Respiratory state recognition method and system based on flexible deformation perception

Abstract

This invention relates to the field of respiratory status monitoring technology, specifically disclosing a respiratory status recognition method and system based on flexible deformation sensing. It utilizes a template inversion process to fabricate a conductive polymer composite flexible sensor with a micro-nano inverted pyramid structure on its surface, causing a significant change in resistance under minute mechanical deformation. The flexible sensor is integrated into a wearable carrier and attached to the human respiratory or joint area to collect time-series electrical signals containing respiratory signals and motion interference. By performing time-series segmentation and feature extraction on the collected signals, and decoupling the respiratory rhythm features and motion interference features in the signals based on a time-series modeling recognition network, stable recognition of states such as calm breathing, deep breathing, rapid breathing, and coughing is achieved under dynamic interference conditions. This invention has the advantages of simple structure, low cost, high sensitivity, and strong resistance to dynamic interference.

Description

Technical Field

[0001] This invention relates to the field of respiratory monitoring technology, and in particular to a respiratory state recognition method and system based on flexible deformation sensing. Background Technology

[0002] With the development of preventive medicine, sports rehabilitation, and human-computer interaction technology, flexible wearable devices that can continuously monitor human physiological parameters and movement status are gradually attracting attention. Among them, physiological indicators such as respiratory rate, respiratory depth, and respiratory pattern have important application value in early disease screening, exercise assessment, and health management.

[0003] Most existing flexible respiratory monitoring technologies assume an approximately linear or weakly nonlinear relationship between sensor output and human body deformation, and suppress motion noise through filtering or thresholding. However, in wearable dynamic scenarios, the viscoelastic hysteresis effect commonly found in flexible conductive composite materials couples with the low-frequency, large-amplitude deformation signals introduced by human movement, resulting in significant temporal asymmetry and state correlation in the sensing signals. These coupled nonlinear characteristics cannot be effectively modeled using traditional filtering or static feature extraction methods, leading to insufficient stability and reliability of existing technologies in dynamic respiratory state recognition.

[0004] Because the mechanical deformation of the facial skin, jaw area, or chest and abdomen caused by human respiration is relatively small, existing flexible sensors with smooth or simple microstructures have limited electrical response changes within the small deformation range, resulting in low signal amplitude and difficulty in obtaining stable respiratory characteristic signals in complex environmental noise backgrounds.

[0005] The polymer elastomer materials used in flexible sensors generally exhibit viscoelastic characteristics. Under dynamic loading or cyclic deformation conditions, the recovery process of the material's internal structure has a time delay, which can easily cause hysteresis in the output signal. In continuous breathing or multi-state recognition processes, this hysteresis effect may distort the timing signal waveform, thereby affecting the accuracy of subsequent recognition results.

[0006] In real-world wearing scenarios, the subjects' walking, limb swinging, or head turning movements introduce large-amplitude mechanical deformation signals. These motion artifacts overlap with respiratory signals in both the time and frequency domains, making it difficult for signal processing methods based on traditional filtering or simple feature extraction strategies to effectively distinguish between respiratory rhythms and motion interference components. This limits the application effectiveness of flexible sensors under dynamic conditions.

[0007] How to stably acquire and accurately identify weak respiratory deformation signals under conditions of human movement interference remains a technical problem that urgently needs to be solved in the field of flexible wearable respiratory monitoring. Summary of the Invention

[0008] This invention provides a breathing state recognition method and system based on flexible deformation sensing. The technical problem it solves is: how to achieve high-sensitivity acquisition and high-precision, high-stability state recognition of weak breathing deformation signals under complex dynamic conditions with strong human motion artifact interference and the inherent viscoelastic hysteresis effect of flexible sensors.

[0009] To address the above technical problems, this invention provides a breathing state recognition method based on flexible deformation sensing, comprising the following steps:

[0010] S1. Surface functionalization graft modification of carbon-based nano-conductive filler is carried out using silane coupling agent; the modified conductive filler is fully mixed and degassed with a polymer elastomer substrate through template reversal process, and then composited. The composite is then heated and cured on the surface of an industrial sandpaper template with random micro-rough protrusions. After peeling, a flexible sensing unit is obtained; a flexible sensor is fabricated based on the flexible sensing unit, and the flexible sensor is integrated into a wearable carrier to obtain a wearable flexible sensing device.

[0011] S2. Attach the wearable flexible sensing device to the human body's mouth and nose breathing airflow area and target joints; under dynamic interference scenarios where the human body is under different exercise loads, collect the original one-dimensional time-series signals under different breathing states in real time.

[0012] S3. After normalizing the original one-dimensional time series signal, it is segmented and reconstructed into a structured three-dimensional time series feature matrix containing phase continuity.

[0013] S4. Construct a hybrid decoupled recognition network based on temporal modeling and train the hybrid decoupled recognition network using the structured three-dimensional temporal feature matrix;

[0014] S5. The wearable flexible sensing device is used to collect sensing signals in real time and process them to obtain a structured three-dimensional temporal feature matrix. The structured three-dimensional temporal feature matrix is ​​continuously input into the hybrid decoupled recognition network after training. The hybrid decoupled recognition network outputs the breathing state recognition result in real time.

[0015] Furthermore, step S1 specifically includes:

[0016] S11. Multi-walled carbon nanotube powder is added to anhydrous ethanol solvent containing silane coupling agent, mixed evenly, and then placed in an ice-water bath. The mixed solution is then subjected to high-frequency pulse ultrasonic treatment and vacuum constant temperature drying to evaporate the anhydrous ethanol, thereby obtaining surface-activated carbon nanotube filler.

[0017] S12. Select industrial sandpaper with random micro-rough protrusions, ultrasonically clean and dry it, and uniformly spray a release agent on the surface of the dried industrial sandpaper to obtain a sandpaper template with a micro-nano structure; uniformly mix and degas the surface-activated carbon nanotube filler with the polymer elastomer substrate material to obtain a conductive slurry; uniformly coat the conductive slurry on the sandpaper template and heat and cure it; after complete heat curing, mechanically peel the flexible film formed by the conductive slurry from the sandpaper template to obtain a flexible sensing unit.

[0018] S13. Cut the prepared flexible sensing unit into a standard rectangular test strip, and connect signal transmission electrodes to both ends of the standard rectangular test strip to obtain a flexible sensor. Then connect the flexible sensor to the hardware acquisition circuit and seamlessly integrate it into the wearable carrier to obtain a wearable flexible sensing device.

[0019] Further, in step S11, the multi-walled carbon nanotube powder has an outer diameter of 8-15 nm, a length of 30-50 μm, and a mass fraction of 5%, the silane coupling agent has a mass fraction of 1.5%, the high-frequency pulse ultrasonic treatment has a power of 400 W, and the time is 120 minutes with 3 seconds of ultrasonic treatment followed by 2 seconds of intermittent treatment. In step S12, 1000-mesh silicon carbide industrial sandpaper is selected, and the A and B components of the polymer elastomer substrate material Ecoflex 00-30 silicone rubber are weighed at a mass ratio of 1:1 and mixed with the surface-activated carbon nanotube filler. The conductive paste is uniformly coated on the sandpaper template with a thickness of 150 μm.

