquisition quality. At this time, since the signal frequency bands of different sensors are different, the current execution end uses a frequency domain filtering module to separate the composite signal. The original acquired signal is S(t) = S ECG (t) + S EMG (t) + S ACC (t). After separation, it is expressed as S ECG (t) = H ECG (f) · S(t), S EMG (t) = H EMG (f) · S(t), S ACC (t) = H ACC (f) · S(t), where H(f) is a band-pass filter. In addition, the sensor can be encapsulated using a flexible circuit (Flexible PCB), embedding the sensor in a flexible material to adapt to the dynamic deformation of the human body, enabling the composite sensor to reduce the volume of the wearable device, improve the signal acquisition quality, achieve synchronous acquisition of multi-modal signals, and enhance data consistency.

[0093] In the embodiments of the present application, for the user's motion state, the sensitivity of the sensor can also be dynamically adjusted to improve the accuracy of signal acquisition. First, real-time state recognition is performed, and the user's motion state is judged in real time through the acceleration signal, expressed as where ||A(t)|| is the modulus of the acceleration, and ε is the threshold of the motion state. According to the state, the sensor gain is switched, expressed as Thus, real-time adjustment is achieved using an automatic gain control (AGC) module. The current execution end monitors the signal quality in real time. At this time, if the noise level is high, the sensitivity is increased to reduce the interference of motion artifacts, dynamically optimize the signal acquisition accuracy, and improve the device power consumption management efficiency. The power consumption is reduced in the stationary state. In addition, for the sensor, an integrated noise suppression function can also be adopted to reduce the interference of motion artifacts and environmental noise on the signal. At this time, for the hardware filtering layer, an electromagnetic shielding layer is added to the surface of the heart rate and electromyography sensors to reduce electromagnetic interference, and a common-mode noise suppression circuit (such as a differential amplifier) is added at the sensor interface. For the dynamic noise suppression algorithm, a reference signal (such as baseline drift ultrasound) can be collected, and the noise component is eliminated through real-time modeling, expressed as where is the noise estimate value, and h(·) is the noise modeling function. In addition, a collaborative correction algorithm can be introduced between the acceleration sensor and the electromyography sensor to detect and filter the artifacts caused by motion, expressed as: S corrected (t) = S raw (t) - α·S motion (t), which can suppress the environmental noise and motion artifacts in real time, improve the purity of the collected signal, and enhance the anti-interference ability of the sensor. Finally, multiple sensors are integrated through a local sensor network (LSN) to achieve data collaborative acquisition and intelligent processing. Specifically, for the network architecture, each sensor node can have independent data acquisition and processing capabilities, communicate with the central node (Central Hub) wirelessly, and the data synchronization protocol can be t node = t hub - Δt. Furthermore, distributed processing is performed, and data processing (filtering and noise reduction) is carried out at each node execution part to reduce the load of the central node. At this time, based on the central node fusion algorithm, the signals from different sensors are uniformly fused at each central node to optimize the consistency of multimodal data, realize the convenience of sensor maintenance and expansion, and improve the overall reliability of data acquisition.

[0094] In another embodiment of the present application, for further limitation and explanation, after the step of retrieving the monitored prediction model that matches the monitoring task and has completed model training to perform monitoring processing on the physiological parameters and the user basic information to obtain a monitoring result, the method further includes:

[0095] Map the electrocardiogram signal, the electromyogram signal, and the acceleration signal of the heart into wavelet scales, and extract the dynamic features at the wavelet scales based on the correlation between the time domain and the frequency domain;

[0096] Calculate the correlation and the minimum correlation deviation of the dynamic features, and use sparse regularization constraints to constrain the attention network;

[0097] Extract the dynamic features through the constrained attention network, the correlation, and the minimum correlation deviation to obtain analysis features, and perform comparison based on the analysis features to obtain classification features.

