Flexible sensing array and adaptive warning multi-parameter data processing method and system

Abstract

The present application belongs to the technical field of medical flexible electronic technology and intelligent physiological parameter technology monitoring technology, and specifically provides a flexible sensing array and a multi-parameter data processing method and system for adaptive early warning, which comprises the following steps: acquiring multi-dimensional physiological original signals; adopting a VMD-CEEMDAN joint algorithm combined with spectral purity analysis to map and reconstruct weights, and executing Bayesian filter bias compensation by using a redundant calibration framework; constructing an adaptive dynamic reference benchmark based on physical state recognition of a monitoring object, and extracting a multi-scale time sequence feature vector; applying a hierarchical decision-making framework, identifying an abnormal propagation mode of parameters by a GNN model, executing dynamic weight correction by using attention mechanism fusion technology context metadata, combining a time sequence framework to predict an evolution trend, and outputting differentiated hierarchical early warning instruction signals. The present application effectively suppresses flexible sensing artifacts and drift interference, reduces the false alarm rate, and realizes stable and highly reliable, intelligent quantitative evaluation feedback of a monitoring object.

Description

Technical Field

[0001] This invention relates to the fields of medical flexible electronics technology and intelligent physiological parameter monitoring technology, specifically to a multi-parameter data processing method and system for flexible sensor arrays and adaptive early warning. Background Technology

[0002] In the field of modern physiological information monitoring and safety surveillance, real-time acquisition and accurate analysis of the physiological parameters of monitored subjects are of significant technical value for assessing the evolution of physiological states. Traditional monitoring devices mostly employ rigid sensor designs, whose physical form makes it difficult to achieve good conformal contact with human tissue, resulting in poor wearing comfort and low subject compliance, and failing to meet the requirements for high-fidelity, long-term continuous data acquisition. In recent years, advancements in flexible electronics technology have provided new pathways for the development of wearable monitoring terminals; however, several technical bottlenecks still remain to be overcome in practical engineering applications and complex signal processing logic.

[0003] First, the stability of flexible sensor signals is highly susceptible to interference from the physical environment and dynamic behavior. Due to the complex dynamic interaction between flexible sensors and the skin interface, motion artifacts introduced by the daily activities of the monitored object, fluctuations in environmental temperature and humidity, and changes in contact impedance can all lead to a significant decrease in the signal-to-noise ratio of the acquired raw physiological signals, resulting in severe data quality fluctuations. Existing data analysis logic largely relies on fixed-threshold early warning strategies, which lack the ability to adaptively process non-stationary signals, easily triggering large-scale false alarms and causing "alarm fatigue," which severely weakens the response efficiency of the monitoring system in identifying extreme risk characteristics and the authenticity of data feedback.

[0004] Secondly, existing physiological monitoring algorithms have technical limitations in terms of individualized baseline adaptation. The basic physiological characteristics, data distribution patterns, and physiological regulatory benchmarks of different monitoring subjects exhibit significant heterogeneity, making it difficult for static evaluation models using uniform population distribution thresholds to accurately capture individualized feature shifts. For example, for subject groups with different metabolic levels or respiratory and circulatory characteristics, fixed threshold logic often leads to an imbalance between the sensitivity and specificity of feature recognition, easily resulting in technical biases such as missed feature detections or over-warnings, and failing to achieve accurate extraction of individualized physiological state characteristics.

[0005] Furthermore, the dimensions of multi-parameter collaborative analysis and deep feature fusion are still insufficient. Existing monitoring systems generally adopt a single-parameter independent analysis architecture, failing to effectively establish mathematical models of the intrinsic correlations and collaborative change patterns among physiological parameters. In the process of physiological system evolution, the synchronous drift of multi-dimensional indicators often contains deeper information on state changes. Simple single-parameter logic cannot capture such complex feature sequences, resulting in limited accuracy in identifying complex physiological evolution processes and difficulty in providing comprehensive assessment data with high confidence.

