Data processing system and physiological sign monitoring system for monitoring physiological signs

Abstract

This application provides a data processing system and a physiological sign monitoring system for physiological sign monitoring. The data processing system for physiological sign monitoring includes: a synchronization and conditioning unit, an adaptive elimination unit, a feature extraction unit, an artificial intelligence inference unit, and a data flow management and scheduling unit. Through the coordinated operation of these units, motion artifacts are eliminated in real time and effectively, improving the signal-to-noise ratio of the signal waveform used to characterize physiological signs and the accuracy of the final monitoring results.

Description

Technical Field

[0001] This application relates to the field of physiological signal monitoring technology, and more specifically, to a data processing system and a physiological sign monitoring system for monitoring physiological signs. Background Technology

[0002] In sensor-based (e.g., fiber optic) non-contact physiological monitoring systems, the performance of the back-end signal processing unit directly determines the monitoring accuracy and reliability of the entire system. In practical applications, the acquired raw physiological signals typically include multiple types: high-sampling-rate signals reflecting dynamic signs such as cardiac impact and respiration, and low-sampling-rate signals reflecting static signs such as weight and body movement. These two types of signals differ significantly in amplitude, frequency, and noise characteristics, posing a challenge to synchronous processing.

[0003] More seriously, the user's physical activities during monitoring (such as turning over or slight movements) generate motion artifacts. These artifacts, as a type of non-stationary strong noise, often have amplitudes far exceeding the weak physiological signals themselves. Traditional signal processing methods, such as using filters with fixed parameters or simple software filtering algorithms running on general-purpose processors, are difficult to effectively cope with this strong interference that has an unknown and time-varying relationship with the body motion reference signal.

[0004] Therefore, when processing mixed physiological signals, existing technologies often fail to effectively separate motion artifacts from real physiological signals, resulting in a significant decrease in monitoring accuracy or even monitoring failure when the user moves. This leads to technical defects such as insufficient robustness and difficulty in guaranteeing real-time performance. Summary of the Invention

[0005] This application provides a data processing system and a physiological sign monitoring system for physiological sign monitoring, aiming to solve the technical problem in the prior art that it is difficult to eliminate strong motion artifacts in real time and effectively when processing multi-rate heterogeneous physiological signals, resulting in insufficient monitoring accuracy and robustness.

[0006] To address the aforementioned problems, embodiments of this application provide a data processing system for monitoring physiological signs, comprising:

[0007] The synchronization and conditioning unit is used to receive two digital signal streams with different sampling rates from the analog-to-digital conversion module, and to perform adaptive piecewise polynomial interpolation and fine phase synchronization processing on the low sampling rate signal stream in the digital signal streams with different sampling rates, so that the two digital signal streams achieve time base alignment; wherein, the digital signal streams include a first signal representing body weight or body movement signal and a second signal representing cardiopulmonary mixed signal;

[0008] An adaptive cancellation unit is used to use one of the time-base aligned signals as a reference noise input and employ a dual-mode variable-step-size least mean square algorithm to adaptively filter the other signal to eliminate motion artifacts and output a clean physiological signal. The adaptive filtering using the dual-mode variable-step-size least mean square algorithm includes: selecting either a fixed-step-size fine-locking mode or a normalized variable-step-size fast-tracking mode for filtering based on the energy of the noise signal. The noise signal is an estimate of unknown noise.

[0009] The feature extraction unit is used to perform parallel time-domain, frequency-domain, and nonlinear-domain feature extraction on the pure physiological signal and output a multi-dimensional physiological feature vector.

[0010] An artificial intelligence inference unit is used to receive the multidimensional physiological feature vector, and perform inference based on a preset lightweight neural network with attention mechanism to output physiological state classification results.

[0011] The data stream management and scheduling unit is used to encapsulate, schedule, and output a mixed data stream of the feature data extracted by the feature extraction unit and the physiological state classification results inferred by the artificial intelligence inference unit.

[0012] Optionally, the adaptive piecewise polynomial interpolation includes: dynamically selecting zero-order hold, first-order linear interpolation, or higher-order spline interpolation mode based on the comparison result between the difference between adjacent sampling points of the low-sampling-rate digital signal stream and a preset threshold; the fine phase synchronization processing includes: performing phase offset correction at the sub-sampling period level on the interpolated signal.

[0013] Optionally, the feature extraction unit includes components that operate in parallel:

[0014] The temporal feature extraction subunit is used to detect the J-peak of the cardiac impact signal in the pure physiological signal in real time and calculate the heart rate, J-peak amplitude, and rise slope, as well as to detect the zero-crossing point of the respiratory signal in the pure physiological signal and calculate the respiratory rate and respiratory cycle variation coefficient.

[0015] The frequency domain feature extraction subunit integrates a fast Fourier transform kernel to calculate the power spectral density of the pure physiological signal in a preset physiological frequency band.

[0016] A nonlinear feature extraction subunit is used to extract the entropy value of the pure physiological signal;

[0017] Also includes:

[0018] The context-aware post-processing subunit is used to perform validity verification, smoothing processing based on the current user state, and dynamic feature selection on each feature extracted by the time-domain feature extraction subunit, the frequency-domain feature extraction subunit, and the nonlinear feature extraction subunit.

[0019] Optionally, the step size in the fast tracking mode Calculated based on the following formula:

[0020]

[0021] in, This is a noise signal. For reference noise signal, For gain factor, is the regularization constant.

[0022] Optionally, the data flow management and scheduling unit is configured to: assign different transmission priorities to the feature data extracted by the feature extraction unit and the physiological state classification results inferred by the artificial intelligence inference unit according to the importance and timeliness of the data; predict future bandwidth requirements based on historical data traffic and adaptively adjust the parameters of the physical communication interface.