[0098] In order to improve the signal analysis ability, achieve efficient feature selection, thereby improve the multi-modal embedding and dynamic weight comparison mechanism, and enhance the reliability and applicability of signal comparison analysis, the current execution end maps the electrocardiogram signal, the electromyogram signal, and the acceleration signal of the heart into wavelet scales, and extracts the dynamic features at the wavelet scales based on the correlation between the time domain and the frequency domain, that is, an adaptive scale parameter adjustment mechanism is introduced for wavelet transform, and the scale is adaptively adjusted according to the instantaneous frequency of the signal. The formula is expressed as where a(t) is the scale function for adaptive adjustment of the signal instantaneous frequency, and ψ a,b (t) is the adaptive wavelet function, realizing the goal of dynamically mapping the instantaneous frequency of the signal to the wavelet scale, and extracting the non-linear dynamic features of different frequency bands. In addition, the current execution end can evaluate the correlation between the signal time domain and frequency domain distributions through the time-frequency domain feature cross-entropy model, and obtain the correlation and the minimum correlation deviation of the dynamic features. At this time, the correlation is expressed as where P x (i) and Q y (i) respectively represent the normalized probability distributions in the time domain and the frequency domain, and the coupling relationship of the signal is captured through the cross-entropy H(x,y), depicting the time-frequency relationship of the physiological signal from the feature dimension and improving the modeling accuracy. It is also possible to learn the importance of different features of the signal through the dynamic attention module and calculate the attention weight φ(f i ) = MLP(f i ), φ(f i ) is the feature weight extracted by the multi-layer perceptron (MLP), and α is the dynamic weight adjustment factor, thus fusing the time domain, frequency domain, and cross-domain features into a unified weight model to highlight the importance of signal features.

[0099] It should be noted that in order to enhance multimodal information, the current execution end uses a constrained attention network as an enhancement module, combines it with relevance and minimum relevance deviation to extract dynamic features, obtains analysis features, and conducts a comparison based on the analysis features to obtain classification features. At this time, by constructing a correlation matrix C between multimodal features ij , the correlation enhancement screening objective between multimodal features is defined, denoted as where C ij represents the correlation between features f i , f j . δ ij is the Kronecker delta, used to maintain the independence between features, and J corr is the minimum relevance deviation, used to retain independent features, thereby enhancing the feature complementarity between different signal modalities and automatically eliminating redundant features. Among them, the current execution end selects kernel sparse regularization feature screening, that is, introduces a sparse regularization constraint on the feature weights based on the kernel function, denoted as where Φ(X) is the kernel mapping of the input features, w is the feature weight, and λ 1 , λ 2 are regularization coefficients, serving as the constrained attention network to further improve the screening accuracy. Furthermore, features are gradually screened through multimodal feature interaction selection, and the iterative formula is expressed as: J sparse = J corr + J sparse , η is the learning rate, and f t+1 is the screened feature set, serving as the analysis feature. Finally, in order to enhance the signal classification and clustering capabilities, the current execution end conducts a comparison on the analysis features based on the signal contrast learning model based on multimodal embedding to obtain classification features. Specifically, the multimodal features are mapped to a unified feature space through the embedding network, and the formula is expressed as: where f ECG , f EMG , f ACC are feature extraction functions for different modalities, is feature concatenation, and then it is calculated through the contrast learning loss function, expressed as where d(z i , z j ) = ||z i - z j || 2 is the Euclidean distance, and τ is the temperature parameter.

[0100] It should be noted that the current execution end can introduce a modality weight adjustment factor to dynamically adjust the loss weight according to the importance of the signal:

[0101] L weighted = α ECG L ECG + α EMG L EMG + α ACC L ACC , where α ECG + α EMG + α ACC = 1, and is dynamically adjusted through the attention mechanism, thereby improving the performance of signal classification and clustering and optimizing the results.

[0102] In another embodiment of the present application, for further limitation and illustration, the steps further include:

[0103] Generate dynamic labels according to the user basic data and the analysis features, and store the monitoring results, the electrocardiogram signal, the electromyogram signal, and the acceleration signal based on the dynamic labels and the signal modality.

[0104] To improve the label accuracy of multi-dimensional features, the current execution end first generates dynamic labels according to the user basic data and the analysis features to achieve accurate dynamic classification. Specifically, the dynamic label formula is L t = f(C ECG , C EMG , C ACC ) + g(age, gender), where C ECG , C EMG , C ACC are the classification features of the electrocardiogram, electromyogram, and acceleration signals, and f(·), g(·) are classification functions, which generate labels according to the physiological data and the user characteristics. In a specific embodiment, the current execution end can perform multi-level classification storage based on a hierarchical structure, that is, adopt a hierarchical storage structure to classify the data in multiple levels according to categories and time series. For example, the first-level classification is the signal modality (ECG, EMG, ACC), and the second-level classification is the feature label (normal / abnormal, movement / rest, etc.), so the hierarchical classification improves the orderliness of data management and the retrieval efficiency. In addition, an adaptive classification model based on deep learning can be introduced, that is, use a deep neural network (DNN) to automatically learn the classification rules of the data and dynamically adjust the classification strategy. For example, by clustering the historical data and learning the classification rules, new classification labels are generated adaptively, making the classification results more intelligent and closer to the actual needs.

