A wearable respiratory belt and early warning method for respiratory rate and tidal volume monitoring

Abstract

The application discloses a wearable breathing belt and early warning method for respiratory rate and tidal volume monitoring. The wearable breathing belt comprises a breathing belt body, a piezoelectric sensitive unit, an inertial measurement unit, a signal acquisition module, a processing module, a storage module, a communication module and an early warning output module. The piezoelectric sensitive unit is used for collecting original breathing signals corresponding to abdominal fluctuations, and the inertial measurement unit is used for collecting acceleration signals and angular velocity signals. The method pre-processes the original breathing signals and inertial signals, obtains fusion breathing features through a monotone sensing mapping layer, a selective state space model backbone network based on Mamba and a posture modulation module, and then outputs continuous breathing volume waveforms, tidal volume of each breath and respiratory rate by a physical decoder. In the training stage, a RAP-Loss total loss function including amplitude loss, rhythm loss and physical mechanics residual loss is adopted to improve the stability of tidal volume estimation under different body positions and different individual conditions. The system performs hierarchical early warning according to the two-dimensional joint threshold value of respiratory rate and tidal volume and the duration window rule, and executes local early warning prompt and remote reporting.

Description

Technical Field

[0001] This invention relates to the fields of biomedical signal processing, wearable medical monitoring, intelligent health early warning and edge artificial intelligence technology, and in particular to a wearable breathing belt and early warning method for monitoring respiratory rate and tidal volume, which is used to monitor changes in human respiratory status in real time, provide early warning of respiratory problems such as abnormal respiratory rate, insufficient tidal volume and ventilation abnormalities, and improve the convenience, accuracy and timeliness of continuous respiratory monitoring. Background Technology

[0002] Respiratory monitoring is an important component of chronic respiratory disease management, postoperative recovery observation, sleep health assessment, and home health management. Existing wearable respiratory monitoring devices typically characterize respiratory movements through signals such as chest or abdominal displacement, pressure, strain, and acceleration, and further extract respiratory-related parameters.

[0003] In existing technologies, the first type of approach mainly detects respiratory rate, respiratory rhythm, or respiratory events, but lacks the ability to continuously and stably estimate tidal volume. When only respiratory rate is obtained without simultaneously obtaining tidal volume, although it can reflect changes in the speed of breathing, it is difficult to comprehensively characterize the effective ventilation per unit breath. Therefore, in cases of insufficient ventilation, shallow and rapid breathing, or fatigue breathing, it is easy to fail to identify the actual decline in ventilation in a timely manner based solely on a normal or near-normal respiratory rate.

[0004] While the second type of approach incorporates deep learning models to regress or classify respiratory signals, it often employs convolutional networks, recurrent networks, or standard attention networks, directly regressing the output from the raw signals. This type of purely black-box data-driven approach typically has a large number of parameters, resulting in high computational and storage costs when modeling long-term sequences. Furthermore, it primarily relies on statistical correlation learning, failing to fully utilize the intrinsic connections between abdominal deformation, trunk posture, changes in chest and abdominal respiratory components, and respiratory system mechanics. This leads to the model output easily deviating from the physiologically reasonable range under conditions of individual differences, variations in wearing tightness, changes in body position, and body movement disturbances.

[0005] In the third approach, human respiration is simplified to a single one-dimensional displacement process. The output of the abdominal breathing belt, besides being affected by abdominal wall undulations, is also influenced by changes in body position, turning over, walking, sitting up, speaking, coughing, belt tightness, and changes in local wearing position. If different postures such as standing, sitting, supine, and lateral lying cannot be effectively identified, or if artifacts caused by body movement cannot be suppressed, the direct estimation of tidal volume from local abdominal deformation is prone to distortion, resulting in insufficient estimation stability under different postures and usage scenarios.

[0006] In the fourth category of solutions, some devices only provide simple warnings when the signal exceeds the limit, failing to jointly determine respiratory rate, tidal volume, duration window, posture, and signal quality. This leads to false alarms, missed alarms, or untimely warnings, making it difficult to meet the needs of continuous clinical monitoring and early risk detection in home settings.

[0007] Therefore, there is an urgent need for a wearable breathing belt and early warning method that can simultaneously monitor respiratory rate and tidal volume, take into account the real-time operation capability of the edge side, improve the estimation stability by combining posture and physical constraints, and provide graded early warning for ventilation abnormalities. Summary of the Invention

[0008] I. Technical problems to be solved To address the problems of existing wearable respiratory monitoring devices, such as the inability to accurately estimate tidal volume by only measuring respiratory rate, the heavy computational burden of pure black-box models that ignore human physical laws, insufficient generalization ability under different body positions and movement conditions, and the lack of a real-time early warning mechanism based on joint determination of respiratory rate and tidal volume, this paper provides a wearable breathing belt and early warning method for monitoring respiratory rate and tidal volume.