[0023] Optionally, the lightweight neural network with attention mechanism adopts an architecture combining CNN, BiLSTM and additive attention. The CNN layer contains 2 convolutional kernels, the BiLSTM layer contains 64 hidden units, and the physiological state classification results include sleep stages, cardiovascular risk level and respiratory abnormality warning.

[0024] Secondly, embodiments of this application provide a physiological sign monitoring system, which includes:

[0025] A distributed fiber optic sensing array, wherein each fiber optic sensor in the fiber optic sensing array includes a rigid reference arm, a flexible sensing fiber wound around the outer wall of the rigid reference arm, and at least one reference fiber passing through the internal cavity of the rigid reference arm; the distributed fiber optic sensing array is used to collect the user's physiological signals.

[0026] The photoelectric demodulation circuit is used to demodulate and output a first analog signal representing body weight or body movement signal and a second analog signal representing cardiopulmonary mixed signal based on the physiological signal.

[0027] The analog-to-digital conversion module includes a first ADC and a second ADC. The first ADC is used to sample the first analog signal at a dynamically adjustable first preset frequency and output a first digital signal. The second ADC is used to sample the second analog signal at a dynamically adjustable second preset frequency and output a second digital signal. The second preset frequency is greater than the first preset frequency.

[0028] The data processing system, as described in the first aspect, is a data processing system for monitoring physiological signs, used to receive the first digital signal and the second digital signal from the analog-to-digital conversion module, and output feature data and physiological state classification result data after processing.

[0029] The terminal host is used to generate control commands based on user operations, and to acquire and display the data output by the data processing system.

[0030] Optionally, the photoelectric demodulation circuit includes a DC servo loop, a high-frequency phase-locked loop, and an adaptive gain adjustment module;

[0031] The operation process of the photoelectric demodulation circuit includes:

[0032] In zero-bias mode, the photodiode converts the received optical signal into a photocurrent containing a static DC component and an AC carrier component.

[0033] The DC servo loop forms a negative feedback with the voltage-controlled current source through an integrator, generating a compensation current with the same amplitude and opposite direction as the static DC component. The static DC component is physically neutralized at the virtual ground bus node, and the output voltage of the integrator is used as the first analog signal output.

[0034] The remaining AC carrier component after neutralization enters the high-frequency phase-locked loop, is amplified by the transimpedance amplifier, demodulated by the synchronous phase-sensitive demodulator, and then outputs the second analog signal after low-pass filtering; wherein, the adaptive gain adjustment module is used to dynamically switch the feedback resistor value of the transimpedance amplifier.

[0035] Optionally, the adaptive gain adjustment module includes a switchable feedback resistor network and a signal amplitude monitoring unit. The switchable feedback resistor network contains multiple precision resistors with different resistance values, which are switched by an analog switch. The signal amplitude monitoring unit consists of a peak detector and a comparator, which dynamically triggers the range switching based on the peak value of the transimpedance amplifier output signal.

[0036] Optionally, the flexible sensing fiber is laid on the substrate in a continuous serpentine path, and the fiber sensing array is arranged in a three-level distributed layout along the longitudinal axis. The three-level distributed layout includes different layout methods in the core area, transition area, and background area. The core area corresponds to the thoracic spine and ischial tuberosities of the buttocks when the human body is supine, and the row spacing of the flexible sensing fiber in the core area is 1.2cm to 2.0cm. The transition area corresponds to the thoracic spine and the root of the thigh, and the row spacing of the flexible sensing fiber in the transition area linearly transitions from 2.0cm to 3.5cm along the longitudinal axis. The background area corresponds to the outer shoulder and outer thigh region of the human body, and the row spacing of the flexible sensing fiber in the background area increases exponentially with a non-linear spacing, with the maximum spacing being 3 to 4 times that of the core area.

[0037] Optionally, the terminal host is configured to: during the calibration phase, collect the user's resting physiological data and calculate the user's personalized baseline vector; during the monitoring phase, send the personalized baseline vector to the data processing system for physiological sign monitoring, so that the feature extraction unit of the data processing system for physiological sign monitoring can perform feature normalization on the extracted features.

[0038] The technical solution provided in this application has at least the following beneficial effects:

[0039] Improved monitoring accuracy and robustness: Through adaptive piecewise polynomial interpolation and fine phase synchronization processing, accurate alignment of multi-rate signals is achieved. Combined with the dual-mode variable step size least mean square (LMS) algorithm, it can deeply and adaptively suppress strong and non-stationary motion artifacts, solving the core technical problem of monitoring failure during body movement, and significantly improving the signal-to-noise ratio of physiological waveforms and the accuracy of the final monitoring results.

[0040] It achieves low latency and high real-time performance: The core algorithms such as signal synchronization, adaptive filtering, and feature extraction are implemented in hardware in the data processing system in a parallel or pipelined manner, which reduces the processing latency from tens of milliseconds in traditional software solutions to sub-milliseconds, and can meet the needs of fine physiological index analysis with extremely high real-time requirements such as heart rate variability (HRV).

[0041] Reduced system power consumption: By adopting a variety of energy-saving technologies such as adaptive filtering, dual-mode variable step size algorithm, and adaptive communication bandwidth adjustment, the overall power consumption is much lower than that of software solutions using general-purpose processors or DSPs while completing complex signal processing tasks, making it more suitable for long-term, portable monitoring equipment powered by batteries. Attached Figure Description

[0042] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. Furthermore, these drawings and textual descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this application to those skilled in the art through reference to specific embodiments.