[0105] It should be noted that the monitoring results, electrocardiogram signals, electromyogram signals, and acceleration signals are stored based on dynamic tags and signal modalities. When storing, at least one of a time window, dynamic sorting priority, multi-modal sorting, and data editing is used for storage. Specifically, for efficient sorting of the time window, the data is arranged in blocks according to the time window. After block division, efficient querying and batch processing are supported. For example, the data is time-blocked in minutes to construct a time index. The time window formula is expressed as W i ={x t |t∈[t i ,t i +Δt]}, where t i is the starting point of the time window, and Δt is the length of the time window. For dynamic sorting of priorities, that is, the data sorting is dynamically adjusted by combining data importance and access frequency. The priority formula is P(x i ) = α·F(x i ) + β·R(x i ), where F(x i ) is the data access frequency, R(x i ) is the importance score of the data, and α and β are weight factors that control their relative influence. For multi-modal signals, independent sorting is performed according to different signal modalities to generate a modal index table. For example, the electrocardiogram signals are arranged in ascending order of heart rate, and the electromyogram signals are arranged according to muscle activation level, thereby improving the independent management ability of data in different modalities.

[0106] In the embodiments of the present application, due to the lack of real-time performance and flexibility in data editing, especially in dealing with large-scale data, delays are likely to occur, that is, the editing operations are mostly manual and lack intelligent support. Therefore, the current execution end also introduces a batch editing mechanism to simplify the data modification process through batch operations. For example, when batch-adjusting the heart rate abnormality label, the operation is expressed as: x' i =x i +Δx i , x i ∈ the abnormal range, so batch editing improves the operation efficiency of large-scale data. At the same time, a correction algorithm based on rules and learning can also be used to automatically identify and correct abnormal data. The correction formula is expressed as: where f(x i ) is the correction function, which corrects the abnormal data according to adjacent data and model prediction values, thereby improving the accuracy of abnormal data correction. Finally, the current execution end can also introduce version control in data editing to record each editing operation and support backtracking and revocation. For example, a unique version number is generated for each edit. At this time, the backtracking formula is expressed as where For the current version of the data, Δx represents the editing changes to ensure the traceability of editing and operational safety.

[0107] In another embodiment of the present application, for further limitation and explanation, the steps further include:

[0108] After obtaining the monitoring results, personalized analysis can also be performed based on different monitoring tasks. At this time, the analysis and application scenarios, in addition to real-time motion analysis, fatigue monitoring suggestions, and anomaly detection alarms, can also include fields such as health management, chronic disease monitoring, and postoperative rehabilitation. The embodiments of the present application do not make specific limitations.

[0109] The embodiments of the present application provide a method for monitoring physiological signals. After stably fixing the wearable device on the user, accurate physiological signals are collected to realize the monitoring of the user's physiological conditions, greatly improving the monitoring accuracy of physiological signals, so as to provide personalized suggestions with different monitoring features for the user.

[0110] Further, as an implementation of the above Figure 1 shown method, the embodiments of the present application provide a wearable device, as Figure 3 shown, the wearable device includes: a processor 31, a multimodal sensor 32, and a strip-shaped connection component 33, a wearable body 1,

[0111] Specifically, as Figure 4 、 5 shown on the wearable body 204, the connection component includes a plurality of magic tapes. A collar 2 is provided at the top of the wearable body 1, and a plurality of openings are provided on the wearable body 1. The magic tapes are connected through the openings to fix the user's body;

[0112] The multimodal sensor 29 is disposed on the wearable body 1 for collecting multimodal physiological data;

[0113] The processor is disposed on the wearable garment body 1 and is used to obtain the user's basic information and the physiological signals collected in real time. The physiological signals include the electrocardiogram signal, electromyogram signal, and acceleration signal of the heart; perform multimodal fusion on the physiological signals based on a multimodal fusion algorithm to obtain physiological parameter data, and determine a monitoring task. The multimodal fusion algorithm includes one of parallel fusion, serial fusion, and attention fusion; retrieve a monitoring prediction model that matches the monitoring task and has completed model training to perform monitoring processing on the physiological parameters and the user's basic information to obtain a monitoring result, so as to perform feature comparison based on the monitoring result. Furthermore, through the real-time analysis of the deep learning algorithm, the user can understand their own exercise performance and physiological condition in real time. The data and suggestions are intuitively presented through the front-end interface of the application program to help the user monitor their health condition, perform efficient exercise training, and adjust the exercise plan in a timely manner. For example, when the user is going for a morning run, the electrode garment monitors the heart rate and muscle activity in real time. After the App analyzes the data, it gives an exercise load assessment and suggestions to avoid overtraining. Another example is that when the user is performing fitness training, the electrode garment monitors the muscle activity and heart rate and provides real-time feedback to guide the user to control the exercise intensity and perform the movements correctly. In addition, the multimodal sensor 29 may include a strip-shaped acceleration sensor, an electromyogram sensor, and a heart rate sensor, and the embodiments of the present application do not make specific limitations.