[0009] II. Technical Solution To address the aforementioned technical problems, this invention provides a wearable breathing belt early warning method for monitoring respiratory rate and tidal volume, comprising the following steps: acquiring raw respiratory signals corresponding to abdominal undulations via a piezoelectric sensing unit, acquiring acceleration and angular velocity signals via an IMU inertial unit, and obtaining user static parameters; performing time synchronization, denoising, segmentation, baseline correction, and signal quality assessment on the raw respiratory signals and IMU inertial signals, and inputting the raw respiratory signals into a monotonic sensing mapping layer to obtain an abdominal expansion characterization signal; inputting the abdominal expansion characterization signal into a Mamba-based selective state-space model backbone network for temporal feature extraction, and extracting based on the IMU inertial signals... Attitude modulation factor and bias modulation factor are used to conditionally modulate the temporal features to obtain fused respiratory features; the fused respiratory features are input to a physical decoder to output continuous respiratory volume waveform, tidal volume per breath, respiratory phase information, and respiratory cycle boundary information; respiratory frequency is calculated based on the respiratory cycle boundary information and / or the continuous respiratory volume waveform; the respiratory frequency and tidal volume per breath are input to an anomaly determination module to determine whether a respiratory anomaly has occurred based on a preset threshold and continuous determination rules, and a corresponding warning level is generated; when the anomaly determination module determines that the preset warning conditions have been met, a local warning and / or a remote warning are triggered, and the corresponding respiratory frequency, tidal volume, and warning result are output.

[0010] In the above method, the monotonic sensing mapping layer is used to map the original respiratory signal s(t) into a measure of abdominal distension. And satisfy:

[0011] This ensures that within the preset effective working range, the corresponding abdominal expansion measure does not change in the opposite direction when the original respiratory signal increases.

[0012] In the above method, the step of extracting attitude modulation factors and bias modulation factors based on IMU inertial signals includes: extracting attitude vectors, body dynamics intensity, gravitational direction stability, and activity state information based on acceleration and angular velocity signals; and generating attitude modulation factors based on the information. and bias modulation factor ; and in accordance with:

[0013] Intermediate features output by the backbone network of the Mamba-based selective state-space model Modulation was performed to obtain the modulated fusion respiratory characteristics. .

[0014] In the above method, the Mamba-based selective state-space model backbone network adopts an input-dependent state update mechanism, and its hidden state update satisfies:

[0015] in, Let be the input features at time t. It is in a hidden state. For output features, , , A parametric function that varies with the input.

[0016] In the above method, the physical decoder outputs at least a continuous respiratory volume waveform. tidal volume per breath Abdominal weight ratio Elastic parameters Resistance parameters and potential respiratory drive One or more of the following; wherein the total respiratory volume and the abdominal component volume satisfy:

[0017] Tidal volume per breath It is obtained from the difference in continuous respiratory volume waveforms between adjacent end-expiratory and end-inspiratory phases.

[0018] In the above method, the parameters of the Mamba-based selective state-space model backbone network, attitude modulation module and physical decoder are obtained through training. The training adopts the RAP-Loss total loss function, which includes at least an amplitude loss term, a rhythm loss term and a physical mechanical residual loss term.

[0019] This invention also provides a wearable breathing belt device for monitoring respiratory rate and tidal volume, comprising: a breathing belt body; a piezoelectric sensing unit disposed on the breathing belt body for acquiring raw respiratory signals corresponding to abdominal rise and fall; a central detachable housing module disposed on the breathing belt body, comprising an IMU inertial unit, a processor, a memory, a communication module, and an early warning output module; wherein, the IMU inertial unit is used to acquire acceleration signals and angular velocity signals; the early warning output module is used to provide local early warning prompts based on early warning results; the communication module is used for remote reporting; the memory stores a computer program, which, when executed by the processor, causes the processor to perform the aforementioned early warning method.

[0020] III. Beneficial Effects Compared with the prior art, the present invention has at least the following beneficial effects:

[0021] 1. This invention enables simultaneous monitoring of respiratory rate and tidal volume. It not only obtains respiratory rate from abdominal respiratory signals but also estimates continuous respiratory volume waveforms and tidal volume per breath. Therefore, compared to methods that only monitor respiratory rate or rhythm, this invention provides a more comprehensive reflection of the subject's effective ventilation status, facilitating the identification of abnormalities such as shallow and rapid breathing and insufficient ventilation that are difficult to detect using only respiratory rate.

[0022] 2. This invention improves the stability of tidal volume estimation under different body positions and movement conditions. It introduces IMU inertial signals to extract posture information, body movement intensity, and gravity direction stability, and generates posture modulation factors and bias modulation factors to conditionally modulate respiratory timing characteristics. This allows for correction of the mapping relationship between abdominal movement and respiratory volume under conditions such as sitting, standing, supine, lateral, turning over, or slight body movement, reducing the impact of positional changes and body movement artifacts on tidal volume estimation results.

[0023] 3. It can improve the interpretability and physiological rationality of tidal volume estimation results. This invention converts the original respiratory signal into a representation of abdominal expansion through a monotonic sensing mapping layer, maintaining a clear monotonic correspondence between the original respiratory signal and the degree of abdominal expansion. At the same time, it outputs continuous respiratory volume waveforms, tidal volume per breath, and related physical surrogate parameters through a physical decoder. In the training process, it introduces RAP-Loss constraints, including amplitude loss, rhythm loss, and physical mechanical residual loss, which helps to reduce the risk of physiologically unreasonable outputs in pure black-box models in cross-individual and cross-positional applications.