[0020] Furthermore, in step S13, the hardware acquisition circuit adopts a high-precision miniature Wheatstone bridge, with the flexible sensor serving as the dynamic resistor of one arm of the bridge, and the other three arms configured as high-precision fixed surface mount resistors. A DC constant voltage source is used as the excitation voltage of the bridge. The differential voltage signal output by the bridge is amplified by a pre-amplifier and then input to an analog-to-digital converter chip to obtain a digital signal. The digital signal is then wirelessly transmitted in real time to a host computer or mobile terminal in the form of a data packet with a microsecond-level timestamp through a microcontroller with integrated Bluetooth protocol.

[0021] Furthermore, in step S2, different exercise loads include sitting still, walking at a constant speed, turning the head left and right, and nodding up and down; different breathing states include calm breathing, deep breathing, rapid breathing, or coughing.

[0022] Furthermore, step S4 specifically includes:

[0023] S41. Construct a multi-scale one-dimensional convolutional neural network module as a feature extraction module for the hybrid decoupled recognition network, used to extract local morphological features and reduce dimensionality;

[0024] S42. Construct a bidirectional long short-term memory network module as a virtual compensator for the hybrid decoupled recognition network, which is used to perform bidirectional temporal rhythm decoupling and material hysteresis error compensation on the output of the multi-scale one-dimensional convolutional neural network module.

[0025] S43. Construct a fully connected layer and a Softmax activation function layer as the output layer of the hybrid decoupled recognition network, which is used to map the high-dimensional features output by the bidirectional long short-term memory network module to the dimension and the total number of target categories. Strictly corresponding classification vectors It calculates the classification probability of each category and outputs the category label corresponding to the maximum probability as the final state determination result of the current sliding window sequence;

[0026] S44. Determine the model generalization constraints, loss function, and training strategy for training the hybrid decoupled recognition network.

[0027] Further, in step S41, the multi-scale one-dimensional convolutional neural network module is composed of two layers of one-dimensional convolutional layers and max pooling layers with different receptive fields, which are alternately cascaded. The kernel size of the first convolutional layer is set to 1×9, the number of output channels is 32, and the stride is 1. The kernel size of the second convolutional layer is set to 1×3 and the number of output channels is 64. After each convolutional operation, a max pooling layer with a kernel size of 1×2 and a stride of 2 is connected.

[0028] In step S42, the bidirectional long short-term memory network module receives the high-dimensional temporal feature vector after dimensionality reduction and flattening by the multi-scale one-dimensional convolutional neural network module, and its forward hidden layer follows the time sequence. Scanning, reverse hidden layer along time series The system scans and uses future contextual information to verify and correct the hysteresis bias at the current moment. This represents the total time step;

[0029] In step S44, the model constraints are as follows: a Dropout regularization layer is introduced between the bidirectional long short-term memory network module and the fully connected layer, with a dropout rate of 0.4; the loss function is a cross-entropy loss function with an L2 weight decay penalty term; and the training strategy is as follows: during the network gradient update stage, the Adam adaptive momentum optimizer is used, with an initial learning rate set to 0.001, and a cosine annealing learning rate decay strategy is employed.

[0030] Furthermore, step S3 specifically includes the following steps:

[0031] S31. Perform frequency domain denoising on the received original signal sequence to obtain a denoised signal sequence;

[0032] S32. An adaptive multi-scale overlapping sliding window is used to segment and reconstruct the denoised signal sequence to obtain a structured three-dimensional temporal feature matrix. The data overlap rate of the sliding window is 75%.

[0033] S33. Normalize each of the structured three-dimensional temporal feature matrices to obtain temporal feature samples, and combine all temporal feature samples to obtain a temporal feature sample set.

[0034] The present invention also provides a breathing state recognition system based on flexible deformation sensing, the key feature of which is: an intelligent agent is provided to implement the breathing state recognition method based on flexible deformation sensing.

[0035] The present invention also provides a computer-readable storage medium, the key feature of which is that computer-executable code is stored thereon, and when the executable code is executed, the breathing state recognition method based on flexible deformation sensing or the breathing state recognition system based on flexible deformation sensing is executed.

[0036] This invention provides a breathing state recognition method and system based on flexible deformation sensing. Addressing key pain points of existing flexible strain sensors, such as insufficient initial sensitivity in monitoring minute breathing signals, significant hysteresis effects of polymer elastomers leading to nonlinear distortion of time-series waveforms, and difficulty in effectively decoupling high-amplitude human motion artifacts in complex real-world wearing scenarios, this invention innovatively constructs a micro-nano hierarchical inverted pyramid structure on the surface of a carbon nanotube / Ecoflex composite conductive elastomer by introducing a low-cost industrial sandpaper template inversion process. This utilizes stress concentration effects to significantly enhance the contact resistance variation under extremely small deformations, greatly improving the physical sensing signal-to-noise ratio and sensitivity coefficient in the initial strain stage. Furthermore, a deep CNN-BiLSTM dual-architecture network is constructed at the signal processing and recognition ends, utilizing one-dimensional convolutional neural networks... The local spatial morphological distortion features of the waveform are accurately scanned and extracted by a network module (CNN). Then, a bidirectional long short-term memory network (BiLSTM) is used to perform forward and reverse time series scanning to deeply learn the nonlinear response law of flexible materials, so as to fit and compensate for the hysteresis effect caused by viscoelasticity from the algorithm end. Finally, the periodic weak respiratory rhythm features hidden in the high amplitude low frequency baseline drift are cleverly decoupled from the non-stationary large body movement interference features by high-dimensional mathematics. This successfully achieves high-precision stable recognition of physiological state under the background of strong artifact noise such as the subject walking, stepping, and violent head turning. It effectively improves the environmental adaptability and long-term operational robustness of flexible wearable electronic devices in complex dynamic monitoring scenarios, and provides reliable technical support for the next generation of intelligent medical and health monitoring and human-computer interaction systems. Attached Figure Description

[0037] Figure 1This is an overall flowchart of the dynamic breathing and human motion state recognition method based on microstructure flexible deformation sensing provided in the embodiments of the present invention;

[0038] Figure 2 This is a schematic diagram illustrating the fabrication process and deformation sensing physical mechanism of the microstructure conductive polymer composite flexible sensor provided in this embodiment of the invention.

[0039] Figure 3 This is a schematic diagram of a dense inverted pyramid-shaped random protrusion structure provided in an embodiment of the present invention;

[0040] Figure 4 This is a diagram of a CNN-BiLSTM-based temporal modeling and decoupling recognition network architecture provided in an embodiment of the present invention. Detailed Implementation

[0041] The embodiments of the present invention are described in detail below with reference to the accompanying drawings. The embodiments are given for illustrative purposes only and should not be construed as limiting the present invention. The accompanying drawings are for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0042] The breathing state recognition method based on flexible deformation sensing provided in this invention embodiment, such as... Figure 1 As shown in the flowchart, this embodiment includes the following steps:

[0043] S1. Fabrication of a wearable flexible sensing device with a micro / nano inverted pyramid structure: Silane coupling agent is used to perform surface functionalization grafting modification on carbon-based nano-conductive fillers to enhance their dispersibility in the substrate; then, through template reversal process, the modified conductive fillers are thoroughly mixed and degassed with a polymer elastomer substrate and then composited. The composite is then heated and cured on the surface of an industrial sandpaper template with random micro-rough protrusions. After peeling, a flexible sensing unit is obtained that completely replicates the micro-geometric features of the template (such as the micro / nano inverted pyramid structure). Then, a flexible sensor is fabricated based on the flexible sensing unit, and the flexible sensor is encapsulated and seamlessly integrated into wearable carriers such as smart masks or flexible joint patches to obtain a wearable flexible sensing device.