[0114] Further, the connection component includes a first Velcro 7, a second Velcro 8, a third Velcro 12, and a fourth Velcro 13. A protective sheet 5 is further provided on the wearable device. A first Velcro 7 is provided on one side of the protective sheet 5 close to the second Velcro 8;

[0115] Two symmetric cuffs 4 are opened on the wearable garment body 1. A first opening 3 is opened on the wearable garment body 1. The inside of the first opening 3 is communicated with the inside of the collar 2 and the cuffs 4; An elastic band 9 is stitched and connected to the front side of the wearable garment body 1. Two symmetric notches 6 are opened on the front side of the wearable garment body 1,

[0116] The second Velcro 8 is provided on the front side of the wearable device. A second opening 28 is opened on the wearable device. A fourth Velcro 13 is fixedly connected to the wearable device. A third Velcro 12 is fixedly connected to one side of the wearable device close to the fourth Velcro 13;

[0117] A transmission and elastic device is provided on the wearable device. A second connecting chain 14 is stitched and connected to the front side of the wearable device. A male buckle 16 is provided on the front side of the wearable device. A female buckle 15 is provided on one side of the second connecting chain 14 close to the male buckle 16. The male buckle 16 and the female buckle 15 are buckled with each other.

[0118] In the embodiments of the present application, through the adhesion of the first magic tape and the second magic tape 8, the position of the protective sheet 5 is positioned, so that the protective sheet 5 can more smoothly protect the electrode film. At the same time, through the adhesion of the first magic tape and the second magic tape 8, the protective sheet 5 can be more conveniently positioned and disassembled, enabling medical staff to more conveniently replace and observe the electrocardiogram monitoring electrode film, and to a certain extent reducing the direct exposure of the electrocardiogram monitoring electrode film to the environment and avoiding the situation that the patient is likely to knock it off. At the same time, through the mutual cooperation of the third magic tape 12 and the fourth magic tape 13, the first connecting piece 11 can fix the fabrics on both sides of the first opening 3, so that the wearable body 1 is sleeved on the patient's arm. Through the mutual cooperation of the male buckle 16 and the female buckle 15, the fabrics on both sides of the second opening 28 can be positioned and fixed, so that the wearable body 1 on the patient's chest is connected, and the wearable body 1 is more smoothly sleeved on the patient.

[0119] Further, as Figure 6 , 7 shown, the transmission tensioning device includes a fixed block 19,

[0120] One side of the fixed block 19 close to the wearable device is fixedly connected to the surface of the wearable device. A fixed piece 17 is fixedly connected to the front side of the wearable device, and a power piece 18 is fixedly connected to the side of the fixed piece 17 away from the wearable device; a sliding groove 25 is opened on one side of the fixed block 19 close to the power piece 18, and the surface of the power piece 18 is slidably connected to the inside of the sliding groove 25. A power groove 21 is opened inside the fixed block 19, and an installation frame 23 is slidably connected to the inside of the power groove 21. An installation groove 24 is opened on one side of the power piece 18 close to the installation frame 23. One side of the installation frame 23 close to the installation groove 24 slidably penetrates the inside of the power groove 21 and extends to the inside of the installation groove 24, and the surface of the installation frame 23 is slidably connected to the inside of the installation groove 24;

[0121] The top of the installation frame 23 is fixedly connected with a spring 22, and the top of the spring 22 is fixedly connected to the inner wall of the power groove 21;

[0122] One side of the installation frame 23 is fixedly connected with a transmission frame 26 that is slidably connected to the inside of the power groove 21. The side of the transmission frame 26 away from the installation frame 23 slidably penetrates the inside of the power groove 21 and extends to the outside of the fixed block 19, and a handle 27 is fixedly connected to the side of the transmission frame 26 away from the installation frame 23.

[0123] In the embodiment of the present application, the fixing piece 17 drives the fabric on the front side of the wearable body 1 to move in the direction of the fixing block 19 with the folding line 10 as the folding line 10. When the power piece 18 moves to a suitable position, the spring 22 resets to make the mounting frame 23 slide into the interior of the mounting groove 24, thereby realizing the position fixation of the power piece 18 and changing the size of the wearable body 1, so that the wearable body 1 can adapt to patients of different sizes, and both male and female patients can wear the wearable body 1 comfortably, so that the wearable body 1 can more smoothly protect the electrocardiogram monitoring electrode film of the patient. In a specific embodiment, the cross-section of the power groove 2121 is Z-shaped, and the cross-section of the mounting groove 2424 is a right trapezoid. The embodiment of the present application does not make specific limitations.