[0024] 4. This invention reduces the computational burden of modeling long-term respiratory signals, facilitating deployment in wearable devices. It employs a Mamba-based selective state-space model backbone network to extract features from respiratory time-series signals, enabling the modeling of long-term continuous respiratory sequences. Furthermore, it selectively retains respiratory-related information and suppresses body movement interference information through an input-related state update mechanism. Compared to complex time-series models with higher computational overhead, this structure helps reduce the inference burden at the edge, facilitating continuous operation in wearable respiratory monitoring devices.

[0025] 5. This invention enables graded early warning based on a combined assessment of respiratory rate and tidal volume. It incorporates respiratory rate, tidal volume per breath, signal quality, and duration window rules into the anomaly detection module. A two-dimensional threshold joint assessment mechanism using both frequency and tidal volume is employed to generate different early warning levels, triggering local alerts and / or remote reporting. Compared to early warning methods based solely on a single parameter exceeding its limit, this approach helps reduce the risk of false alarms and missed alarms, improving the timeliness of detecting abnormal respiratory states in home monitoring and continuous monitoring scenarios. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall structure and module connection of the wearable breathing belt device of the present invention; Figure 2 This is a schematic diagram of the wearable breathing belt of the present invention being worn on the human abdomen; Figure 3 This is a flowchart illustrating the overall processing flow of the method of the present invention; Figure 4 This is a diagram of the Mamba network and physical constraint decoding algorithm architecture of the present invention; Figure 5 This is a schematic diagram of the combined early warning state machine of respiratory rate and tidal volume of the present invention.

[0027] The components in the attached diagram are labeled as follows: 1-Wearable breathing belt body; 2-Piezoelectric sensitive unit; 3-Signal acquisition module; 4-Processing module; 5-Storage module; 6-Communication module; 7-Inertial measurement module; 8-Terminal device; 9-Server; 10-Power supply module; 11-Central detachable box module; 12-Abdominal wearing area; 13-Human torso; 14-Early warning output module. Detailed Implementation

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the concept of the present invention, those skilled in the art can adjust or replace the signal acquisition frequency, network structure parameters, threshold setting method, communication method, and early warning output method.

[0029] The core of this invention lies in: acquiring abdominal respiratory signals using a piezoelectric sensing unit, acquiring posture and body movement information using an IMU inertial unit, forming an abdominal expansion characterization quantity through a monotonic sensing mapping layer, extracting respiratory temporal features through a Mamba-based selective state-space model, and outputting continuous respiratory volume waveforms, tidal volume per breath, and respiratory frequency using IMU posture modulation and a physical decoder, thereby achieving graded early warning based on the joint determination of respiratory frequency and tidal volume.

[0030] The tidal volume referred to in this invention refers to the gas volume corresponding to a single respiration. The abdominal expansion characterization quantity refers to an intermediate characterization quantity related to the degree of abdominal expansion, obtained by monotonic sensing mapping of the piezoelectric sensing unit's output signal. The abdominal volume surrogate characteristic refers to a surrogate quantity related to the amplitude, shape, or volume change of abdominal respiratory movements; it can be a calibrated volumetric dimensional characteristic or a dimensionless characteristic monotonically corresponding to volume changes. The elastic parameters, drag parameters, and potential respiratory drives can be physical surrogate variables or latent variables obtained during model training, without requiring direct measurement of intrathoracic pressure, pleural pressure, or airway pressure by hardware.

[0031] Example 1: Structure and Signal Flow of Wearable Breathing Belt Device like Figure 1 and Figure 2 As shown, this embodiment provides a wearable breathing belt device for monitoring respiratory rate and tidal volume. The device includes a breathing belt body, a flexible breathing sensing unit, a signal acquisition module, a processing module, a storage module, a communication module, a power supply module, an IMU inertial unit, a central detachable housing module, and an early warning output module.