[0044] S2. Acquire multimodal time-series sensing signals: Attach the wearable flexible sensing device to the human body's mouth and nose breathing airflow area and target joints; under complex dynamic interference scenarios where the human body is under different motion loads (such as constant speed walking, head turning, talking, etc.), use a high-precision analog-to-digital conversion circuit to collect in real time mixed time-series resistance change rate signals containing weak periodic breathing characteristics and high-amplitude non-stationary motion artifacts under different breathing states (calm breathing, deep breathing, rapid breathing, and coughing, etc.) at a set sampling frequency;

[0045] S3. Preprocessing multimodal time-series sensing signals: The original one-dimensional time-series signal is sequentially subjected to digital low-pass filtering and Z-Score normalization to filter out high-frequency environmental noise and eliminate baseline resistance drift error caused by individual wearing differences; then, an adaptive sliding window technique with a specific high overlap rate is used to segment and reconstruct the continuous signal into a structured three-dimensional time-series feature matrix containing phase continuity according to the set window length (which needs to cover the entire respiratory cycle) and sliding step size.

[0046] S4. Construct and train a hybrid decoupled recognition network based on time-series modeling: Construct a deep learning model that includes a multi-layer one-dimensional convolutional neural network (1D-CNN) module and a bidirectional long short-term memory network (Bi-LSTM) module, and train it using a structured three-dimensional time-series feature matrix under different breathing states; wherein, the multi-scale receptive field in the one-dimensional convolutional neural network module is used to accurately extract the local spatial morphological distortion features of the signal (such as high-frequency mutations caused by coughing), while the bidirectional long short-term memory network module is used to fit the nonlinear hysteresis effect of flexible materials through forward and reverse time series scanning and extract the long-range time-series dependence of the signal, thereby achieving deep mathematical decoupling of periodic respiratory rhythm features and non-stationary motion interference features in the feature space;

[0047] S5. Real-time Respiratory Status Identification: After real-time acquisition of sensor signals using a wearable flexible sensing device and the same preprocessing procedure, the preprocessed structured three-dimensional temporal feature matrix is ​​continuously input into the trained hybrid decoupled recognition network. The fully connected layer at the end of the hybrid decoupled recognition network and the Softmax classification activation function are used to map it into the probability distribution vector of the target category. The label corresponding to the maximum probability value is selected, and the respiratory status identification results of calm breathing, deep breathing, rapid breathing or coughing are output in real time.

[0048] This invention addresses the problems of insufficient sensitivity of existing flexible sensors in monitoring minute respiration, significant waveform distortion caused by the hysteresis effect of polymer elastomer materials, and difficulty in effectively decoupling human motion artifacts in complex dynamic scenarios. It provides a breathing state recognition method based on flexible deformation sensing. This method significantly improves initial sensitivity by introducing an industrial sandpaper template reversal process to fabricate sensing units with a micro-nano hierarchical inverted pyramid structure on the surface. Then, multimodal time-series signals are acquired and a feature matrix is ​​constructed using overlapping sliding windows. Subsequently, a one-dimensional convolutional neural network is used to accurately extract the local spatial morphological distortion of the waveform, and a bidirectional long short-term memory network is used to fit and compensate for the hysteresis effect of the flexible material through forward and reverse time-series scanning. Finally, this method deeply decouples the periodic, weak respiratory rhythm features from the non-stationary, large-amplitude body motion interference features through deep mathematical decoupling, outputting accurate state recognition results even against a background of strong artifact noise.

[0049] The following provides a more detailed explanation of each step.

[0050] (1) Step S1: Fabricate a wearable flexible sensing device with a micro-nano inverted pyramid structure

[0051] This step aims to fundamentally solve the problem of traditional flexible sensors' limitations in minute strain ranges (such as strain caused by facial muscle micro-movements or respiratory airflow) from the perspectives of physical materials and microstructure. The technical bottleneck is low sensitivity. Figure 2 This is a flowchart illustrating the fabrication of a flexible sensing unit. (Reference) Figure 2 Step S1 specifically includes the following steps:

[0052] S11. Preparation of surface-activated carbon nanotube filler: Multi-walled carbon nanotube powder is added to anhydrous ethanol solvent containing silane coupling agent, mixed evenly, and then placed in an ice-water bath. The mixed solution is then subjected to high-frequency pulse ultrasonic treatment and vacuum constant temperature drying to evaporate the anhydrous ethanol, thereby obtaining surface-activated carbon nanotube filler.

[0053] Step S11 specifically involves selecting high-purity multi-walled carbon nanotubes (MWCNTs) with an outer diameter of 8-15 nm and a length of 30-50 μm as conductive fillers. 5% (w / w) of MWCNT powder is directly added to anhydrous ethanol solvent containing 1.5% (w / w) of silane coupling agent (KH-550, i.e., 3-aminopropyltriethoxysilane). The mixture is placed in an ice-water bath and subjected to high-frequency pulsed ultrasonic treatment at 400W power using an ultrasonic cell disruptor (3 seconds of sonication followed by a 2-second pause) for 120 minutes. Under this high-intensity cavitation effect, the silane coupling agent fully coats the surface of the carbon nanotubes. Through physical adsorption and partial hydrogen bonding, the thermodynamic compatibility and dispersion uniformity of MWCNTs in the polymer elastomer substrate are significantly improved. After ultrasonication, the suspension is placed in an 80℃ vacuum drying oven for constant temperature drying to evaporate the anhydrous ethanol, yielding surface-activated carbon nanotube fillers.

[0054] This step effectively reduces the conductive percolation threshold, enhances the physical entanglement between the filler and the elastomer polymer network, and reduces internal frictional dissipation during stretching and release from the intrinsic material level.

[0055] S12. Preparation of flexible sensing unit: Select industrial sandpaper with random micro-rough protrusions, ultrasonically clean and dry it, and uniformly spray a release agent on the surface of the dried industrial sandpaper to obtain a sandpaper template with micro-nano structure; uniformly mix and degas the surface-activated carbon nanotube filler with the polymer elastomer substrate material to obtain a conductive slurry; uniformly coat the conductive slurry on the sandpaper template and heat and cure it to form a mold; after complete thermal curing, carefully mechanically peel the flexible film formed by the conductive slurry from the sandpaper template to obtain the flexible sensing unit.

[0056] Step S12 abandons complex and costly photolithography or chemical etching processes, innovatively selecting 1000-mesh silicon carbide industrial sandpaper as the negative electrode physical template for the micro / nano structure. This mesh size of sandpaper has moderate surface roughness and a dense distribution of natural peaks, making it ideal for enhancing the micro-stress concentration effect. The sandpaper template is ultrasonically cleaned for 15 minutes with deionized water and anhydrous ethanol respectively, then dried with nitrogen gas, and a very thin layer of release agent is uniformly sprayed onto its rough surface to prevent polymer adhesion. Components A and B of the polymer elastomer substrate material Ecoflex 00-30 silicone rubber are weighed in a 1:1 mass ratio, and the aforementioned dried activated MWCNTs are added. Using a planetary vacuum degassing mixer, under an extremely high negative pressure environment of -98 kPa, the mixture is alternately stirred at 1500 rpm for 15 minutes to ensure a uniform distribution of the three-dimensional conductive network and the absence of any microscopic bubble defects.