[0043] Figure 1 This is a schematic diagram of the structure of a data processing system for monitoring physiological signs provided in an embodiment of this application;

[0044] Figure 2 This is a schematic diagram of the physiological sign monitoring system provided in the embodiments of this application;

[0045] Figure 3 This is a flowchart illustrating the physiological monitoring method provided in the embodiments of this application. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the embodiments of this application. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0047] Example 1

[0048] This embodiment provides a data processing system for monitoring physiological signs, capable of extracting weak physiological features in real time from sensor signals filled with noise and motion artifacts with high accuracy and robustness, and performing intelligent analysis. This system can be widely applied in non-invasive health monitoring devices, such as smart mattresses, monitoring chairs, or wearable devices. The data processing system can be integrated into a chip containing programmable logic and a processor system. The system internally includes multiple cooperating hardware units.

[0049] like Figure 1 As shown, the system specifically includes: a synchronization and conditioning unit, an adaptive elimination unit, a feature extraction unit, an artificial intelligence inference unit, and a data flow management and scheduling unit. The functions of each unit are as follows:

[0050] The synchronization and conditioning unit is the first stage of data processing. It receives two digital signal streams with different sampling rates from the analog-to-digital converter (ADC) and performs adaptive piecewise polynomial interpolation and fine-phase synchronization on the lower-sampling-rate signal stream to achieve time-base alignment between the two streams. The digital signal streams include a series of digital signals transmitted sequentially, specifically a first signal representing body weight or motion and a second signal representing a mixed cardiopulmonary signal. The first and second signals have different sampling rates: the first is a low-sampling-rate signal, and the second is a high-sampling-rate signal. The core function of the synchronization and conditioning unit is to resolve the time-base misalignment problem between the two signals. It increases the sampling rate of the lower-sampling-rate signal stream through adaptive piecewise polynomial interpolation and corrects the inherent slight time delay between different channels of the ADC through fine-phase synchronization, ultimately outputting two signal streams with the same sampling rate and strict time alignment.

[0051] The adaptive cancellation unit, connected to the synchronization and conditioning unit, uses a time-base aligned signal as a reference noise input and employs a dual-mode variable-step least mean square (LMS) algorithm to adaptively filter at least one other signal to eliminate motion artifacts and output a clean physiological signal. Specifically, the adaptive cancellation unit's task is to eliminate the main interference—motion artifacts—in the cardiopulmonary mixed signal (the second signal). It uses the aligned body weight or body movement signal (the first signal) as a reference noise input and employs a dual-mode variable-step least mean square (LMS) adaptive filtering algorithm to subtract relevant noise components from the other cardiopulmonary mixed signal (the second signal). In this way, even if the user turns over or moves during sleep, the system can effectively filter out significant interference and output purified physiological waveforms with a high signal-to-noise ratio (such as cardiac impulse signals and respiratory waves).

[0052] The feature extraction unit is connected to the adaptive elimination unit and is used to perform parallel time-domain, frequency-domain, and nonlinear-domain feature extraction on the purified physiological signal (i.e., the signal after motion artifact removal), and output a multidimensional physiological feature vector. Specifically, the task of the feature extraction unit is to perform deep feature mining on the purified physiological waveform. This unit adopts a parallel pipeline architecture, simultaneously performing feature extraction in the time domain, frequency domain, and nonlinear domain. For example, the feature extraction unit may include a time-domain feature extraction subunit, a frequency-domain feature extraction subunit, and a nonlinear feature extraction subunit. Specifically, the time-domain feature extraction subunit detects the J-peak of the cardiac impulse signal in the purified physiological signal to calculate the heart rate; the frequency-domain feature extraction subunit calculates the power spectral density of the purified physiological signal in a specific physiological frequency band using FFT; and the nonlinear feature extraction subunit calculates the entropy value of the purified physiological signal to assess its complexity. These raw features are further integrated and filtered to ultimately form a multidimensional physiological feature vector.

[0053] The AI ​​inference unit is connected to the feature extraction unit and receives multi-dimensional physiological feature vectors. It then performs inference based on a pre-built lightweight neural network with an attention mechanism, outputting a physiological state classification result. Specifically, the AI ​​inference unit is the core processing component for achieving advanced intelligent analysis of the system. It receives a sequence of multi-dimensional physiological feature vectors generated by the feature extraction unit in real time. Internally, this unit contains a pre-trained lightweight neural network model with an attention mechanism. Based on the input feature sequence, the model performs inference and outputs a classification result for the user's current physiological state, such as "awake," "light sleep," "deep sleep," or "REM sleep."

[0054] The data flow management and scheduling unit is connected to both the feature extraction unit and the artificial intelligence inference unit. It is used to encapsulate, schedule, and output a mixed data flow of feature data extracted by the feature extraction unit and physiological state classification results inferred by the artificial intelligence inference unit. Specifically, the data flow management and scheduling unit is responsible for the data scheduling and management of the entire data processing system. It is responsible for encapsulating, scheduling, and outputting a mixed data flow of feature data extracted in real time by the feature extraction unit (e.g., a feature vector per second) and physiological state classification results inferred by the artificial intelligence inference unit (e.g., sleep stage results per 30 seconds). For example, it can assign high priority to feature data packets representing real-time heart rate to ensure low-latency transmission, while assigning lower priority to physiological state classification results with lower update frequencies, thereby optimizing the communication bandwidth and response capability of the entire system.

[0055] In one optional implementation, adaptive piecewise polynomial interpolation includes: dynamically selecting zero-order hold, first-order linear interpolation, or higher-order spline interpolation mode based on the comparison result between the difference between adjacent sampling points of the low-sampling-rate digital signal stream and a preset threshold; fine phase synchronization processing includes: performing phase offset correction at the sub-sampling period level on the interpolated signal.