[0032] The breathing belt body is a flexible belt structure suitable for wearing around the abdomen and can be fixed around the abdomen. The piezoelectric sensitive unit is disposed on the breathing belt body and is used to sense changes in electrical signals caused by abdominal wall undulations and belt tension changes during respiration. In one embodiment, the piezoelectric sensitive unit is positioned corresponding to the periumbilical region, so that the deformation caused by abdominal expansion and contraction during inhalation and exhalation can be stably transmitted to the piezoelectric sensitive unit. The piezoelectric sensitive unit can be a PVDF piezoelectric film, a piezoelectric composite film, or other flexible sensitive elements capable of converting abdominal deformation into electrical signals. The central detachable housing module is disposed on the breathing belt body and contains an IMU inertial unit, a processing module, a storage module, a communication module, a power supply module, and a warning output module. The IMU inertial unit is used to collect triaxial acceleration signals and triaxial angular velocity signals to obtain the wearer's posture information, body dynamics intensity, gravitational stability, and activity status information. The processing module is used for time synchronization, preprocessing, feature extraction, tidal volume estimation, respiratory rate calculation, and early warning determination of the raw respiratory signal and IMU inertial signal. The storage module stores user static parameters, model parameters, historical respiratory data, and early warning records. The power supply module supplies power to the piezoelectric sensing unit, signal acquisition module, processing module, IMU inertial unit, communication module, and early warning output module. The signal acquisition module is connected to the piezoelectric sensing unit and amplifies, filters, and performs analog-to-digital conversion on the analog respiratory signal output by the piezoelectric sensing unit to form a raw respiratory signal that can be processed by the processing module. The signal acquisition module can be located within the central detachable housing module or adjacent to the piezoelectric sensing unit and connected to the processing module via an electrical connection cable. The raw respiratory signal characterizes the periodic fluctuations in the human abdomen caused by respiration. The communication module preferably uses a 4G Cat.1 wireless communication module to upload data such as respiratory rate, tidal volume, early warning level, signal quality status, posture status, and timestamps to the terminal device or server. In other embodiments, the communication module may also employ wireless communication methods such as Bluetooth, Wi-Fi, NB-IoT, or LoRa to adapt to local transmission or remote reporting needs in different usage scenarios. The warning output module is used to execute local prompts based on the warning results. In one embodiment, the warning output module includes one or more of a buzzer, a vibration motor, and an LED indicator. When the processing module determines that a preset warning condition has been met, the warning output module outputs different forms of local prompts according to the warning level; simultaneously, the communication module can remotely report the warning results to a terminal device, a server, or a doctor's terminal platform.

[0033] The signal flow process in this embodiment is as follows: the piezoelectric sensing unit senses the abdominal wall undulations and belt tension changes, and outputs a simulated breathing signal; the signal acquisition module amplifies, filters, and performs analog-to-digital conversion on the simulated breathing signal to form a raw breathing signal; the IMU inertial unit synchronously acquires triaxial acceleration and triaxial angular velocity signals; the processing module performs time synchronization, preprocessing, and quality assessment on the raw breathing signal and the IMU inertial signal, and inputs the processed data into the subsequent algorithm module to calculate the respiratory rate, continuous respiratory volume waveform, tidal volume per breath, and warning result; finally, the warning result is locally prompted through the warning output module and remotely reported through the communication module.

[0034] With the above structure, this embodiment can achieve joint estimation of respiratory rate and tidal volume using abdominal respiratory motion information and IMU attitude information without relying on oral and nasal airflow sensors, and provides hardware foundation and data input for subsequent graded early warning based on respiratory rate and tidal volume.

[0035] Example 2: Signal preprocessing, monotonic mapping, Mamba coding, and IMU attitude modulation example like Figure 3 and Figure 4 As shown, the inputs to the algorithm include two categories: dynamic signals and static parameters.

[0036] The dynamic signals include at least the original breathing signal and its first and second derivatives, the triaxial acceleration and triaxial angular velocity signals output by the IMU inertial unit, and auxiliary quantities such as the dynamic baseline, envelope signal, and signal quality indicators calculated from the original breathing signal. The user's static parameters include at least one or more of the following: height, weight, gender, age, waist circumference, and wearing pretension level. These static parameters are normalized to form a static parameter vector.

[0037] The preprocessing of the raw respiratory signal includes DC bias removal, bandpass filtering, dynamic baseline updating, and signal quality assessment. The dynamic baseline preferably uses an end-expiratory updating method, meaning the baseline is updated only when low body movement, acceptable signal quality, and the current position is near the end of expiration. This updating method can be expressed as:

[0038] in, As the baseline update factor, in one implementation, It can be set according to the sampling frequency, signal drift speed and target response time.

[0039] The preprocessing of the IMU inertial signal includes gravity component extraction, attitude angle calculation, body motion intensity estimation, and rollover event detection. Body motion intensity can be obtained by weighting the magnitude variance of acceleration, the root mean square of angular velocity, and the rate of change of acceleration at adjacent sampling points. The relevant weights can be set based on calibration data, validation set performance, or the target application scenario.

[0040] 1. Monotonic Sensing Mapping Layer The monotonic sensing mapping layer is used to transform the raw respiratory signal into a physically interpretable representation of abdominal expansion, and combines it with user static parameters to form a proxy feature of abdominal volume, which serves as a priori input for subsequent tidal volume estimation.

[0041] In a preferred embodiment, the monotonic sensing mapping layer employs a monotonic spline network with positive weight constraints, a piecewise linear monotonic network, or a cumulative positive basis function network. Let the original respiratory signal be... The static parameter vector is u Abdominal expansion characterization after monotonic mapping It can be represented as:

[0042] And satisfy:

[0043] In one implementation, the use J The summation of positive basis functions forms:

[0044] in, It is a monotonic basis function. For weights controlled by static parameters, J The preferred value is 8 to 16.

[0045] After obtaining the abdominal expansion characterization, abdominal volume surrogate features are further formed by combining user static parameters. It can take the following form:

[0046] in, , , This refers to the individualized coefficients determined by static parameters. These individualized coefficients can be obtained from calibration data, training data, or historical monitoring data, and are used to correct the correspondence between abdominal deformation and abdominal volume proxy features in different individuals.