[0124] Further, as Figure 8 shown, heat dissipation holes 20 extending to the outside of the wearable device are provided inside the wearable device; a folding line 10 is provided on the front side of the wearable device.

[0125] In the embodiment of the present application, the breathability of the wearable body 1 can be enhanced to a certain extent through the heat dissipation holes 20, so that the patient is not prone to sweating when wearing the wearable body 1, and the patient can wear the wearable body 1 more comfortably.

[0126] In a specific implementation scenario, real-time feedback and adaptive adjustment mechanisms can also be introduced to dynamically adjust the parameters of the monitoring electrode film protection garment (such as pressure distribution, position of the sensor 29, data acquisition frequency, etc.) to improve comfort and monitoring accuracy, making it have real-time feedback, intelligent self-adaptability, multi-scenario adaptability, providing dynamic adjustment and real-time assessment of health status, improving the user experience, automatically optimizing the performance of the protective garment according to the user's state, taking into account both comfort and monitoring accuracy, so as to be applicable to various scenarios such as sports training, rehabilitation monitoring, and cardiovascular disease early warning.

[0127] Specifically, in the real-time data feedback scenario, after adopting multi-modal signals, the signal instructions are analyzed in real time, and the current working state of the sensor 29 is fed back. At this time, the signal quality Q s is defined as the signal quality score, with a range of [0, 1], SNR is the signal-to-noise ratio of the current signal, and ExpectedSNR is the signal-to-noise ratio in the ideal case. When Q s is lower than the threshold of 0.8, the calibration or data acquisition of the sensor 29 is triggered. Furthermore, through dynamic pressure adjustment, according to the user's motion state or static state, the pressure distribution of the wearable body 1 can be adjusted to ensure the stability and comfort of the sensor 29. At this time, the pressure control formula is: P(t) = P base + K·α(t), where P(t) is the pressure distribution at the current time t, P baseis the base pressure value (which can be predicted based on static tests), K is the acceleration sensitivity coefficient, α(t) is the acceleration value real-time monitored by sensor 29, so as to integrate flexible airbags on the wearable body 1, and realize dynamic pressure adjustment through a micro air pump to adjust the pressure in real time. At the same time, it can also be adaptively adjusted through the data acquisition frequency, that is, the signal acquisition frequency is dynamically adjusted according to the intensity obtained by the user to achieve the purpose of optimizing power consumption and data quality. At this time, the sampling frequency adjustment formula is f s = f base ·(1 + α·ActivityLevel), f s is the current sampling frequency, f base is the base sampling frequency, α is the adjustment coefficient, and ActivityLevel is the activity intensity calculated from the acceleration signal, expressed as Therefore, a dynamic sampling mechanism is added to the sensor 29 management module to automatically adjust the sampling frequency by real-time monitoring the activity intensity. Finally, the processor also provides real-time feedback for user interaction, which can be evaluated in real time according to the user monitoring status, and can be fed back according to the real-time health status (such as abnormal heart rate alarm, excessive exercise indication). At this time, the health risk index calculation formula is expressed as: where R is the health risk index, HR current is the current heart rate, HR rest is the resting heart rate, HR max is the maximum heart rate. When R exceeds the safety threshold of 0.8, the user can be reminded to reduce the exercise intensity through the application. Of course, personalized exercise training suggestions can also be provided for the user, that is, according to the user's current monitoring status and activity history, personalized exercise training suggestions are generated, and a personalized model can be trained using historical data, expressed as: TrainingAdice = f ML (X history ), f ML is a machine learning-based model, and X history is the user's historical exercise and physiological data.