[0056] Specifically, the synchronization and conditioning unit may include an adaptive piecewise polynomial interpolator to implement adaptive piecewise polynomial interpolation processing, which may include a detection module, a finite state machine, and an output multiplexer. The detection module calculates the difference between adjacent sampling points in real time. The finite state machine then uses this difference value. With preset threshold and ( The comparison results are used to dynamically control the output multiplexer to select different interpolation cores for the output. More specifically, when the signal changes smoothly ( When the signal changes slowly, a zero-order hold interpolation kernel is selected, which simply copies the previous sampling point, resulting in the lowest power consumption; when the signal changes slowly... When the signal changes drastically, a first-order linear interpolation kernel is selected to perform linear interpolation between two sampling points, balancing accuracy and efficiency; when the signal changes drastically ( For example, when a user quickly turns over, a more complex but smoother third-order spline interpolation kernel is activated. This adaptive mechanism optimizes overall power consumption while maintaining accuracy in the interpolation process.

[0057] In addition, to achieve fine phase synchronization at the sub-sampling period level, the synchronization and conditioning unit may also include a fractional delay filter, which is connected after the adaptive piecewise polynomial interpolator to perform a small phase offset correction on the interpolated signal, ensuring that the two signals are precisely aligned in time.

[0058] In one alternative implementation, the adaptive filtering process of the adaptive elimination unit is further described in detail.

[0059] In this embodiment, the adaptive filtering using the dual-mode variable step size least mean square algorithm includes: selecting either a fine-locking mode with a fixed step size or a fast tracking mode with a normalized variable step size for filtering based on the energy of the noise signal.

[0060] Specifically, in this embodiment, the adaptive cancellation unit includes a dual-mode variable step-size LMS filter. The core of this filter is a mode-switching logic that determines the operating mode based on the energy of the noise signal. When the error energy is low, indicating a stable monitoring state, a fine-locking mode is activated. This module uses a fixed, very small step size for filtering to obtain a lower steady-state error and a cleaner output waveform. When the noise signal energy suddenly increases, indicating strong motion artifacts, the mode-switching logic promptly switches to a fast tracking mode. In this mode, the step size... It is dynamically changing, and its calculation formula is:

[0061]

[0062] in, For noise signals, this refers to the estimated value of unknown noise (mainly motion artifacts), which can be estimated using methods such as FIR filtering. For reference noise signal, The preset gain factor, This is the regularization constant (to avoid the denominator being zero).

[0063] This variable step size strategy enables the filter to quickly adjust its weights, rapidly track and eliminate non-stationary motion disturbances.

[0064] To further improve the stability of the filter, the adaptive cancellation unit may also include a frequency domain constraint module. This module periodically transforms the time-domain weights of the dual-mode variable-step LMS filter to the frequency domain using an FFT (Fast Fourier Transform) module, applies a preset spectral mask (e.g., forces zero gain outside the heartbeat band), and then transforms it back to the time domain using an IFFT (Inverse Fast Fourier Transform) module. This periodic "calibration" prevents unwanted noise amplification in certain frequency bands by the dual-mode variable-step LMS filter.

[0065] In a preferred embodiment, the parallel processing capability of the feature extraction unit is enhanced. The time-domain feature extraction subunit not only calculates heart rate and respiratory rate, but also calculates the amplitude and rise slope of the J-peak of the cardiac impact signal in real time, as well as the coefficient of variation of the respiratory cycle, all of which are important indicators for assessing cardiopulmonary function. The frequency-domain feature extraction subunit utilizes an integrated fast Fourier transform kernel to efficiently calculate the power spectral density of the signal in a preset physiological frequency band (such as the respiratory band of 0.1-0.4 Hz and the cardiac impact band of 0.8-3.0 Hz), thereby decomposing the complex signal fluctuations in the time domain into different frequency ranges for quantitative analysis, thus revealing the physiological system and pathological state information hidden behind the signal (for example, calculating the power spectral density of the respiratory signal can be used to analyze respiratory rhythm and its variability). In addition, the feature extraction unit may also include a context-aware post-processing subunit, which is responsible for validating the extracted raw features (e.g., the heart rate value must be within a reasonable range of 40-200 BPM), smoothing the features according to the current user's state (e.g., stationary or moving), and finally dynamically selecting the most relevant features to form a feature vector, which is then sent to the artificial intelligence inference unit.

[0066] Furthermore, as a specific implementation, a lightweight neural network with an attention mechanism can employ an architecture combining CNN, BiLSTM, and additive attention. The CNN layers (e.g., containing two layers of convolutional kernels of different sizes) are used to automatically extract local temporal patterns, the BiLSTM layers (e.g., containing 64 hidden units) are used to capture long-range temporal dependencies, and the attention mechanism allows the model to focus on the feature time steps that contribute most to the outcome during decision-making. The physiological state classification results output by this network can be very rich, including sleep stages (e.g., awake, light sleep, deep sleep, REM), cardiovascular risk levels (e.g., normal, concern, high risk), and warnings of respiratory abnormalities (e.g., apnea, bradykinesia), etc.

[0067] In one alternative implementation, the data flow management and scheduling unit employs a more intelligent strategy. It is configured to assign different transmission priorities to feature data extracted by the feature extraction unit (such as data containing heart rate and respiratory rate) and physiological state classification result data packets (such as sleep staging results) based on the importance and timeliness of the data. For example, when a suspected sleep apnea event is detected, the relevant feature data packets are given the highest priority to ensure that the external host receives the alarm information immediately.

[0068] In an alternative implementation, the data flow management and scheduling unit can also predict future bandwidth requirements based on historical data traffic and adaptively adjust the parameters (such as baud rate or packet size) of the physical communication interface (such as SPI or UART) accordingly to achieve efficient and reliable data transmission.