[0047] 2. Dual-branch Mamba encoder The time-synchronized input is divided into fixed-length data segments. The segment length can be set to 24-64 sampling points; the segment step size can be set to 8-16 sampling points. During training or batch inference, the batch size is set to... B The number of time segments is T The breathing belt inlet channel is The number of IMU input channels is Then it can form:

[0048] In one embodiment, the breathing band input channel includes at least the raw respiratory signal, first derivative, second derivative, dynamic baseline, signal quality index, and abdominal volume proxy features; the IMU input channel includes at least triaxial acceleration, triaxial angular velocity, and attitude and body motion assistance quantities calculated therefrom. After linear projection, both branches are mapped to a unified hidden dimension D. In one embodiment, the hidden dimension D can be set to 64–128; in other embodiments, it can also be adjusted according to the sampling frequency, processor computing power, and target inference latency.

[0049] Each Mamba module includes a normalization unit, a local convolution unit, a selective state space update unit, and a residual connection unit. The selective state space update can be represented as:

[0050] in, For the current input, In hidden state, For output features, This is a parameter function dependent on the input. The number of layers, hidden dimensions, and local convolutional kernel size of the Mamba module can be configured according to the sampling frequency, processor computing power, and target inference latency. In one embodiment, the Mamba module adopts a multi-layer stacked structure and combines local convolutional units, selective state space update units, and residual connection units to extract respiratory temporal features.

[0051] 3. IMU Attitude Modulation Factor and Bias Modulation Factor The IMU branch output is mapped to an attitude modulation factor via a multilayer perceptron. and bias modulation factor This is used for element-level modulation of the intermediate features of respiratory band branches. Let the respiratory band branches be at the [missing information - likely a specific point or sequence]. l Layer output If the IMU branch outputs the following features at the corresponding layer, then:

[0052]

[0053] in, For the sigmoid function, To scale the range, It is the translation constant. The bias range is defined as follows. In one implementation, preset value ranges can be set for the attitude modulation factor and the bias modulation factor to avoid excessive amplification or excessive offset of the feature channels. For example, the attitude modulation factor can be limited to the range of 0.5 to 1.5, and the bias modulation factor can be limited to the range of -0.5 to 0.5. Element-level modulation is then performed:

[0054] in, This indicates element-wise multiplication. If the IMU determines that the current attitude is stable and the body motion is small, then... It tends to retain more respiratory features; if the IMU determines that the patient is currently in the stage of rolling over, sitting up, walking, or vigorous physical activity, then some channels will show signs of respiration. This will decrease, in order to suppress respiratory characteristics with greater receptor dynamic interference, while by Correct the offset.

[0055] In one implementation, the system can also output an abdominal component proportion prior based on the IMU attitude category. The abdominal component proportion prior for different attitudes can be set to different intervals and can be updated using calibration data or historical monitoring data. For example, different proportion prior intervals can be set for standing, supine, and lateral positions, and the proportion prior can then be used by the physical decoder.

[0056] Example 3: Physical Decoder and RAP-Loss Implementation Example like Figure 4 As shown, after dual-branch Mamba encoding and IMU attitude modulation, fused respiratory features are obtained. The fused respiratory features are input to a physical decoder, which outputs at least the continuous respiratory volume waveform and the tidal volume per breath; in one embodiment, the physical decoder also outputs one or more of the following: abdominal component proportion, elasticity parameter, drag parameter, potential respiratory drive, respiratory phase, and respiratory cycle boundary.

[0057] Wherein, the elastic parameter Resistance parameters and potential respiratory drive These are physiological latent variables or surrogate parameters obtained by fitting the neural network during training. They are used to constrain the estimation results to conform to the mechanical laws of the respiratory system, aiming to improve the physiological rationality of tidal volume estimation. These parameters do not require direct measurement of intrathoracic pressure, pleural pressure, or airway pressure by hardware; rather, they serve as intermediate surrogate quantities to constrain the continuous respiratory volume waveform and tidal volume estimation results. In this invention, the... , , Normalized values, dimensionless surrogate quantities, or characterization quantities that monotonically correspond to the real physical quantities can be used.

[0058] For volume estimation, a combination of abdominal volume surrogate features and abdominal component proportions can be used:

[0059] in, The abdominal volume proxy feature generated by the monotonic sensing mapping layer in Example 2. The proportion of abdominal components varies over time. To prevent tiny constants with a denominator of zero. It can be output by sigmoid, hardtanh or other bounded activation functions, and its value is preferably limited to 0.10 to 0.95.

[0060] When continuous respiratory volume waveform Once calculated, the interval of a single breath can be determined using a phase head or boundary head. For example, the tidal volume of each breath can be calculated using the peak-to-trough difference between the start and end of inspiration:

[0061] in, This represents the end of the inhalation phase of the k-th breath. This corresponds to the end of expiration. Respiratory rate can be obtained from the interval between adjacent end-expiratory phases or the interval between adjacent inspiratory peaks.