[0057] A high-precision thin-film coating machine was used to evenly coat the mixed and degassed conductive slurry onto a sandpaper template, with the coating thickness precisely controlled to 150μm. To eliminate internal residual stress and ensure the uniformity of the cross-linked network, a stepped curing process was adopted: first, it was allowed to stand at room temperature for 30 minutes to level, and then it was placed in a constant temperature drying oven at 60℃ for 4 hours to heat cure and allow it to fully cross-link and form.

[0058] Scanning electron microscopy (SEM) morphology characterization confirmed that the contact surface of this flexible sensing unit perfectly replicated the micro-nano morphology of 1000-grit sandpaper through pure physical imprinting, exhibiting a dense, inverted pyramid-shaped, randomly protruding structure with heights ranging from 10-30 μm. Figure 3 As shown. After the electrodes are brought out, a 30μm thick pure Ecoflex insulating film is coated on both the surface and the back as a protective layer to completely isolate interference from human sweat, environmental humidity and external oxidation, ensuring long-term signal stability.

[0059] Its underlying physical sensing mechanism lies in the synergistic sensing mechanism of Holm contact resistance theory and quantum tunneling effect. When the surface of the sensor with a micro-nano inverted pyramid structure is subjected to extremely weak respiratory airflow pressure or facial muscle traction deformation, local stress concentration easily occurs at the tip of the protruding structure. This minute deformation causes the originally separated micro-nano protrusions to approach each other and make mechanical contact, resulting in a dramatic nonlinear increase in the effective contact area. At the same time, the tunnel spacing between adjacent carbon nanotubes decreases sharply, leading to an exponential decrease in the overall contact resistance. This purely physical "micro-contact control mechanism" enables the sensor to exhibit not only an extremely low detection limit (e.g., 0.1% micro-strain response) within the 0-5% micro-strain range, but also a Gauge Factor (GF) that is more than ten times higher than that of traditional smooth structure sensors, perfectly meeting the stringent hardware requirements for non-invasive, high-precision dynamic respiratory monitoring.

[0060] S13. Fabrication of a wearable flexible sensing device: The fabricated flexible sensing unit is cut into a standard rectangular test strip, and signal transmission electrodes are connected to both ends of the standard rectangular test strip to obtain a flexible sensor. The flexible sensor is then connected to the hardware acquisition circuit and seamlessly integrated into wearable carriers such as smart masks or flexible joint patches to obtain a wearable flexible sensing device.

[0061] The specific steps are as follows:

[0062] The flexible sensing unit prepared above was precisely cut into standard rectangular test strips, each 10 mm long and 30 mm wide. To eliminate rigid stress concentration at the electrode connections, extremely fine multi-strand flexible enameled copper wires were connected at both ends using highly conductive silver conductive adhesive as signal transmission electrodes. A pure Ecoflex polymer elastomer with a thickness of approximately 20 μm was then coated around the wires using a spin-coating process for secondary insulation and anti-oxidation encapsulation, completely isolating them from the corrosive effects of human sweat and environmental humidity, thus obtaining a flexible sensor.

[0063] To deeply filter out environmental electromagnetic interference and amplify weak respiratory electrical signals without distortion, the hardware acquisition circuit employs a high-precision miniature Wheatstone Bridge. The flexible sensor serves as the dynamic resistor for one arm of the bridge, while the other three arms are configured with high-precision fixed surface-mount resistors with extremely low temperature coefficients. The system uses an extremely low DC constant voltage source (e.g., 3.3V) as the bridge excitation voltage to avoid the Joule heating effect in the carbon nanotube conductive network caused by high input voltage, which would lead to thermal resistance drift. The weak differential voltage signal output by the bridge is pre-differentially amplified by a low-noise instrumentation amplifier with high common-mode rejection ratio (CMRR) (e.g., INA333) before being input to a 24-bit high-resolution analog-to-digital converter chip (e.g., ADS1256). The 24-bit ADC provides extremely high quantization resolution, capable of accurately resolving extremely weak voltage fluctuations at the microvolt (μV) level. The system's digital sampling frequency is strictly set to 50Hz. According to the Nyquist sampling theorem, this frequency can not only completely and without distortion reproduce the normal respiratory rhythm fluctuations in the 0.2-0.5Hz range, but also accurately capture transient signals such as coughs and speech containing high-frequency harmonic components of 10-20Hz. The acquired high-fidelity digital signals are wirelessly transmitted in real time to the host computer or mobile terminal via a microcontroller integrating the Bluetooth Low Energy (BLE) 5.0 protocol, in the form of data packets with microsecond-level timestamps. The timestamp synchronization mechanism effectively prevents packet loss and out-of-order issues during wireless transmission.

[0064] Then, the hardware and software acquisition circuitry is attached to the inside of the smart mask, in the area directly facing the airflow from the mouth and nose (primarily sensing the minute strain on the thin film caused by weak airflow impact), or at key joints of the human body (such as elbows and knees, areas of high strain). The use of medical tape effectively avoids additional artifact noise caused by sensor slippage or partial peeling during strenuous exercise.

[0065] Step S1 constructs a micro-nano inverted pyramid structure through template inversion technology, which significantly improves the sensitivity and signal-to-noise ratio of the flexible sensor in the small deformation range (<1%) by utilizing the stress concentration effect, thus solving the problem of insufficient response of traditional sensors to weak breathing deformation from a physical perspective.

[0066] (2) Step S2: High-fidelity acquisition of multimodal time-series sensing signals

[0067] This step aims to convert physical deformations into digital signals without distortion and to construct a robust training dataset containing strong noise features, specifically:

[0068] To obtain a deep learning recognition model with extremely high robustness and generalization ability, an experimental data acquisition protocol incorporating multiple extreme dynamic disturbances was designed. Multiple subjects of different genders, ages, and body mass indexes (BMI) were recruited and performed the following task sequence while wearing smart masks (each task was recorded for 3 to 5 minutes to obtain sufficient time-window samples):

[0069] (1) Static baseline group: Subjects performed calm breathing, deep breathing, rapid breathing and intermittent coughing in a quiet sitting state to construct a baseline physiological characteristic sample database;

[0070] (2) Walking Interference Group: Subjects walked at a constant speed of 5 km / h on a commercial treadmill while maintaining normal, calm breathing. The mechanical impact force generated by the feet hitting the ground during walking was transmitted to the face through the human skeleton, producing regular vibrations with a frequency of approximately 1.5-2 Hz. This vibration caused a strong periodic low-frequency baseline drift in the sensor with an amplitude far exceeding that of the breathing signal, which was used to simulate regular motion interference during daily commutes;

[0071] (3) Complex head movement interference group: Subjects, while sitting upright, continuously and irregularly performed large-amplitude left and right head turning and up and down head nodding movements at a frequency of about 1 Hz. Due to the direct and intense traction of the neck and maxillofacial muscles, this movement not only changed the contact boundary conditions between the mask and the face, but also generated a huge instantaneous nonlinear strain pulse signal, which could easily drown out the real weak breathing waveform.

[0072] During all dynamic interference data acquisition processes, high-definition cameras were used simultaneously to record the subjects' real actions and physiological states, providing extremely accurate ground truth timestamp alignment for temporal segmentation and supervised learning in subsequent deep network model training.