[0069] In one specific hardware implementation, the aforementioned data processing system is fully integrated into a Field-Programmable Gate Array (FPGA). The synchronization and conditioning unit, adaptive elimination unit, feature extraction unit, and data flow management and scheduling unit are implemented using the FPGA's programmable logic (PL) resources through a hardware description language (such as Verilog or VHDL), forming a highly parallel processing pipeline. The artificial intelligence inference unit can be implemented using a dedicated deep learning processing unit (DPU) embedded within the FPGA or as a custom IP core within the PL. The processor core integrated on the FPGA chip runs a lightweight operating system or bare-metal program, responsible for top-level control of the entire system, communication with external systems, and performing some non-real-time management tasks.

[0070] Based on the above system, precise alignment of multi-rate signals is achieved through adaptive piecewise polynomial interpolation and fine phase synchronization processing. Combined with the dual-mode variable step size least mean square (LMS) algorithm, it can deeply and adaptively suppress strong and non-stationary motion artifacts, solving the core technical problem of monitoring failure during body movement and significantly improving the signal-to-noise ratio of physiological waveforms and the accuracy of the final monitoring results. Furthermore, core algorithms such as signal synchronization, adaptive filtering, and feature extraction are implemented in hardware in parallel or pipelined manner on processors such as FPGAs, reducing processing latency from tens of milliseconds in traditional software solutions to sub-millisecond levels. This meets the high real-time requirements of detailed physiological indicator analysis, such as heart rate variability (HRV). In addition, by employing adaptive interpolation, dual-mode variable step size algorithms, and adaptive communication bandwidth adjustment, the overall power consumption is far lower than that of software solutions using general-purpose processors or DSPs while completing complex signal processing tasks, making it more suitable for long-term, portable monitoring devices powered by batteries.

[0071] Example 2

[0072] Based on Example 1, this example provides a physiological sign monitoring system that can monitor and process human physiological signs.

[0073] like Figure 2 As shown, in addition to the data processing system (i.e., the data processing system for monitoring physiological signs in Embodiment 1), it also includes: a distributed fiber optic sensor array, a photoelectric demodulation circuit, an analog-to-digital conversion module, and a terminal host. The specific structure and function of each part are as follows:

[0074] The distributed fiber optic sensor array comprises multiple distributed fiber optic sensors, which are used to acquire the user's physiological signals.

[0075] Specifically, the fiber optic sensing array is used to convert physiological pressures applied by the human body (including static weight, dynamic respiration, and heartbeat) into composite optical signals. This sensing array employs a rigid-flexible heterogeneous helical coupling structure. Specifically, each fiber optic sensor includes a rigid reference arm, a flexible sensing fiber wound around the outer wall of the rigid reference arm, and at least one reference fiber passing through the internal cavity of the rigid reference arm. The Young's modulus of the rigid reference arm is much larger (at least 10 times greater) than that of the flexible sensing fiber, and an air gap of 0.4 mm to 0.9 mm is provided between the internal cavity and the dual-core reference fiber. When external pressure is applied to the fiber optic sensing array, the rigid reference arm acts like a "rigid anvil," causing the flexible sensing fiber to slightly bend, resulting in light loss in the optical signal proportional to the pressure (P). The internal reference fiber, protected by the structure of the rigid reference arm, is isolated from external mechanical stress, but is in the same temperature field as the flexible sensing fiber; its optical signal primarily reflects changes in ambient temperature. By differentially processing the signals from the sensing fiber and the reference fiber, common-mode noise such as temperature drift can be effectively canceled at the physical level.

[0076] Furthermore, the internal reference fiber can adopt a "dual-core parallel" structure, that is, two reference fibers are set with a spacing of 0.2mm~0.3mm between them, both suspended in the internal cavity of the rigid reference arm. Based on this, temperature drift noise can be further suppressed by differential operation of the two reference signals, and the common mode rejection ratio (CMRR) can be improved by 15dB~20dB.

[0077] In some embodiments, the inner wall of the rigid reference arm may also be coated with a heat-insulating coating (made of aerogel) with a thickness of 10μm to 20μm to reduce the rapid conduction of human body heat to the reference optical fiber through the rigid arm and reduce the impact of instantaneous temperature fluctuations on the reference signal.

[0078] Building upon the above embodiments, to enhance the durability and sensitivity of the sensor, in some embodiments, the cladding surface of the flexible sensing fiber can be coated with a 2μm~5μm thick polyimide (PI) wear-resistant coating. The coating surface is treated with micro-nano texture (through laser etching or molding processes, with a texture depth of 0.5μm~1μm), which improves wear resistance (extending service life by more than 3 times) and enhances friction with the human body, reducing signal fluctuations caused by relative displacement. Furthermore, in some embodiments, the fiber core is doped with 0.1%~0.3% graphene quantum dots. By controlling the quantum dot concentration, the optical transmission efficiency and micro-bending sensitivity of the fiber are optimized, increasing the sensitivity coefficient k by 20%~30% and ensuring effective modulation of weak signals. Simultaneously, to maximize flexibility, the flexible sensing fiber can be stripped of its traditional PVC / PE protective sheath, retaining only a thin cladding layer less than 10μm thick, controlling its outer diameter to 0.5mm±0.05mm.

[0079] Furthermore, the flexible sensing fiber is laid on the substrate in a continuous serpentine path. The sensing array can be arranged in a three-level distributed layout of "core area - transition area - background area" along its longitudinal axis (e.g., along the length of the mattress) to optimize the allocation of sensing resources.