[0006] Furthermore, the long-term service stability and calibration drift of flexible sensing units are key factors limiting their widespread application. Factors such as the degradation of the sensitive layer activity of biochemical sensors, the aging of optoelectronic devices, and biocontamination at electrode interfaces can all lead to nonlinear shifts in measurement data over time. Systems lacking adaptive dynamic calibration mechanisms struggle to maintain measurement consistency and data output accuracy throughout their entire lifespan. Therefore, developing a depth monitoring technology that integrates a high-performance flexible sensing architecture, adaptive signal enhancement algorithms, and multi-dimensional risk assessment strategies to address core issues such as signal distortion, logic rigidity, and performance drift has become an important research direction for improving the reliability and accuracy of physiological monitoring systems. Summary of the Invention

[0007] The purpose of this application is to provide a multi-parameter data processing method and system for flexible sensor arrays and adaptive early warning, so as to solve the above-mentioned technical problems.

[0008] In a first aspect, this application proposes a multi-parameter data processing method for flexible sensor arrays and adaptive early warning, which includes the following steps: S1. Acquire multi-dimensional raw physiological signals continuously collected by a flexible multimodal sensor array; S2. A joint algorithm based on variational mode decomposition and adaptive noise complete set empirical mode decomposition is adopted for multi-dimensional physiological raw signals. Spectral purity analysis is performed on each component to map dynamic reconstruction weights. Online drift detection is performed using redundant measurements between multi-dimensional physiological raw signals. Bayesian filtering is applied to estimate sensor bias and gain error. Selective calibration and update are performed based on online confidence index. Finally, the denoised physiological feature signal is reconstructed by combining dynamic reconstruction weights and calibration parameters. S3. Based on physiological feature signals, determine a dynamic reference benchmark that adapts to the state of the monitored object, and combine the dynamic reference benchmark to perform multi-scale time series feature extraction to obtain a multi-scale time series feature vector. S4. An adaptive early warning algorithm based on a hierarchical decision architecture is used to detect anomalies in multi-scale time-series feature vectors. Anomaly propagation patterns among multiple parameters are identified through parameter correlation topology. An attention mechanism is used to fuse multi-dimensional technical context metadata to perform dynamic weight correction and time-series trend prediction, and output differentiated hierarchical early warning command signals.

[0009] In the above technical solution, by constructing an adaptive benchmark and hierarchical early warning architecture, high-precision deconstruction and intelligent risk assessment of multimodal physiological signals are achieved in complex dynamic environments, significantly reducing early warning errors caused by individual differences and environmental noise.

[0010] Furthermore, the flexible multimodal sensor array includes: a metabolite sensor based on microchannel sweat collection, an interstitial fluid biochemical sensor based on microneedle electrodes, a hemodynamic sensor based on a flexible ultrasonic transducer, and an electrocardiogram / electromyogram sensor based on conductive polymers.

[0011] In the above technical solution, by integrating heterogeneous sensing units, the simultaneous sensing of biochemical and physical indicators is achieved, providing a complete data foundation for subsequent cross-validation and multi-parameter collaborative analysis.

[0012] Furthermore, in step S2, a joint algorithm is used to separate the trend term and the oscillation term, specifically including: By presetting the number of modes, balance parameters, and convergence tolerance, the alternating direction multiplier method is used to analyze the original physiological signal into multiple bandwidth-limited intrinsic mode components, among which the center frequency is lower than that of the intrinsic mode components. The first mode is defined as the trend term representing the baseline drift, and the remaining modes are defined as oscillation terms covering the ultra-low frequency to high frequency range; For the oscillation term, a set of intrinsic mode functions with clear time scale characteristics is extracted by introducing controlled Gaussian white noise to perform recursive decomposition and using the ratio of the noise set number to the standard deviation of the additional noise for mean processing, so as to eliminate the mode aliasing between physiological components and motion noise.

[0013] In the above technical solution, the joint decomposition algorithm is used to achieve accurate separation of trend term and small oscillation term in non-stationary signal, effectively suppressing the interference of baseline drift and mode mixing on feature extraction.

[0014] Furthermore, the expression for the set of intrinsic mode functions is:

[0015] In the formula, For the stripped-out first An intrinsic mode function component with a single characteristic frequency range; The total number of noise sets; The remaining energy sequence after the previous decomposition serves as the physical input for the current decomposition. These are the noise weighting factors for each order; It is a residual signal; For the first A Gaussian white noise sequence; Extracting the first from the synthesized signal Empirical mode decomposition operator for features at each time scale.