[0073] Step S2 involves collecting mixed signals containing breathing and motion artifacts under various dynamic interference scenarios to construct an original dataset covering extreme conditions, providing high-fidelity data support for subsequent feature decoupling and model training.

[0074] (3) Step S3: Time series signal preprocessing and multi-scale feature window construction

[0075] This step aims to transform the raw one-dimensional digital signal, containing massive amounts of environmental and motion noise, into a high-quality, structured feature matrix suitable for ingestion by deep learning networks. Specifically, it includes:

[0076] S31. Zero-phase-shift digital filtering denoising: The received original signal sequence (containing a one-dimensional time-series resistance change rate with complex dynamic environmental noise) is denoised in the frequency domain to obtain a denoised signal sequence.

[0077] To avoid frequency domain aliasing during feature extraction in subsequent deep learning networks, step S31 performs frequency domain denoising using a designed high-order (e.g., fourth-order) Butterworth low-pass digital filter. The Butterworth filter is chosen to leverage its amplitude-frequency response with maximum flatness within its passband, effectively avoiding spurious feature injection caused by filter ripple. Its cutoff frequency is set to 20Hz. This threshold effectively filters out 50Hz power frequency noise caused by environmental electromagnetic interference and sensor thermal noise (random white noise), while fully preserving normal breathing rhythms (0.1-0.5Hz), walking vibration noise (1-3Hz), and critical transient signals such as coughing and speech containing high-frequency harmonic components up to approximately 15Hz. Furthermore, to prevent signal distortion on the time axis caused by filtering, this step employs a forward-backward filtering algorithm to eliminate the phase delay introduced by traditional causal filters, ensuring absolute alignment of the time-series signal on the time axis.

[0078] S32. Multi-scale overlapping sliding window matrix reconstruction: Adaptive multi-scale overlapping sliding window is used to segment and reconstruct the denoised signal sequence to obtain a structured three-dimensional temporal feature matrix.

[0079] To address the dynamic pattern recognition requirements of non-stationary breathing and sudden movements in the human body, step S32 employs an adaptive multi-scale overlapping sliding window technique to segment and structurally reconstruct continuous digitized one-dimensional time series. Considering that a complete adult deep breathing cycle (including inhalation, exhalation, and a brief pause) can last up to approximately 4 seconds, the sliding window length W is strictly set to 200 sampling points (corresponding to a 4-second time span at a 50Hz sampling rate). This theoretically ensures that each independent time window contains at least one complete breathing fluctuation waveform cycle, effectively avoiding the loss of temporal features due to waveform truncation. Simultaneously, to capture instantaneous changes in physiological state (such as a sudden violent cough or head rotation impact during calm breathing), the sliding step size S is set to 50 sampling points (i.e., a 1-second step). This parameter configuration maintains a data overlap rate of up to 75% between adjacent time windows, greatly ensuring the phase continuity and smooth transition of the feature space during state transitions in non-stationary signals. After this segmentation process, the length is... A one-dimensional continuous-time signal sequence is reconstructed into a dimension of The structured three-dimensional temporal feature matrix, in which This is the total number of window samples generated, thus satisfying the dimensionality requirements of the input tensor for deep temporal convolutional networks.

[0080] S33. Local Z-Score Standardization and Covariate Shift Suppression: Normalize each structured 3D temporal feature matrix to obtain temporal feature samples, and combine all temporal feature samples to obtain a temporal feature sample set.

[0081] Step S33 performs statistical local Z-Score zero-mean standardization on the data matrix X of each independently segmented time window. The calculation formula is as follows:

[0082] ,

[0083] in, and These are the local signal mean and local standard deviation of the signal sequence of 200 sampling points within the current single time window, respectively.

[0084] In real-world flexible wearable applications, differences in facial bone contours among subjects, the randomness of the tightness of the mask straps each time a mask is worn, changes in the ambient temperature field, and the static stress relaxation of polymer elastomer materials under continuous stress all contribute to significant and unpredictable low-frequency drift in the absolute baseline of the sensor's output resistance. Traditional global normalization methods cannot eliminate this local drift; however, the local Z-score normalization operation employed in this invention mathematically forces the removal of this absolute resistance error caused by physical baseline drift, uniformly transforming the feature distribution entering the subsequent neural network into a standard normal distribution with a mean of 0 and a variance of 1. This core processing not only forces the network model to focus solely on the "relative fluctuation morphology and dynamic distortion rate" of the waveform during feature extraction, fundamentally shielding the absolute resistance magnitude from interference with classification weights, but also effectively suppresses internal covariate shift, significantly accelerating the gradient descent convergence speed during backpropagation and preventing gradient explosion or vanishing phenomena in deep training.

[0085] Step S3 eliminates baseline drift by filtering and Z-score normalization, and reconstructs the structured temporal feature matrix using a high overlap sliding window, providing high-quality, phase-continuous feature input for the deep network.

[0086] (4) Step S4: Construct and train a hybrid decoupled recognition network based on time-series modeling.

[0087] This step aims to mathematically decouple aliased high-amplitude motion artifacts from low-amplitude physiological signals in the feature space by constructing an end-to-end deep temporal model, and to compensate for nonlinear physical errors of materials from the algorithm end. Figure 4 This is a structural diagram of a hybrid decoupled recognition network. (Reference) Figure 4 This step specifically includes:

[0088] S41. Construct a feature extraction module for a hybrid decoupled recognition network: Construct a multi-scale one-dimensional convolutional neural network (1D-CNN) module for efficient extraction and dimensionality reduction of local morphological features.

[0089] This module consists of two alternating cascaded one-dimensional convolutional layers and max-pooling layers with different receptive fields, in order to filter multi-band noise and anchor key morphological features.

[0090] The first convolutional layer employs a large receptive field design, with a kernel size of 1×9, 32 output channels, and a stride of 1. Its main function is to act as a feature-level low-pass filter, smoothing out spikes in the signal caused by extremely fine muscle tremors or environmental noise, and accurately extracting the large-scale trend contours of the waveform (such as the envelope of a breathing wave).

[0091] The second convolutional layer employs a small receptive field design, with a kernel size of 1×3 and the number of output channels doubled to 64. Its main function is to keenly capture local abrupt changes in the smoothed feature map caused by instantaneous stress pulses (such as the steep slope produced when coughing or the high-frequency subtle ripples produced when speaking).

[0092] The forward propagation mathematical operations of convolutional layers are uniformly represented as follows:

[0093] ,

[0094] in, For the first Each output feature map at location activation value, The total number of channels for the input features. For the first The one-dimensional convolution kernel acts on the first... The weight coefficient vector or matrix corresponding to each input channel; The input structured 3D feature matrix is ​​in the first... On each channel, from the location Crossing to Local continuous data slices, i.e., sliding window feature data covered by the receptive field of the convolution kernel; The spatial size of the one-dimensional convolution kernel. For the corresponding to the first The first one-dimensional convolution kernel determines the... The bias term for the overall baseline offset of each output feature map; The Rectified Linear Unit (RLU) is an activation function used to introduce nonlinear mappings into neural networks, thereby giving the network the ability to process complex nonlinear deformation signals.