[0062] In one implementation, to ensure the non-negativity of the elasticity and drag parameters, softplus, exponential functions, or ReLU variants can be used as the output activation function. For the normalized surrogate variable, preset upper and lower bound constraints can also be set to avoid non-physiological jumps in the latent variables. The output of the latent respiratory drive can be set to bounded or unbounded form according to the model training requirements.

[0063] The RAP-Loss total loss function includes at least: an amplitude loss term, a rhythm loss term, and a physical / mechanical residual loss term. To enhance training stability, it may further include a boundary constraint term, a smoothing term, a signal quality weighting term, and a self-supervised pre-training term.

[0064] The amplitude loss term is used to constrain the deviation between the estimated tidal volume and the reference tidal volume.

[0065] in, The weighting coefficient and the boundary parameter of the Huber loss for the k-th breath tidal volume provided for the reference device can be set according to the quality of the reference tidal volume label, the scale of the training data, and the target application scenario.

[0066] Rhythm loss term, used to constrain respiratory phase and inspiratory-expiratory duration:

[0067] Among them, the estimated phase and reference phase are used to constrain respiratory phase consistency, and the estimated inspiratory duration, estimated expiratory duration and their corresponding reference values ​​are used to constrain respiratory duration consistency; the phase loss weight and duration loss weight can be set according to the reliability of the phase label, the quality of the respiratory cycle boundary labeling and the training stability.

[0068] The physical and mechanical residual loss term is used to constrain the model output to satisfy low-order respiratory system mechanical relationships.

[0069] The corresponding residual term is defined as:

[0070] P0 is the baseline term, which can be set as a constant or as a trainable bias. When P0 is set as a trainable bias, it can be updated along with the model parameters during training.

[0071] To prevent non-physiological jumps in latent variables, boundary and smoothing constraints can be further set:

[0072]

[0073] In one implementation, It can be set to 0.10; It can be set to 0.95.

[0074] The total loss function can be written as:

[0075] in, , , , and These are the weight coefficients for the amplitude loss term, rhythm loss term, physical and mechanical residual loss term, boundary constraint term, and smoothness constraint term, respectively. Each weight coefficient can be set according to the training data scale, reference label quality, model complexity, and target application scenario.

[0076] During training, Adam, AdamW, RMSProp, or other first-order gradient optimization algorithms can be used to update the model parameters. The learning rate, batch size, number of training epochs, and weights of each loss term can be set according to the training data scale, reference label quality, model complexity, and target application scenario. The continuous respiratory volume waveform, tidal volume per breath, and physical proxy variables are calculated through forward propagation. Then, the RAP-Loss total loss is calculated, and the network parameters are updated based on this total loss, thereby reducing tidal volume error, rhythm error, and physical residuals.

[0077] Example 4: Implementation of a frequency-quantity dual-parameter joint early warning mechanism like Figure 5 As shown, before making a warning judgment, the system first establishes a user-specific baseline. This individualized baseline can be established during the initial device wear or in a stable resting state, and includes baseline respiratory rate, baseline tidal volume, tidal volume fluctuation range, signal quality status, and wearing status information. When the signal quality does not meet preset requirements, the system prioritizes outputting a "signal unreliable" or "wearing check" prompt, rather than directly outputting the physiological abnormality level.

[0078] The current respiratory rate is RR, and the tidal volume of the current k-th breath is The anomaly detection module employs a two-dimensional joint early warning method combining a respiratory rate threshold, a tidal volume threshold, and a duration window. The respiratory rate threshold can be determined using a fixed medical reference range, the user's baseline respiratory rate, or a combination of both; the tidal volume threshold can be adaptively determined based on the user's baseline tidal volume, body shape parameters, posture, wearing status, and historical monitoring data.

[0079] In one implementation, the system converts the user's baseline tidal volume into Level 1, Level 2, and Level 3 tidal volume thresholds. The Level 1 threshold identifies mild tidal volume decline, the Level 2 threshold identifies more significant tidal volume insufficiency, and the Level 3 threshold identifies severe tidal volume insufficiency or suspected apnea risk. Each level of tidal volume threshold can be derived proportionally from the user's baseline tidal volume or adjusted based on age, body size parameters, postural status, and historical monitoring data.

[0080] When the respiratory rate exceeds the first frequency threshold or the tidal volume falls below the first tidal volume threshold, and the abnormal state meets the preset persistence criteria, the system enters a Level 1 warning state. When the degree of abnormality in respiratory rate or tidal volume further increases, or when the Level 1 warning state fails to recover within a subsequent time window, the system enters a Level 2 warning state. When both respiratory rate and tidal volume are abnormal, or either parameter reaches a severely abnormal condition, or when no effective respiratory cycle is detected within a preset time window, the system enters a Level 3 warning state.

[0081] The continuous judgment criteria may include at least one of the following: the number of abnormal breaths reaches a preset proportion in N consecutive breaths; the duration of abnormality reaches a preset duration within a sliding time window of length τ; and M consecutive breaths all meet the same warning level condition. By using the above continuous judgment criteria, false alarms caused by occasional body movements, single waveform abnormalities, or short-term wearing disturbances can be reduced.