[0016] The above technical solution provides an iterative mathematical framework for signal decomposition, which ensures that intrinsic components with single time scale characteristics can still be extracted even in noisy environments, thereby improving the fidelity of signal reconstruction.

[0017] Furthermore, step S2 involves performing spectral purity analysis on each component to map the dynamic reconstruction weights. Specifically, this includes: performing a frequency domain transformation on each intrinsic mode function to obtain its spectral distribution; combining this with a pre-defined motion artifact feature spectral library; and calculating the energy belonging to the physiological signal band and the energy belonging to the motion artifact band respectively through power spectral density integration. Finally, the sigmoid logic function is used to map the spectral purity ratio to generate the dynamic reconstruction weights for each component. The calculation formula is as follows:

[0018] In the formula, For dynamic reconfiguration of weights; This is the slope control parameter; The energy carried by physiological signals, This represents the energy of the motion artifact frequency band.

[0019] In the above technical solution, dynamic weight allocation is achieved through energy distribution characteristics, which can identify and suppress motion artifact interference in a specific frequency band online, ensuring the spectral purity of the reconstructed signal.

[0020] Furthermore, step S2 includes: A heterogeneous redundancy calibration architecture consisting of a main sensing channel and redundant reference channels is established, and a state transition equation is constructed. With observation equation The evolution of the sensor's physical performance is dynamically and recursively tracked; where, at time... state vector Includes the measurement bias term and gain error term to be estimated, and the observation vector. Reflects the real-time physical residual between the main sensing channel and the redundant reference channel; Performing selective calibration updates based on online confidence metrics includes: Calculation time The standardized deviation distance between the measurements of the master and slave sensors is used as an online confidence index, and the calculation formula is as follows: In the formula, As a confidence level indicator; The main sensor at time Real-time measurement values; For redundant sensors at any time Real-time measurement values; The variance is measured by the main sensor. Variance is measured for redundant sensors; According to the confidence index Real-time distributed implementation of hierarchical adaptive updates: when Perform standard recursive updates to correct sensor biases. With gain parameters ;when When, restricted calibration is implemented by reducing the update weights; when The system will reject the parameter update and trigger a sensor self-test. The formula for physiological characteristic signals is expressed as follows:

[0021] In the formula, The final reconstructed output is a high signal-to-noise ratio physiological feature signal; Discrete update time index for Bayesian filter calibration; For a moment Estimated sensor measurement bias; For a moment Estimated sensor gain parameters; The time-domain variable corresponding to the physiological signal; This represents the upper limit of the number of modes; For the first One intrinsic mode component; For the set of valid intrinsic mode function indices; For the first One intrinsic mode function; For dynamic reconfiguration of weights; The denoised physiological feature signal is used as a clean data source and input into step S3 to perform dynamic reference benchmark modeling and feature extraction.

[0022] In the above technical solution, the established online calibration mechanism can dynamically adjust the calibration weight according to the stability of the sensor's operating status, effectively compensating for the sensitivity decay and zero-point offset of the hardware during long-term monitoring.

[0023] Furthermore, step S3 includes: S31, Preprocessing physiological feature signals based on physical logic thresholds; S32, apply Gaussian mixture model or kernel density estimation to fit the historical data distribution characteristics of each physiological parameter in order to establish a probability distribution function that characterizes the random evolution of the signal and delineate the dynamic distribution range; S33 identifies the real-time physical state labels of the monitored object, including resting, active, and sleeping states, and performs differentiated distribution modeling of physiological signal characteristics under different physical state labels to generate a dynamic reference benchmark that adaptively aligns with the real-time physiological behavior state of the monitored object; wherein, the physical state labels are obtained by extracting macroscopic envelope features from the physiological feature signals output in step S2, and the macroscopic envelope features include the instantaneous amplitude sequence obtained by performing Hilbert transform on the trend component, and the time-varying distribution of the energy proportion of each component; S34 uses an exponentially weighted moving average or recursive filtering algorithm to continuously track the statistical center shift of the probability distribution function in order to achieve adaptive compensation for baseline drift. S35, based on a dynamic reference benchmark, extracts multi-scale time-series feature vectors. The multi-scale time-series feature vectors include instantaneous features for characterizing immediate response features, short-term statistical features for describing short-period fluctuation attributes, and long-term trend features for characterizing the long-term evolution of physiological homeostasis.