[0080] Specifically, the three-tiered distributed layout refers to employing different layout methods in the core area, transition area, and background area. The core area corresponds to critical stress areas such as the thoracic spine (T1-T8) and ischial tuberosities of the buttocks when the human body is supine. The row spacing of the flexible sensing fibers in the core area is 1.2cm~2.0cm to capture weak cardiac impact signals. The transition area corresponds to the thoracic spine (T9-T12) and groin. The row spacing of the flexible sensing fibers in the transition area linearly transitions from 2.0cm to 3.5cm along the longitudinal axis. The background area corresponds to non-critical areas such as the outer shoulder and outer thigh. The row spacing of the flexible sensing fibers in the background area increases non-linearly exponentially, with a maximum spacing 3 to 4 times (5.0cm~8.0cm) of the row spacing in the core area. This layout effectively saves on sensing material costs and system data processing burden while ensuring monitoring accuracy in critical areas.

[0081] In addition, the photoelectric demodulation circuit is used to demodulate and output a first analog signal to characterize body weight or body movement signals based on physiological signals and a second analog signal to characterize mixed cardiopulmonary signals.

[0082] As one specific implementation, the photoelectric demodulation circuit includes a DC servo loop, a high-frequency phase-locked loop, and an adaptive gain adjustment module. Its operation is as follows:

[0083] The optical signal output from the fiber optic sensor is converted into a photocurrent (i.e., input current) containing static DC and dynamic AC components by a photodiode in zero-bias mode (also known as photovoltaic mode, meaning the photodiode operates with an applied bias voltage of 0V). A DC servo loop forms negative feedback with an integrator and a voltage-controlled current source (VCCS). This loop generates a sink current (compensation current) with the same amplitude but opposite direction to the static DC component, which cancels out the DC component in the input current at the virtual ground junction (physical neutralization). At this point, the integrator's output voltage is the first analog signal representing body weight or large-amplitude body movement. After DC component neutralization, the remaining net current mainly consists of an AC carrier component carrying weak physiological information such as cardiopulmonary function. This AC carrier component then enters a transimpedance amplifier (TIA) for high-gain amplification, and after synchronous phase-sensitive demodulation and low-pass filtering, outputs a second analog signal representing the mixed cardiopulmonary signal.

[0084] In addition, the opto-demodulation circuit also includes an adaptive gain adjustment module, which can dynamically switch the feedback resistor value of the transimpedance amplifier (TIA) according to the signal amplitude, thereby maintaining the best signal-to-noise ratio over a wide dynamic range and avoiding signal saturation or insufficient gain.

[0085] Specifically, the adaptive gain adjustment module may include a switchable feedback resistor network and a signal amplitude monitoring unit. The switchable feedback resistor network contains multiple precision resistors with different resistance values, which are switched by analog switches. The signal amplitude monitoring unit consists of a peak detector and a comparator, which dynamically triggers the range switching based on the peak value of the transimpedance amplifier output signal.

[0086] For example, this switchable feedback resistor network can include precision resistors with three resistance levels: 100MΩ, 200MΩ, and 500MΩ, controlled by an analog switch (such as an optocoupler relay). The signal amplitude monitoring unit monitors the peak value of the transimpedance amplifier (TIA) output signal in real time. When the peak value falls below a certain lower threshold, it triggers a switch to a higher resistance value to increase gain; when the peak value exceeds a certain upper threshold, it triggers a switch to a lower resistance value to prevent saturation.

[0087] In addition, the analog-to-digital conversion module includes a first ADC and a second ADC. The first ADC is used to sample a first analog signal at a dynamically adjustable first preset frequency and output a first digital signal (corresponding to the first signal received by the synchronization and conditioning unit). The second ADC is used to sample a second analog signal at a dynamically adjustable second preset frequency and output a second digital signal (corresponding to the second signal received by the synchronization and conditioning unit). The second preset frequency is greater than the first preset frequency.

[0088] Specifically, to address the different characteristics of the two analog signals, the analog-to-digital converter (ADC) module employs a heterogeneous ADC architecture. The first analog signal, weight or body motion, is characterized by slow changes but requires high precision. Therefore, it is sampled by a low-sampling-rate, high-resolution first ADC (e.g., a Sigma-Delta ADC) to output the first digital signal. The second analog signal, a cardiopulmonary mixed signal, is characterized by a higher frequency but slightly lower absolute precision requirements. Therefore, it is sampled by a high-sampling-rate, medium-resolution second ADC (e.g., a SAR ADC) to output the second digital signal. Typically, the sampling frequency of the second ADC (the second preset frequency) is much greater than (50 times or more) the sampling frequency of the first ADC (the first preset frequency). The two digital signals are transmitted to the subsequent processing unit via an isolation barrier and a standard bus (such as an SPI bus). Furthermore, in practical applications, both the first and second preset frequencies can be dynamically adjusted to reduce power consumption while meeting accuracy requirements.

[0089] More specifically, the first ADC, such as a Sigma-Delta ADC, has a resolution greater than or equal to 24 bits to ensure the accuracy of weight measurement. Its sampling rate can be dynamically switched in the range of 1SPS to 10SPS, controlled by the terminal host according to the user's status: when the user is at rest, it switches to 1SPS to reduce power consumption; when body movement is detected, it automatically switches to 5SPS to 10SPS to improve the accuracy of capturing body movement signals.

[0090] The second ADC, such as a SAR-type ADC, has a resolution greater than or equal to 12 bits, and its sampling rate can be dynamically adjusted within the range of 100Hz to 1000Hz, based on the adaptive setting of the signal bandwidth: when the respiratory rate is <10 breaths / minute (signal bandwidth <0.17Hz), the sampling rate is set to 100Hz; when the heart rate is >180 beats / minute (signal bandwidth >3Hz), the sampling rate is set to 1000Hz, satisfying the Nyquist sampling theorem, while optimizing power consumption.