[0128] In a specific implementation scenario, the monitoring mode can also be dynamically switched according to the user's activity scenario (such as stationary, exercising, resting, etc.) or health status (such as normal, abnormal, etc.) by configuring a scenario recognition module to optimize system performance and user experience. Among them, deep learning technology can be combined with multi-modal signals (electrocardiogram, electromyogram, acceleration, etc.) to identify the user's current scenario and trigger corresponding mode adjustments. For example, adjusting the data acquisition frequency, activating specific sensors 29 or starting an alarm mechanism is not specifically limited in this application embodiment. In the construction of the scenario recognition module, first, the input signals include: X ECG is the electrocardiogram signal, X EMG is the electromyogram signal, X ACCis the acceleration signal, each signal length is T, and the feature dimension is d. Secondly, data preprocessing can be performed, that is, the standardized signal is expressed as i∈{ECG,EMG,ACC},μ i , σ i is the mean and standard deviation of the signal. When extracting features, the time domain features (such as signal peak value and mean value) are represented as Feature time =[max(X),min(X),mean(X)], frequency domain features (such as main frequency, bandwidth) are expressed as: Feature freq = [FFT(X)], after concatenation, the feature concatenation is: Feature combined =Concat(Feature time ,Feature freq ). Among them, for scene classification, the scene recognition network based on LSTM can be expressed as: h t =σ(W h X t +U h h t-1 +b h ), h t is the hidden state at time step t, W h , U h , b h is the network parameter, output scene category C∈{stationary, moving, resting, abnormal}, output probability P(C)=softmax(W out H T +b out ), P(C) is the probability distribution of the scenario category. In addition, for the switching of dynamic modes, the switching rule that can be adopted is to dynamically adjust the monitoring mode according to the scenario category C, including static scenario (reducing the sampling frequency to save power), sports scenario (activating all sensors 29 to capture dynamic changes), rest scenario (focusing on monitoring ECG signals to evaluate recovery), and abnormal scenario (triggering alarms and uploading data). The sampling frequency adjustment for the above scenarios is expressed as f s =f base ·(1+α·P(C=motion)), f base is the basic sampling frequency, and α is the adjustment factor. At this time, for the activation control of sensor 29, S active is the set of activated sensors 29. When C = abnormal and P(C = abnormal)>τ, TriggerAlarm = 1, τ is the alarm triggering threshold. Finally, the system can display user scenario and mode adjustment information in real time through the application APP interface. For example, the current state is sports mode, and the sampling frequency is prompted to 200Hz. It can also be optimized automatically, that is, based on the recorded historical scenario data, Hhistory = [C 1 , C 2 ,..., C T .

[0129] In another embodiment of the present application, on the basis of scenario recognition and dynamic mode switching, a multi-modal data prediction and health risk assessment module is further designed. Combining the multi-modal signals and historical data collected in real time by the user, through time series prediction and risk scoring, potential health risks are actively identified, so as to provide scientific health management suggestions. Specifically, for multi-modal time series prediction, the input signals are is the electrocardiogram signal in the past T time steps, is the electromyogram signal in the past T time steps, is the acceleration signal in the past T time steps. Feature normalization is i ∈ {ECG, EMG, ACC}, and the historical time window data integration is represented as Among them, a multi-modal time series prediction model is constructed through Transfoemer or Bi LSTM, and the self-attention mechanism is expressed as: Q, K, V are the query, key, and value matrices of the input signal respectively, and the output prediction result is: The predicted future signal trend is Among them, the loss function uses weighted mean square error (MSE) as the optimization objective, ω i is the weight of different signals, which is dynamically adjusted according to the signal importance. For the monitoring risk score, the health risk score R = α 1 R ECG + α 2 R EMG + α 3 R ACC , R ECG , R EMG , R ACC are the sub-risk scores based on electrocardiogram, electromyogram, and acceleration signals respectively, and α i is the signal importance weight. At this time, the sub-score calculation of each signal includes: the electrocardiogram signal score is expressed as HR t is the heart rate at time step t, HR expected is the expected heart rate according to age and activity level, and the electromyogram signal score

[0130] Score R Classification of Health Risk Levels: Low Risk (Normal): R < 0.5, Medium Risk (Warning): 0.5 ≤ R < 0.8; High Risk (Abnormal): R ≥ 0.8. At this time, the current health risk level can be displayed, which can be distinguished by colors (green, yellow, red), and real-time suggestions can also be provided. For example, in the normal state: maintain the current activity state, in the warning state: it is recommended to reduce the exercise intensity or enter the rest mode, and in the abnormal state: trigger an alarm and notify the medical service. For personalized health intervention, a personalized intervention plan can be generated in advance according to the prediction signal and risk score. For example: increase the rest time, reduce the exercise intensity, and regularly review cardiovascular indicators.

[0131] In a specific implementation scenario, when in use, after the wearable body 1 is unfolded and placed on the hospital bed, the patient lies on the wearable body 1, and then the wearable body 1 wraps the patient. Through the mutual cooperation of the third magic tape 12 and the fourth magic tape 13, the first connecting piece 11 can fix the fabrics on both sides of the first opening 3, so that the wearable body 1 is sleeved on the patient's arm. Through the mutual cooperation of the male buckle 16 and the female buckle 15, the fabrics on both sides of the second opening 28 can be positioned and fixed, so that the wearable body 1 in front of the patient's chest is connected, and the wearable body 1 is more smoothly sleeved on the patient, making it more convenient for the patient to wear the wearable body 1, and making it more smooth for the patient's neck and arm to enter the inside of the collar 2 and the cuff 4 respectively. Through the adhesion of the first magic tape 7 and the second magic tape 8, the position of the protective piece 5 is positioned, so that the protective piece 5 can more smoothly protect the electrode film. At the same time, through the adhesion of the first magic tape 7 and the second magic tape 8, the protective piece 5 can be more conveniently positioned and disassembled, making it more convenient for medical staff to replace and observe the electrocardiogram monitoring electrode film. After the staff wears the wearable body 1 on the patient, the power piece 18 moves inside the chute 25. Since the cross-section of the installation groove 24 is a right trapezoid, when the power piece 18 moves, through the inclined surface of the installation groove 24, the installation frame 23 can be pushed upward to move, compressing the spring 22, and driving the fabric on the front side of the wearable body 1 to move in the direction of the fixed block 19 with the fold line 10 as the fold line. When the power piece 18 moves to a suitable position, the spring 22 resets and the installation frame 23 slides into the inside of the installation groove 24, thus realizing the position fixation of the power piece 18 and changing the size of the wearable body 1, so that the wearable body 1 can adapt to patients of different sizes, and both male and female patients can comfortably wear the wearable body 1.