[0095] After each convolutional operation, a max-pooling layer with a kernel size of 1×2 and a stride of 2 is applied. This layer not only achieves a 50% downsampling of the feature space dimension to significantly reduce computational overhead, but more importantly, it endows the network with local translation invariance. In real-world wearable scenarios, slight slippage of the mask or patch can cause a small shift in the waveform along the time axis. Max-pooling ensures that as long as key features (such as peaks) appear within the pooling window, the network can extract them stably, thereby greatly enhancing the model's robustness to slight sensor displacement.

[0096] S42. Construct a virtual compensator for a hybrid decoupled identification network: Construct a bidirectional long short-term memory network (Bi-LSTM) module as an algorithm-level virtual compensator for bidirectional temporal rhythm decoupling and material hysteresis error compensation.

[0097] Due to the inherent viscoelasticity of polymer elastomers (such as Ecoflex substrates), when a flexible sensor is mechanically stretched and then released, the network rearrangement of the polymer chains requires relaxation time. This causes the resistance recovery trajectory to not coincide with the stretching trajectory, but instead form a nonlinear hysteresis loop. In continuous high-frequency dynamic respiratory monitoring, the residual internal stress from the previous deformation is superimposed on the next deformation, resulting in severe distortion of the waveform baseline.

[0098] The Bi-LSTM module receives high-dimensional temporal feature vectors after dimensionality reduction and flattening by 1D-CNN. Its forward hidden layers follow the time series... ( The total time step is scanned to capture the dynamic evolution during the deformation loading process; the inverse hidden layer follows the time series. The system scans and extracts the recovery features of the deformation unloading process, and uses future context information to verify and correct the hysteresis deviation at the current moment.

[0099] Its internal units control the flow and interruption of residual deformed memories through a sophisticated gating mechanism. Specifically, this is achieved through the forgetting gate. Control the proportion of residual deformation memory from the previous time step:

[0100] ,

[0101] via input gate With candidate cell state Filter and write the new strain features at the current time step:

[0102] ,

[0103] ,

[0104] Current state of deep cells Evolution and update of output gate Final hidden state The calculation is expressed as:

[0105] ,

[0106] ,

[0107] ,

[0108] in, To constrain the output to the (0,1) interval using the sigmoid activation function, It is the hyperbolic tangent function. Enter the current time. This is the hidden state from the previous moment. This represents the state of the deep cells inside the cell layer at the previous moment. , , , These are the weight matrices and bias vectors learned through backpropagation of the forget gate, input gate, candidate cell state, and output gate, respectively. The output at the final time step is formed by concatenating the forward and backward hidden states.

[0109] Relying on this bidirectional compensation and gated timing mining capability, the network can autonomously learn and fit the complex physical laws of polymer viscoelasticity, achieve virtual hysteresis compensation without hardware addition at the algorithm level, and completely strip away the weak respiratory rhythm superimposed on low-frequency large-amplitude walking vibration interference.

[0110] S43. Construct the output layer of the hybrid decoupled recognition network: Determine that the output layer at the end of the hybrid decoupled recognition network is a fully connected layer and a softmax activation function layer.

[0111] The fully connected layer at the end of the hybrid decoupled recognition network performs a linear spatial transformation on the high-dimensional temporal fusion feature vector from the Bi-LSTM module, mapping the high-dimensional features to the dimension and the total number of target categories. Strictly corresponding Logits classification vector Subsequently, the vector is input into the Softmax activation function layer, where it is normalized and transformed into a relative probability distribution for various physiological and motor states (such as calm breathing, deep breathing, rapid breathing, and coughing). Its mathematical conversion formula is:

[0112] ,

[0113] in, The first of the output vectors of the fully connected layer One portion, This indicates that the current time window belongs to the [number]th [time window]. The system predicts the probability of physiological states. It uses the Argmax function to extract the category label corresponding to the maximum value in the probability distribution vector, which is used as the final state determination result of the current sliding window sequence. The result is then output to the mobile monitoring terminal or cloud medical system in real time with extremely low latency via wireless communication protocols such as Bluetooth, triggering corresponding state recording or abnormal alarms.

[0114] S44. Determine the model generalization constraints, loss function, and training strategy.

[0115] To prevent deep networks from overfitting in complex dynamic environments (i.e., rote memorization of specific walking cadences or head-shaking noise patterns from the training set), a Dropout regularization layer is forcibly introduced between the Bi-LSTM layer and the final Softmax classification fully connected layer, with a dropout rate of 0.4. This operation randomly severs 40% of the neuron connections in each forward propagation training iteration, breaking the co-adaptation between neurons and forcing the network to learn more generalized and essential physiological mapping features.

[0116] Simultaneously, a cross-entropy loss function with an L2 weight decay penalty term is adopted:

[0117] ,

[0118] in, The total number of samples to be processed in a batch (Batch Size). The total number of status categories. It is a sample In category The One-Hot tag on the platform. It is the probability distribution predicted by the model. Let be the regularization penalty coefficient, set to 1e-4. L2 regularization term. middle, The hybrid decoupled recognition network represents the multi-scale one-dimensional convolutional neural network module, the bidirectional long short-term memory network module, and all trainable neuron connection weights in the fully connected layers. The penalty term calculates the sum of squares of all weight parameter values ​​in the network, i.e., the squared L2 norm of the weight matrix. By introducing this additional penalty based on weight size in addition to the cross-entropy loss, the algorithm continuously applies a decay constraint to larger weight values ​​during gradient descent, effectively limiting the overall network weight matrix. The norm growth forces the model to learn smoother, more dispersed feature mapping relationships, thereby mathematically avoiding the extreme overfitting or overexertion of individual neuron parameters in the model due to excessively large values ​​in response to occasional high-amplitude motion artifact noise.

[0119] During the network gradient update phase, the Adam (Adaptive Moment Estimation) adaptive momentum optimizer is employed. Compared to the traditional SGD algorithm, Adam combines unbiased estimations of first-order and second-order momentum, enabling the adaptive allocation of independent update step sizes for different parameters. The initial learning rate is set to 0.001, coupled with a cosine annealing learning rate decay strategy. This strategy causes the learning rate to decrease periodically in the form of a cosine function with each training epoch. The larger learning rate in the early stages helps the model quickly escape the local minima trap introduced by drastic motion noise, while the gradually decreasing minimum learning rate in the later stages ensures that the model accurately converges to the global optimum at the bottom of the complex high-dimensional error surface.

[0120] S45. Based on the determined hybrid decoupled recognition network architecture, model generalization constraints, loss function and training strategy, the hybrid decoupled recognition network is trained using the temporal feature sample set obtained in step S3.

[0121] Step S4 constructs a hybrid decoupled recognition network cascaded with 1D-CNN and BiLSTM, which decouples high-dimensional features of motion artifacts and respiratory rhythms, and effectively compensates for the nonlinear hysteresis error of flexible materials from the algorithm side, thus solving the problem of poor stability of respiratory state recognition under dynamic interference.

[0122] (5) Step S5: Accurate determination and real-time output of physiological and motor states

[0123] This step aims to apply the trained hybrid decoupled recognition network to actual breathing state recognition. By collecting and preprocessing sensor signals in real time, the signals are input into the trained hybrid decoupled recognition network, and the network outputs the recognition result of whether the current breathing state is calm breathing, deep breathing, rapid breathing, or coughing.

[0124] It should be noted that the various processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved. This embodiment does not impose any limitations on these steps.