[0091] The analog-to-digital conversion module samples the data using the sampling principle described above and outputs a first digital signal and a second digital signal to the data processing system. The data processing system receives the first and second digital signals from the analog-to-digital conversion module and processes them according to the principle and process described in Example 1. After processing, it outputs feature data and physiological state classification result data, which are then displayed to the terminal host.

[0092] In addition, the terminal host serves as the system's user interaction and control center, used to generate control commands based on user operations, and to acquire and display data output from the data processing system used for physiological sign monitoring. This terminal host can be an embedded device, such as a Raspberry Pi.

[0093] Furthermore, considering the significant differences in physiological baselines among individual users, algorithms using uniform thresholds or models are difficult to guarantee accuracy. Therefore, in some embodiments, the terminal host implements personalized monitoring for different individual users through a two-stage "calibration-monitoring" mechanism.

[0094] During the calibration phase, the terminal host guides the user to remain at rest, collects resting physiological data for a period of time, and calculates and generates the user's personalized baseline vector (PBV) based on this data. During the monitoring phase, the terminal host sends this personalized baseline vector to the data processing system used for physiological sign monitoring at the start of the task. After feature extraction, the feature extraction unit of the data processing system uses the baseline vector to normalize the feature vector, thereby eliminating the influence of individual differences and making subsequent neural network inference more accurate.

[0095] Example 3

[0096] This embodiment is used to illustrate a physiological monitoring method based on the above-mentioned physiological sign monitoring system.

[0097] like Figure 3 As shown, the complete workflow of this method is as follows:

[0098] First, during the S101 system initialization process, the system is powered on, and the data processing system and terminal host complete self-tests and program loading.

[0099] Next, the user will proceed to step S102, "User Identification and Calibration Selection," where the terminal host interface will prompt the user to select their identity or proceed with calibration. If the user selects calibration or the system identifies them as a new user, the user will proceed to step S103, "Personalized Baseline Calibration."

[0100] In step S103, the terminal host guides the user to rest for a period of time (e.g., 120 seconds), while the system collects physiological data during this period. The terminal host calculates and stores a personalized baseline vector (PBV) representing the user's health status based on the collected data.

[0101] Next, the process proceeds to step S104, loading and distributing the personalized baseline vector. Whether the vector is newly calibrated or retrieved from the database, the terminal host will distribute the corresponding PBV to the data processing system via the communication interface before monitoring begins.

[0102] Next, in the S105 monitoring startup step, the system enters real-time monitoring mode. A distributed fiber optic sensor array continuously collects optical signals modulated by human physiological activity. The photoelectric demodulation circuit converts the optical signals into two analog signals. The analog-to-digital converter digitizes these two signals at a dynamically adjusted sampling rate. The data processing system executes the core algorithm: upsampling the low-sampling-rate digital body motion signal to generate a reference noise signal; using a variable-step-size LMS adaptive filtering algorithm to cancel motion artifacts in the high-speed cardiopulmonary digital signal, outputting a clean physiological waveform signal; then, multi-domain feature extraction is performed on the clean waveform to obtain a feature vector; next, the feature vector is normalized using the distributed PBV; finally, the normalized feature vector is input into a lightweight neural network with an attention mechanism, outputting multi-dimensional physiological state classification results.

[0103] In the S106 terminal host processing steps, the data processing system continuously uploads real-time feature data (e.g., encapsulated as 0xF1 packets) and physiological state classification results (e.g., encapsulated as 0xF2 packets) to the terminal host through a mixed data output interface. The terminal host is responsible for parsing the data packets and dynamically displaying information such as heart rate, respiratory waveform, and sleep stage on the screen. During this process, the terminal host can also execute a dynamic baseline update algorithm to fine-tune the baseline based on recent monitoring data to adapt to the slow changes in the user's physiological state.

[0104] Finally, when the monitoring ends, proceed to step S107 to end and generate a report. The system will stop collecting data and compile all the data and analysis results from this monitoring into a health report for users to review or share.

[0105] In summary, this application constructs a complete closed-loop system from perception to decision-making through ingenious hardware design and intelligent integration at the algorithm level. It can effectively overcome the two major challenges of motion artifacts and individual differences, and achieve high-precision and highly robust non-invasive physiological sign monitoring.

[0106] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0107] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0108] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A data processing system for monitoring physiological signs, characterized in that, include: The synchronization and conditioning unit is used to receive two digital signal streams with different sampling rates from the analog-to-digital conversion module, and to perform adaptive piecewise polynomial interpolation and fine phase synchronization processing on the low sampling rate signal stream in the digital signal streams with different sampling rates, so that the two digital signal streams achieve time base alignment; wherein, the digital signal streams include a first signal representing body weight or body movement signal and a second signal representing cardiopulmonary mixed signal; An adaptive cancellation unit is used to use one of the time-base aligned signals as a reference noise input and employ a dual-mode variable-step-size least mean square algorithm to adaptively filter the other signal to eliminate motion artifacts and output a clean physiological signal. The adaptive filtering using the dual-mode variable-step-size least mean square algorithm includes: selecting either a fixed-step-size fine-locking mode or a normalized variable-step-size fast-tracking mode for filtering based on the energy of the noise signal. The noise signal is an estimate of unknown noise. The feature extraction unit is used to perform parallel time-domain, frequency-domain, and nonlinear-domain feature extraction on the pure physiological signal and output a multi-dimensional physiological feature vector. An artificial intelligence inference unit is used to receive the multidimensional physiological feature vector, and perform inference based on a preset lightweight neural network with attention mechanism to output physiological state classification results. The data stream management and scheduling unit is used to encapsulate, schedule, and output a mixed data stream of the feature data extracted by the feature extraction unit and the physiological state classification results inferred by the artificial intelligence inference unit.