[0024] In the above technical solution, by identifying the physical behavior state and dynamically compensating for the distribution center shift, the false alarm problem caused by slow baseline drift is solved, and the feature extraction is accurately normalized relative to the individual's physiological homeostasis.

[0025] Furthermore, the adaptive early warning algorithm adopts a hierarchical decision-making architecture, specifically including: The single-parameter risk assessment layer calculates the statistical deviation of real-time monitoring values ​​from dynamic reference benchmarks and generates risk contribution scores in an independent dimension. A multi-parameter collaborative identification layer is constructed to build a parameter correlation model based on graph neural networks, identify abnormal propagation patterns in the monitoring network, and calculate a system-level risk index. The technology context fusion layer uses an attention mechanism to fuse multi-dimensional technology context metadata and calls context gating rules to perform real-time online correction of warning weights and technology warning thresholds. The temporal evolution prediction layer predicts the trajectory of risk evolution and executes advance warnings when the predicted value reaches the physical safety boundary.

[0026] In the above technical solution, by deeply integrating single-indicator fluctuations, multi-indicator coupling modes and external environmental contexts, advanced risk prediction with context awareness is achieved, which greatly improves the reliability of the monitoring system's decision-making in complex evolution processes.

[0027] Furthermore, the multi-parameter collaborative recognition layer utilizes a parameter association model. Assess the overall steady state; among which, It is a set of nodes composed of different categories of physiological parameters, including frequency feature sources, pressure feature sources, concentration feature sources and temperature feature sources; The set of edges characterizing the biological coupling relationships between parameters includes output power coupling edges, pressure-rhythm correlation constraint edges, respiratory-circulatory compensation correlation edges, and metabolic response activation correlation edges; The feature vector includes the instantaneous sampled values, time-domain rate of change, and nonlinear dynamic exponent of each sensing node; The adjacency matrix, which describes the strength of the correlation between parameters, is calculated in real time by the dynamic correlation weights in the edge set and is used to quantitatively characterize the system-level risk index of the overall steady state of the monitored object. The calculation logic for the correlation weights includes: calculating the weight of the output power coupling edge based on the product of the Pearson correlation coefficient and the physical dynamics estimate between the signal sequences of the two sensing nodes; calculating the weight of the respiratory-circulatory compensation correlation edge based on the deviation of the ratio of the characteristic change rate between the two sensing nodes relative to the dynamic reference baseline; applying the radial basis function to calculate the Euclidean distance of the composite metabolic characteristics from the dynamic reference baseline, and calculating the weight of the metabolic response activation correlation edge based on the negative exponential mapping.

[0028] The aforementioned technical solution represents a leap from "single-point monitoring" to "system topology monitoring." It enables the capture of systemic anomaly propagation patterns hidden behind multiple weakly correlated parameters. Even if a single indicator does not reach the alarm threshold, the system can identify the steady-state collapse trend of the monitored object in advance through the coupling relationship between parameters (such as the decorrelation of pressure and rhythm). Simultaneously, by utilizing dynamic weights constructed using radial basis functions and Pearson correlation coefficients, it adaptively filters out sudden noise that does not conform to physical laws, ensuring that early warning decisions are based on rigorous dynamic correlations.

[0029] Furthermore, the execution logic of the context gating rules includes: when an operating mode involving external equipment is identified, the warning weight of the interference parameter is suppressed to zero weight to shield the false disturbances introduced by physical intervention; when a specific warning sensitive mode is identified, the risk contribution weight of the corresponding feature component is increased, and the tolerance boundary of the technical warning threshold is reduced simultaneously to enhance sensitivity; when an external power support intervention mode is identified, the offset of the target physical quantity benchmark is increased to compensate for the feature phenotypic offset introduced by external power support.

[0030] The aforementioned technical solution endows the early warning system with extremely strong "contextual awareness" and "intelligent correction" capabilities. When external physical interventions (such as dialysis, transportation, and equipment calibration) cause drastic but normal signal deviations, the system can effectively identify and filter out "non-technical fluctuations" through zero-weight suppression and baseline bias adjustment, greatly reducing the false alarm rate of the monitoring system. Furthermore, within known high-risk time windows (such as the early postoperative period or the medication adjustment period), the system achieves "refined focusing" on key physical evolution processes by reducing the technical tolerance boundary, ensuring millisecond-level response sensitivity to potential crisis signals.