[0132] An embodiment of the present application provides a monitoring wearable device for physiological signals. After stably fixing the wearable device on a user, accurate physiological signals are collected to monitor the physiological condition of the user, greatly improving the monitoring accuracy of physiological signals, so as to provide personalized suggestions with different monitoring features for the user.

[0133] Obviously, those skilled in the art should understand that the above-mentioned modules or steps of the present application can be implemented by a general-purpose computing device. They can be concentrated on a single computing device or distributed on a network composed of multiple computing devices. Optionally, they can be implemented by program codes executable by the computing device. Thus, they can be stored in a storage device and executed by the computing device. And in some cases, the steps shown or described can be executed in a sequence different from that here, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module to implement. In this way, the present application is not limited to any specific combination of hardware and software.

[0134] The above are only the preferred embodiments of the present application and are not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for monitoring physiological signals, applied to a wearable device, characterized in that: include: Obtaining basic information of the user and physiological signals collected in real time, wherein the physiological signals include electrocardiogram signals, electromyography signals and acceleration signals of the heart, and the physiological signals are collected based on multimodal sensors on a wearable device, and the wearable device is fixed to the user through a strip-type connecting component; Performing multimodal fusion on the physiological signal based on a multimodal fusion algorithm to obtain physiological parameter data and determine a monitoring task, wherein the multimodal fusion algorithm includes one of parallel fusion, serial fusion and attention fusion; A monitoring prediction model that matches the monitoring task and has completed model training is retrieved to monitor and process the physiological parameters and the user basic information to obtain monitoring results, so as to perform feature comparison based on the monitoring results.

2. The method according to claim 1, characterized in that Before the monitoring prediction model that matches the monitoring task and has completed model training is retrieved to monitor the physiological parameters and the user basic information and obtain the monitoring result, the method further includes: Obtain training samples of physiological parameters that match different monitoring tasks; Creating a multimodal dynamic hierarchical correlation prediction network that matches the monitoring task, wherein the multimodal dynamic hierarchical correlation prediction network includes a modal feature encoding layer, a correlation capture layer, and a modal fusion layer, and the output layer of the multimodal dynamic hierarchical correlation prediction network includes a plurality of branch networks that match the monitoring task; Based on the physiological parameter training samples, model training is performed on the multimodal dynamic hierarchical correlation prediction network to obtain a monitoring prediction model; Among them, the loss function of the multimodal dynamic hierarchical correlation prediction network is adjusted based on the weight value of the monitoring task.

3. The method according to claim 1, characterized in that The multimodal fusion of the physiological signals based on the multimodal fusion algorithm to obtain the physiological parameter data includes: When multimodal fusion is performed on the physiological signal based on the attention fusion, modal feature division is performed on the physiological signal to determine feature types, where the feature types include low-level modal features, global modal features, and cross-modal features; If it is the low-level modal feature, then local timing weights are calculated based on the electrocardiogram signal, the electromyography signal and the acceleration signal of the heart respectively; If it is the global modal feature, calculating the global weight based on the feature mean values ​​in the electrocardiogram signal, the electromyography signal and the acceleration signal of the heart; If it is the cross-modal feature, calculating the interaction weight based on the interaction dot product between the electrocardiogram signal of the heart, the electromyography signal and the acceleration signal; Calculate based on the local temporal weight, the global weight, the interaction weight and the weight coefficient to determine a total attention weight value; The physiological parameter data of feature fusion is determined based on the total attention weight value and the dynamic weight coefficient.

4. The method according to claim 1, characterized in that: Before obtaining the basic information of the user and the physiological signals collected in real time, the method further includes: According to the dynamically adjusted sampling frequency, the electromyographic sensor, heart rate sensor and acceleration sensor in the multimodal sensor are time-stamped and synchronized, and the electrocardiographic signal and electromyographic signal are collected through the synchronized electromyographic sensor and heart rate sensor, and the acceleration signal is collected based on the synchronized acceleration sensor.