[0082] To further reduce false alarms caused by factors such as body movement, turning over, coughing, talking, or loosening of the device, this embodiment may adopt one or more of the following false alarm prevention rules: (1) Majority voting rule: In the most recent N An alert is only triggered when the percentage of abnormal cycles exceeds a threshold within a single respiratory cycle. (2) Dual-window confirmation rule: The warning is triggered only after both the periodic window and the time window abnormal conditions are met simultaneously; (3) Body movement inhibition rule: If the IMU inertial unit determines that the current state is a state of intense body movement, the physiological warning level will be reduced or postponed, and a body movement interference prompt will be output; (4) Wearing validity rules: If the system determines that the strap is loose, the signal is saturated, or the sensor is in abnormal contact, it will output a wearing abnormality prompt first, instead of directly determining it as a breathing abnormality; (5) Posture-specific threshold: For different body positions such as supine, lateral, sitting or standing, the system can correct the tidal volume threshold or judgment result according to the posture state.

[0083] In a Level 1 alert state, the system can trigger a low-intensity local alert and provide a suggestive remote report according to the communication strategy. In a Level 2 alert state, the system can trigger an enhanced local alert and upload the current timestamp, respiratory rate, tidal volume, alert level, posture category, signal quality, and device identifier to the terminal device or server. In a Level 3 alert state, the system can trigger a high-intensity local alert and send a high-priority remote alert message. The local alert can be implemented through one or more of the following methods: beeping, vibration, and light. The alert intensity, frequency, and duration can be configured according to the alert level.

[0084] In one implementation, the device executes the following sequence in a loop: acquiring signals from the piezoelectric sensing unit and the IMU inertial unit; performing preprocessing and quality assessment; obtaining fused respiratory features through a monotonic sensing mapping layer and a dual-branch Mamba; calculating the current continuous respiratory volume waveform, tidal volume per breath, and respiratory rate via a physical decoder; updating the warning status according to two-dimensional joint warning rules and duration windows; outputting local warning prompts and / or remote warning prompts; and updating the user baseline and individualized thresholds in a stable state.

[0085] Through the above-mentioned frequency and quantity dual-parameter joint early warning mechanism, this embodiment can incorporate respiratory rate, tidal volume, signal quality, posture status and duration window into the anomaly judgment process. Compared with the early warning method that relies on a single parameter exceeding the limit, it is beneficial to reduce the risk of false alarms and missed alarms, and improve the timeliness of abnormal respiratory status detection in home monitoring and continuous monitoring scenarios.

[0086] The parameter ranges, time windows, threshold intervals, loss weights, and network structure sizes given in the above embodiments are all feasible preferred ranges. Those skilled in the art can make equivalent substitutions or partial adjustments based on the target population, sensor sensitivity, sampling frequency, computing resources, and clinical application scenarios without departing from the technical concept of this invention. The technical features in each embodiment can be combined with each other without conflict and are all within the scope of the content disclosed in this specification.

Claims

1. A wearable breathing belt early warning method for monitoring respiratory rate and tidal volume, characterized in that, Includes the following steps: S1. Collect the raw breathing signal corresponding to the abdominal rise and fall during human breathing through the piezoelectric sensing unit, collect the acceleration signal and angular velocity signal as IMU inertial signal through the IMU inertial unit, and obtain the user's static parameters; S2. Perform time synchronization, noise reduction, segmentation, baseline correction, and signal quality assessment on the original respiratory signal and IMU inertial signal, and input the original respiratory signal and user static parameters into the monotonic sensing mapping layer to obtain the abdominal expansion characterization signal. S3. Input the abdominal expansion characterization signal into the backbone network of the selective state-space model based on Mamba for temporal feature extraction, and extract attitude information, body motion intensity and gravity direction stability according to the IMU inertial signal to generate attitude modulation factor and bias modulation factor, and conditionally modulate the temporal features to obtain fused respiratory features. S4. Input the fused respiratory features into the physical decoder and output the continuous respiratory volume waveform, tidal volume of each breath, respiratory phase information and respiratory cycle boundary information. S5. Calculate the respiratory rate based on the respiratory cycle boundary information, or extract the respiratory cycle and calculate the respiratory rate based on the continuous respiratory volume waveform; S6. Input the respiratory rate and the tidal volume of each breath into the abnormality determination module, determine whether a respiratory abnormality has occurred based on the preset threshold and continuous determination rules, and generate a corresponding warning level. S7. When the anomaly determination module determines that the preset warning conditions are met, it triggers a local warning and / or a remote warning, and outputs the corresponding respiratory rate, tidal volume per breath and warning result.

2. The early warning method according to claim 1, characterized in that, The monotonic sensing mapping layer is used to map the raw respiratory signal s(t) into a measure of abdominal distension. The mapping satisfies a monotonicity constraint, such that within a preset effective working interval: This ensures that when the original respiratory signal increases, the corresponding abdominal distension measurement does not change in the opposite direction.

3. The early warning method according to claim 1, characterized in that, The step of extracting attitude modulation factors and bias modulation factors based on IMU inertial signals includes: extracting attitude vectors, body motion intensity, gravitational direction stability, and activity state information based on acceleration and angular velocity signals; and generating attitude modulation factors based on the attitude vectors, body motion intensity, gravitational direction stability, and activity state information. and bias modulation factor and in accordance with: Intermediate features output by the backbone network of the Mamba-based selective state-space model Modulation was performed to obtain the modulated fusion respiratory characteristics. .