[0125] In addition, this embodiment of the invention also provides a breathing state recognition system based on flexible deformation sensing, which is equipped with an intelligent agent for implementing the above-mentioned breathing state recognition method based on flexible deformation sensing.

[0126] In addition, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable code thereon. When the executable code is executed, the above-described breathing state recognition method or breathing state recognition system based on flexible deformation sensing is executed.

[0127] The embodiments described in this invention can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with the implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0128] Computer programs for implementing the methods and systems of the present invention may be written in any combination of one or more programming languages ​​and stored in a computer-readable storage medium. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0129] Computer-readable storage media can be tangible media that may contain or store computer programs for use by or in conjunction with an instruction execution system, apparatus, or device. Computer-readable storage media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, computer-readable storage media can be machine-readable signal media. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, optical disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0130] In summary, the breathing state recognition method and system based on flexible deformation sensing provided in this invention, with underlying materials and physical hardware as its core and upper-level data engineering and deep algorithms as its upper-level core, forms a complete adaptive anti-interference monitoring closed loop through extremely tight software and hardware co-design, achieving the following three major technological breakthroughs:

[0131] 1. Extremely high physical sensing resolution and front-end signal-to-noise ratio gain: By introducing 1000-grit industrial silicon carbide sandpaper as a reverse template, a dense micro-nano inverted pyramid hierarchical structure was constructed on the sensing surface of the conductive elastomer. This structure perfectly utilizes the abrupt change mechanism of micro-contact resistance in the initial stage of stress (tunneling effect and nonlinear expansion of micro-protrusion contact area), enabling the system to keenly capture deformation in the deformation initiation sensing stage. The ability to detect extremely subtle strains (such as shallow breathing airflow). This significant improvement in signal-to-noise ratio at the hardware front-end physical level lays an irreplaceable high-quality data foundation for feature extraction in subsequent deep learning algorithms;

[0132] 2. Superior resistance to strong dynamic interference and spatiotemporal feature decoupling: Abandoning traditional bandpass or bandstop hardware filtering, it innovatively adopts a dual network architecture of 1D-CNN and Bi-LSTM. 1D-CNN acts like a precise high-frequency scanner, accurately scanning across the time window to extract local high-frequency morphological spatial distortions (such as the spike pulse at the moment of coughing); while Bi-LSTM acts like a rhythm tracker, tracking low-frequency long-range rhythms across the time window (such as the fluctuating envelope of respiration). The deep fusion of spatiotemporal features in the hidden layer space successfully achieves complete decoupling and separation of severely aliased strong motion artifacts (such as gait vibrations and muscle tension) from genuine, subtle physiological rhythms in a mathematically high-dimensional space.

[0133] 3. Effectively fits and non-destructively compensates for the intrinsic nonlinear hysteresis of flexible polymer materials: By utilizing the Bi-LSTM network's forward and reverse dual scanning and multi-gated memory mechanism on the time series, an algorithmic "virtual viscoelastic hysteresis compensator" is essentially constructed in the model. This not only avoids the cumbersome process of adding complex damping materials to the sensor hardware, but also automatically corrects phase delays and waveform distortions caused by dynamic stress relaxation and residual strain of silicone rubber materials through online learning, ensuring signal fidelity during continuous high-frequency monitoring.

[0134] To fully verify the reliability of the method provided by this invention in extremely complex environments, a systematic comparative experiment was conducted between the method of this invention and traditional mainstream comparison methods (Support Vector Machine (SVM) algorithm, smooth sensor combined with unidirectional LSTM, and microstructure sensor combined with unidirectional LSTM) in various scenarios containing complex dynamic interference. Subjects were tested under static sitting, constant speed walking on a treadmill, and violent head movements, and the results are shown in Table 1.

[0135] Table 1: Comparative test results of the present invention and traditional identification methods

[0136]

[0137] As can be seen from Table 1, the static breathing recognition accuracy of the microstructure + CNN-BiLSTM method provided by this invention is as high as 98.9%, which is significantly better than the pure smooth sensor system (78.5%) and the system of microstructure combined with unidirectional LSTM (94.3%). This strongly proves that the microstructure design at the hardware level greatly improves the absolute amplitude and signal-to-noise ratio of the weak signal at the front end, and achieves the first breakthrough in perception at the physical level.

[0138] Furthermore, under extremely severe gait interference from a uniform 5km / h walking speed and irregular, violent head movements, the accuracy of traditional linear SVM algorithms based on manual feature extraction plummeted to below 50% (42.1% and 45.6%) due to the strong non-stationary and nonlinear characteristics of the motion noise, rendering the system essentially worthless for practical monitoring. In contrast, the method of this invention robustly maintained ultra-high state recognition accuracy of 96.5% and 95.8%, far exceeding other single-dimensional optimization schemes. This definitively demonstrates that the bidirectional joint spatiotemporal modeling of 1D-CNN and Bi-LSTM successfully decoupled and separated high-frequency transient muscle traction noise from severe low-frequency walking baseline drift.

[0139] Furthermore, while possessing the capability to automatically compensate for material hysteresis errors across the entire range, the system model's single real-time inference response time on the edge computing terminal is only 28ms. This extremely low latency fully meets the stringent industrial-grade requirements of modern wearable medical IoT devices for real-time physiological signal monitoring and ultra-robustness.

[0140] The extreme test data in Table 1 fully demonstrate that the method of the present invention has significant advantages over traditional methods in core indicators such as the limit of physical perception accuracy, ability to resist complex dynamic artifacts, lossless correction of hysteresis fitting, and robustness of system operation, and verifies its excellent comprehensive performance in real-world dynamic environments with strong interference.

[0141] In summary, the present invention provides a dynamic breathing and human motion state recognition method based on microstructure flexible deformation sensing. Addressing the shortcomings of existing technologies such as insufficient sensitivity, susceptibility to artifact interference, and material hysteresis, this method introduces a low-cost, mass-producible industrial sandpaper template reversal process to prepare a conductive polymer composite elastomer with a micro-nano hierarchical inverted pyramid structure, significantly improving the signal-to-noise ratio of sensing weak breathing deformations. Simultaneously, it acquires multimodal dynamic signals under all operating conditions and reconstructs the feature matrix using a high-overlap sliding window technique. Subsequently, a deeply optimized CNN-BiLSTM dual-architecture temporal network is constructed at the signal processing core. This network efficiently extracts local spatial morphological distortions of the waveform through a one-dimensional convolution module and performs lossless virtual hysteresis compensation by using a bidirectional temporal cyclic gating mechanism to deeply learn the nonlinear response law of the material. Ultimately, this method ingeniously and extremely efficiently decouples the low-frequency, weak, periodic real respiratory rhythm signal from the high-amplitude, non-stationary motion artifact signal of various parts of the body through deep feature decoupling. It successfully achieves continuous monitoring of physiological and motion states with extremely high precision under harsh conditions with strong interference, such as subjects walking and shaking their heads. This significantly improves the anti-interference ability, recognition stability, and clinical applicability of flexible wearable electronic devices in complex real-world environments.