2. The system according to claim 1, characterized in that, The adaptive piecewise polynomial interpolation includes: dynamically selecting zero-order hold, first-order linear interpolation, or higher-order spline interpolation mode based on the comparison result between the difference between adjacent sampling points of the low-sampling-rate digital signal stream and a preset threshold. The fine phase synchronization processing includes performing phase offset correction at the subsampling period level on the interpolated signal.

3. The system according to claim 1, characterized in that, The feature extraction unit includes components that operate in parallel: The temporal feature extraction subunit is used to detect the J-peak of the cardiac impact signal in the pure physiological signal in real time and calculate the heart rate, J-peak amplitude, and rise slope, as well as to detect the zero-crossing point of the respiratory signal in the pure physiological signal and calculate the respiratory rate and respiratory cycle variation coefficient. The frequency domain feature extraction subunit integrates a fast Fourier transform kernel to calculate the power spectral density of the pure physiological signal in a preset physiological frequency band. A nonlinear feature extraction subunit is used to extract the entropy value of the pure physiological signal; Also includes: The context-aware post-processing subunit is used to perform validity verification, smoothing processing based on the current user state, and dynamic feature selection on each feature extracted by the time-domain feature extraction subunit, the frequency-domain feature extraction subunit, and the nonlinear feature extraction subunit.

4. The system according to claim 1, characterized in that, Step size in the fast tracking mode Calculated based on the following formula: in, This is a noise signal. For reference noise signal, For gain factor, is the regularization constant.

5. The system according to claim 1, characterized in that, The data flow management and scheduling unit is configured to: assign different transmission priorities to the feature data extracted by the feature extraction unit and the physiological state classification results inferred by the artificial intelligence inference unit according to the importance and timeliness of the data; predict future bandwidth requirements based on historical data traffic and adaptively adjust the parameters of the physical communication interface; The lightweight neural network with attention mechanism adopts an architecture combining CNN, BiLSTM and additive attention. The CNN layer contains 2 convolutional kernels, and the BiLSTM layer contains 64 hidden units. The physiological state classification results include sleep stages, cardiovascular risk level and respiratory abnormality warning.

6. A physiological sign monitoring system, characterized in that, include: A distributed fiber optic sensing array, wherein each fiber optic sensor in the fiber optic sensing array includes a rigid reference arm, a flexible sensing fiber wound around the outer wall of the rigid reference arm, and at least one reference fiber passing through the internal cavity of the rigid reference arm; the distributed fiber optic sensing array is used to collect the user's physiological signals. The photoelectric demodulation circuit is used to demodulate and output a first analog signal representing body weight or body movement signal and a second analog signal representing cardiopulmonary mixed signal based on the physiological signal. The analog-to-digital conversion module includes a first ADC and a second ADC. The first ADC is used to sample the first analog signal at a dynamically adjustable first preset frequency and output a first digital signal. The second ADC is used to sample the second analog signal at a dynamically adjustable second preset frequency and output a second digital signal. The second preset frequency is greater than the first preset frequency. The data processing system is the data processing system for monitoring physiological signs as described in any one of claims 1 to 5, used to receive the first digital signal and the second digital signal from the analog-to-digital conversion module, and output feature data and physiological state classification result data after processing; The terminal host is used to generate control commands based on user operations, and to acquire and display the data output by the data processing system.

7. The system according to claim 6, characterized in that, The photoelectric demodulation circuit includes a DC servo loop, a high-frequency phase-locked loop, and an adaptive gain adjustment module. The operation process of the photoelectric demodulation circuit includes: In zero-bias mode, the photodiode converts the received optical signal into a photocurrent containing a static DC component and an AC carrier component. The DC servo loop forms a negative feedback with the voltage-controlled current source through an integrator, generating a compensation current with the same amplitude and opposite direction as the static DC component. The static DC component is physically neutralized at the virtual ground bus node, and the output voltage of the integrator is used as the first analog signal output. The remaining AC carrier component after neutralization enters the high-frequency phase-locked loop, is amplified by the transimpedance amplifier, demodulated by the synchronous phase-sensitive demodulator, and then outputs the second analog signal after low-pass filtering; wherein, the adaptive gain adjustment module is used to dynamically switch the feedback resistor value of the transimpedance amplifier.

8. The system according to claim 7, characterized in that, The adaptive gain adjustment module includes a switchable feedback resistor network and a signal amplitude monitoring unit. The switchable feedback resistor network contains multiple precision resistors with different resistance values, which are switched by an analog switch. The signal amplitude monitoring unit consists of a peak detector and a comparator, which dynamically triggers the range switching based on the peak value of the transimpedance amplifier output signal.

9. The system according to claim 6, characterized in that, The flexible sensing fiber is laid on the substrate in a continuous serpentine path, and the fiber sensing array is arranged in a three-level distributed layout along the longitudinal axis. The three-level distributed layout includes different layout methods in the core area, transition area and background area. The core area corresponds to the thoracic spine and ischial tuberosities of the buttocks when the human body is supine, and the row spacing of the flexible sensing fiber in the core area is 1.2cm~2.0cm. The transition area corresponds to the thoracic spine and the root of the thigh, and the row spacing of the flexible sensing fiber in the transition area linearly transitions from 2.0cm to 3.5cm along the longitudinal axis. The background area corresponds to the outer shoulder and outer thigh area of ​​the human body, and the row spacing of the flexible sensing fiber in the background area increases exponentially non-linearly, with the maximum spacing being 3 to 4 times that of the core area.

10. The system according to claim 6, characterized in that, The terminal host is also configured to: during the calibration phase, collect the user's resting physiological data and calculate the user's personalized baseline vector; During the monitoring phase, the personalized baseline vector is sent to the data processing system for monitoring physiological signs, so that the feature extraction unit of the data processing system for monitoring physiological signs can normalize the extracted features.

Images