4. The early warning method according to claim 1, characterized in that, The Mamba-based selective state-space model backbone network employs an input-dependent state update mechanism, and its hidden state update satisfies: in, Let be the input features at time t. It is in a hidden state. For output features, It is a parameter function that varies with the input to achieve selective memorization of respiratory-related information and selective suppression of body motion artifacts.

5. The early warning method according to claim 1, characterized in that, The physical decoder outputs at least a continuous respiratory volume waveform. Tidal volume per breath In one embodiment, the physical decoder also outputs abdominal volume proxy features. Abdominal weight ratio Elastic parameters Resistance parameters and potential respiratory drive One or more of the following; wherein, in the physical decoder output abdominal volume proxy features and abdominal proportion Under these conditions, the continuous respiratory volume waveform satisfies: Tidal volume per breath It is obtained from the difference in continuous respiratory volume waveform between the end of inspiration and the end of expiration in the k-th respiratory cycle.

6. The early warning method according to claim 1, characterized in that, The respiratory rate is calculated in any of the following ways: (1) Calculate the instantaneous respiratory rate based on the time interval between the boundaries of two adjacent respiratory cycles; (2) Calculate the average respiratory rate based on the average period of multiple respiratory cycles within the sliding time window; (3) Extract the respiratory cycle and calculate the respiratory rate based on the inspiratory peak, end of expiration or phase inflection point in the continuous respiratory volume waveform.

7. The early warning method according to claim 1, characterized in that, The parameters of the Mamba-based selective state-space model backbone network, the attitude modulation module for generating the attitude modulation factor and the bias modulation factor, and the physical decoder are obtained through training. The training adopts the RAP-Loss total loss function, which includes at least an amplitude loss term, a rhythm loss term, and a physical mechanical residual loss term.

8. The early warning method according to claim 7, characterized in that, The amplitude loss term is used to constrain the difference between the estimated tidal volume and the reference tidal volume, and the amplitude loss term is expressed as follows: in, Let k be the estimated tidal volume for the k-th respiration. This is the reference tidal volume for the k-th breath. These are the weighting coefficients. To prevent tiny constants with a denominator of zero.

9. The early warning method according to claim 7, characterized in that, The rhythm loss term is used to simultaneously constrain respiratory phase and respiratory duration, and the rhythm loss term includes at least phase loss and respiratory duration loss, wherein: in, and These are the estimated respiratory phase and the reference respiratory phase, respectively. and These are the estimated inspiratory duration and the reference inspiratory duration for the k-th breath, respectively. and These are the estimated expiratory duration and the reference expiratory duration for the k-th breath, respectively. and These are the weighting coefficients.

10. The early warning method according to claims 5 and 7, characterized in that, The physical and mechanical residual loss term is constructed based on the mechanical relationship of the respiratory system, which satisfies the following: The corresponding physical and mechanical residual loss term is expressed as follows: in, As the benchmark term, and for and Set nonnegativity constraints for Set range constraints within the preset interval.

11. The early warning method according to claim 1, characterized in that, The anomaly detection module employs a two-dimensional threshold joint detection mechanism of respiratory rate and tidal volume, including at least the following: when the respiratory rate exceeds the first frequency threshold range, or the tidal volume is lower than the first tidal volume threshold, it is determined as a level one warning; when the respiratory rate exceeds the second frequency threshold range, or the tidal volume is lower than the second tidal volume threshold, it is determined as a level two warning; when both respiratory rate and tidal volume are abnormal, and / or the degree of abnormality of any one of them reaches a preset severity threshold, it is determined as a level three warning. The second frequency threshold range has a higher degree of abnormality than the first frequency threshold range, and the second tidal volume threshold is lower than the first tidal volume threshold.

12. The early warning method according to claim 1 or 11, characterized in that, The persistent determination rule includes at least one of the following: (1) The number of abnormal breaths reaches a preset proportion in N consecutive breaths; (2) The duration of the anomaly reaches the preset duration within a sliding time window of length τ; (3) All M consecutive breaths meet the conditions for the same warning level; Furthermore, the preset threshold adopts at least one of a fixed threshold and an individualized threshold, wherein the individualized threshold is adaptively determined based on the user's baseline respiratory rate, baseline tidal volume, body shape parameters and / or wearing status parameters.

13. A wearable breathing belt device for monitoring respiratory rate and tidal volume, characterized in that, include: The breathing belt itself; A piezoelectric sensing unit is disposed on the breathing belt body and is used to collect the original breathing signal corresponding to the abdominal rise and fall; A central detachable housing module, mounted on the breathing belt body, includes an IMU inertial unit, a processor, a memory, a communication module, and a warning output module. The IMU inertial unit is used to collect acceleration and angular velocity signals. The warning output module is used to provide local warnings based on the warning results. The communication module is used for remote reporting. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the warning method according to any one of claims 1 to 12.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the early warning method according to any one of claims 1 to 12.

Images