[0142] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A respiratory state recognition method based on flexible deformation sensing, characterized in that, Including the following steps: S1. Surface functionalization grafting modification of carbon-based nano-conductive filler is carried out using silane coupling agent; after degassing by fully mixing and degassing the modified conductive filler with a polymer elastomer substrate through template reversal process, the composite is formed by heating and curing on the surface of industrial sandpaper template with random micro-rough protrusions, and after peeling, a flexible sensing unit is obtained. A flexible sensor is fabricated based on a flexible sensing unit and integrated into a wearable carrier to obtain a wearable flexible sensing device. S2. Attach the wearable flexible sensing device to the human body's mouth and nose breathing airflow area and target joint area; Under dynamic disturbance scenarios where the human body is under different exercise loads, the original one-dimensional time-series signals under different respiratory states are collected in real time. S3. After normalizing the original one-dimensional time series signal, it is segmented and reconstructed into a structured three-dimensional time series feature matrix containing phase continuity. S4. Construct a hybrid decoupled recognition network based on temporal modeling and train the hybrid decoupled recognition network using the structured three-dimensional temporal feature matrix; S5. The wearable flexible sensing device is used to collect sensing signals in real time and process them to obtain a structured three-dimensional temporal feature matrix. The structured three-dimensional temporal feature matrix is ​​continuously input into the hybrid decoupled recognition network after training. The hybrid decoupled recognition network outputs the breathing state recognition result in real time.

2. The breathing state recognition method based on flexible deformation sensing according to claim 1, characterized in that, Step S1 specifically includes: S11. Multi-walled carbon nanotube powder is added to anhydrous ethanol solvent containing silane coupling agent, mixed evenly, and then placed in an ice-water bath. The mixed solution is then subjected to high-frequency pulse ultrasonic treatment and vacuum constant temperature drying to evaporate the anhydrous ethanol, thereby obtaining surface-activated carbon nanotube filler. S12. Select industrial sandpaper with random micro-rough protrusions, ultrasonically clean and dry it, and uniformly spray a release agent on the surface of the dried industrial sandpaper to obtain a sandpaper template with a micro-nano structure; uniformly mix and degas the surface-activated carbon nanotube filler with the polymer elastomer substrate material to obtain a conductive slurry; uniformly coat the conductive slurry on the sandpaper template and heat and cure it; after complete heat curing, mechanically peel the flexible film formed by the conductive slurry from the sandpaper template to obtain a flexible sensing unit. S13. Cut the prepared flexible sensing unit into a standard rectangular test strip, and connect signal transmission electrodes to both ends of the standard rectangular test strip to obtain a flexible sensor. Then connect the flexible sensor to the hardware acquisition circuit and seamlessly integrate it into the wearable carrier to obtain a wearable flexible sensing device.

3. The breathing state recognition method based on flexible deformation sensing according to claim 2, characterized in that, In step S11, the multi-walled carbon nanotube powder has an outer diameter of 8-15 nm, a length of 30-50 μm, and a mass fraction of 5%. The silane coupling agent has a mass fraction of 1.5%. The high-frequency pulse ultrasonic treatment has a power of 400 W and a duration of 120 minutes, consisting of 3 seconds of ultrasonic treatment followed by 2 seconds of intermittent treatment. In step S12, 1000-mesh silicon carbide industrial sandpaper is selected. The A and B components of the polymer elastomer substrate material Ecoflex 00-30 silicone rubber are weighed at a mass ratio of 1:1 and mixed with the surface-activated carbon nanotube filler. The conductive paste is uniformly coated on the sandpaper template with a thickness of 150 μm.

4. The breathing state recognition method based on flexible deformation sensing according to claim 2, characterized in that: In step S13, the hardware acquisition circuit adopts a high-precision miniature Wheatstone bridge, with the flexible sensor serving as the dynamic resistor of one arm of the bridge, and the other three arms configured as high-precision fixed surface mount resistors. A DC constant voltage source is used as the excitation voltage of the bridge. The differential voltage signal output by the bridge is amplified by a pre-amplifier and then input to an analog-to-digital converter chip to obtain a digital signal. The digital signal is then wirelessly transmitted in real time to a host computer or mobile terminal in the form of a data packet with a microsecond-level timestamp through a microcontroller with integrated Bluetooth protocol.

5. The breathing state recognition method based on flexible deformation sensing according to claim 1, characterized in that: In step S2, different exercise loads include sitting still, walking at a constant speed, turning the head left and right, and nodding up and down; different breathing states include calm breathing, deep breathing, rapid breathing, or coughing.

6. The breathing state recognition method based on flexible deformation sensing according to claim 1, characterized in that, Step S4 specifically includes: S41. Construct a multi-scale one-dimensional convolutional neural network module as a feature extraction module for the hybrid decoupled recognition network, used to extract local morphological features and reduce dimensionality; S42. Construct a bidirectional long short-term memory network module as a virtual compensator for the hybrid decoupled recognition network, which is used to perform bidirectional temporal rhythm decoupling and material hysteresis error compensation on the output of the multi-scale one-dimensional convolutional neural network module. S43. Construct a fully connected layer and a Softmax activation function layer as the output layer of the hybrid decoupled recognition network, which is used to map the high-dimensional features output by the bidirectional long short-term memory network module to the dimension and the total number of target categories. Strictly corresponding classification vectors It calculates the classification probability of each category and outputs the category label corresponding to the maximum probability as the final state determination result of the current sliding window sequence; S44. Determine the model generalization constraints, loss function, and training strategy for training the hybrid decoupled recognition network.

7. The breathing state recognition method based on flexible deformation sensing according to claim 6, characterized in that: In step S41, the multi-scale one-dimensional convolutional neural network module is composed of two layers of one-dimensional convolutional layers and max pooling layers with different receptive fields, which are alternately cascaded. The kernel size of the first convolutional layer is set to 1×9, the number of output channels is 32, and the stride is 1. The kernel size of the second convolutional layer is set to 1×3 and the number of output channels is 64. After each convolutional operation, a max pooling layer with a kernel size of 1×2 and a stride of 2 is connected. In step S42, the bidirectional long short-term memory network module receives the high-dimensional temporal feature vector after dimensionality reduction and flattening by the multi-scale one-dimensional convolutional neural network module, and its forward hidden layer follows the time sequence. Scanning, reverse hidden layer along time series The system scans and uses future contextual information to verify and correct the hysteresis bias at the current moment. This represents the total time step; In step S44, the model constraints are as follows: a Dropout regularization layer is introduced between the bidirectional long short-term memory network module and the fully connected layer, with a dropout rate of 0.4; the loss function is a cross-entropy loss function with an L2 weight decay penalty term; and the training strategy is as follows: during the network gradient update stage, the Adam adaptive momentum optimizer is used, with an initial learning rate set to 0.001, and a cosine annealing learning rate decay strategy is employed.

8. The breathing state recognition method based on flexible deformation sensing according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31. Perform frequency domain denoising on the received original signal sequence to obtain a denoised signal sequence; S32. An adaptive multi-scale overlapping sliding window is used to segment and reconstruct the denoised signal sequence to obtain a structured three-dimensional temporal feature matrix. The data overlap rate of the sliding window is 75%. S33. Normalize each of the structured three-dimensional temporal feature matrices to obtain temporal feature samples, and combine all temporal feature samples to obtain a temporal feature sample set.

9. A respiratory state recognition system based on flexible deformation sensing, characterized in that: An intelligent agent is provided to implement the breathing state recognition method based on flexible deformation sensing as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: It stores computer-executable code, and when the executable code is executed, it executes the breathing state recognition method based on flexible deformation sensing as described in any one of claims 1 to 8 or the breathing state recognition system based on flexible deformation sensing as described in claim 9.

Images