5. The method according to claim 1, characterized in that After the monitoring prediction model that matches the monitoring task and has completed model training is retrieved to monitor the physiological parameters and the user basic information and obtain the monitoring results, the method further includes: Mapping the electrocardiogram signal, the electromyography signal and the acceleration signal of the heart into a wavelet scale, and extracting dynamic features under the wavelet scale based on the correlation between the time domain and the frequency domain; Calculating the correlation and minimum correlation deviation of the dynamic features, and constraining the attention network using sparse regularization constraints; The dynamic features are extracted through the constrained attention network, the correlation and the minimum correlation deviation to obtain analysis features, and comparison is performed based on the analysis features to obtain classification features.

6. The method according to claim 5, characterized in that The method further comprises: Generate a dynamic tag according to the user basic data and the analysis feature, and store the monitoring result and the electrocardiogram signal, the electromyography signal and the acceleration signal based on the dynamic tag and the signal modality; When storing, at least one of a time window and dynamic sorting priority, multimodal sorting, and data editing is used for storage.

7. A wearable device, characterized in that: include: Wearable body, processor, multimodal sensor and strip-like structure connection components, The connection assembly includes a plurality of Velcro strips, a collar is provided on the top of the wearable garment body, and a plurality of openings are provided on the wearable garment body, and the Velcro strips are connected through the openings to fix the user's body; The multimodal sensor is disposed on the wearable garment body and is used to collect multimodal physiological data; The processor is arranged on the wearable garment body, and is used to obtain basic information of the user and physiological signals collected in real time, wherein the physiological signals include electrocardiogram signals, electromyography signals and acceleration signals of the heart; multimodal fusion is performed on the physiological signals based on a multimodal fusion algorithm to obtain physiological parameter data, and a monitoring task is determined, wherein the multimodal fusion algorithm includes one of parallel fusion, serial fusion and attention fusion; a monitoring prediction model that matches the monitoring task and has completed model training is called to monitor and process the physiological parameters and the basic information of the user, and a monitoring result is obtained, so as to perform feature comparison based on the monitoring result.

8. The wearable device according to claim 7, characterized in that: The connecting component includes a first Velcro, a second Velcro, a third Velcro and a fourth Velcro, and the wearable device is further provided with a protective sheet, and the first Velcro is provided on a side of the protective sheet close to the second Velcro; The wearable garment body is provided with two symmetrical cuffs, and the wearable garment body is provided with a first opening, the interior of the first opening is communicated with the interior of the collar and the cuffs; an elastic band is sewn and connected to the front side of the wearable garment body, and the front side of the wearable garment body is provided with two symmetrical notches, The second Velcro is provided on the front side of the wearable device, a second opening is opened on the wearable device, a fourth Velcro is fixedly connected to the wearable device, and the third Velcro is fixedly connected to a side of the wearable device close to the fourth Velcro; The wearable device is provided with a transmission elastic device, the front side of the wearable device is sewn with a second connecting piece, the front side of the wearable device is provided with a male buckle, and the side of the second connecting piece close to the male buckle is provided with a female buckle, and the male buckle and the female buckle are buckled with each other.

9. The wearable device according to claim 7, characterized in that: The transmission tensioning device comprises a fixing block; The side of the fixing block close to the wearable device is fixedly connected to the surface of the wearable device, the front side of the wearable device is fixedly connected to a fixing sheet, and the side of the fixing sheet away from the wearable device is fixedly connected to a power sheet; a sliding groove is provided on the side of the fixing block close to the power sheet, the surface of the power sheet is slidably connected to the inside of the sliding groove, a power groove is provided inside the fixing block, and a mounting bracket is slidably connected to the inside of the power groove, a mounting groove is provided on the side of the power sheet close to the mounting bracket, a side of the mounting bracket close to the mounting groove slides through the inside of the power groove and extends to the inside of the mounting groove, and the surface of the mounting bracket is slidably connected to the inside of the mounting groove; A spring is fixedly connected to the top of the mounting frame, and the top of the spring is fixedly connected to the inner wall of the power slot; One side of the mounting frame is fixedly connected to a transmission frame slidably connected to the inside of the power slot, and the side of the transmission frame away from the mounting frame slides through the inside of the power slot and extends to the outside of the fixed block, and the side of the transmission frame away from the mounting frame is fixedly connected to a handle.

10. The wearable device according to claim 7, characterized in that: A heat dissipation hole extending to the outside of the wearable device is opened inside the wearable device; and a fold line is arranged on the front side of the wearable device.

Images