[0031] Furthermore, the metabolite sensor in step S1 employs a microfluidic structure based on enzyme-immobilized hydrogels. The hydrogel matrix is ​​doped with a carbon nanotube network to enhance conductivity, and molecular imprinting technology is used for the selective recognition of target metabolites.

[0032] Furthermore, the early warning command signals include: information recording level commands, corresponding to low risk levels, used to drive the system to archive monitoring data; mobile alert level commands, corresponding to medium risk levels, used to push technical status alerts to mobile monitoring terminals; and highest response level commands, corresponding to high risk levels, used to trigger the highest level of technical response sequence.

[0033] Secondly, this application proposes a multi-parameter data processing system for flexible sensor arrays and adaptive early warning, the system comprising: The signal acquisition module is configured to acquire multi-dimensional raw physiological signals continuously collected by a flexible multimodal sensor array; The signal collaborative reconstruction module is configured to use a joint algorithm based on variational mode decomposition and adaptive noise complete set empirical mode decomposition on multi-dimensional raw physiological signals. It performs spectral purity analysis on each component to map dynamic reconstruction weights; utilizes redundant measurements between multi-dimensional raw physiological signals for online drift detection; applies Bayesian filtering to estimate sensor bias and gain error; performs selective calibration updates based on online confidence index; and finally combines dynamic reconstruction weights and calibration parameters to reconstruct the denoised physiological feature signal. The dynamic benchmark modeling module is configured to determine a dynamic reference benchmark that adaptively adjusts with the state of the monitored object based on physiological feature signals, and performs multi-scale temporal feature extraction in combination with the dynamic reference benchmark to obtain multi-scale temporal feature vectors. The hierarchical early warning decision module is configured to use an adaptive early warning algorithm based on a hierarchical decision architecture to detect anomalies in multi-scale time-series feature vectors, use an attention mechanism to fuse multi-dimensional technical context metadata to perform dynamic weight correction and time-series trend prediction, and output differentiated hierarchical early warning command signals.

[0034] Compared with the prior art, the beneficial results of the present invention are as follows: (1) This application significantly improves the sensing performance and data continuity of the device in complex dynamic scenarios by deeply integrating a flexible multimodal sensing array with an online calibration framework. Compared with traditional rigid monitoring devices, this solution uses a flexible material system to improve the physical compatibility between the sensor and the skin interface and reduce interface impedance fluctuations. At the same time, through a heterogeneous redundant calibration architecture and a Bayesian filtering algorithm, it realizes real-time tracking and online compensation of the physical performance evolution (such as gain shift and zero drift) of the flexible sensor during long-term operation. This integrated design of "sensing, calibration, and compensation" effectively solves the performance degradation problem of flexible devices throughout their entire life cycle and ensures the spatiotemporal consistency and measurement accuracy of the underlying raw data.

[0035] (2) This application proposes a joint denoising mechanism based on physical logic and spectral purity, which greatly enhances the extraction purity of physiological feature signals. Using the VMD-CEEMDAN joint decomposition algorithm, the complex original physiological signals are finely mapped to intrinsic mode spaces at different time scales, effectively eliminating mode aliasing between physiological electrophysiological / biochemical signals and motion artifacts. By introducing spectral purity analysis and Sigmoid mapping weights, the system can identify and suppress non-technical high-frequency interference and low-frequency baseline drift online, achieving dynamic reconstruction of high signal-to-noise ratio physiological feature signals. This mechanism provides clean and robust data input for subsequent high-dimensional feature analysis, significantly improving the system's anti-interference capability in drastic motion scenarios.

[0036] (3) This application solves the technical bottleneck of high false alarm rate and poor scene adaptability of fixed threshold strategy in multi-parameter monitoring by constructing a dynamic reference benchmark that adaptively adjusts with the physical state of the monitored object. The system no longer relies on universal statistical indicators, but captures the historical data distribution characteristics of specific monitored objects based on Gaussian mixture models, and achieves millisecond-level alignment between the benchmark model and the actual behavioral state (resting, active, sleeping) by combining macroscopic envelope recognition technology. This dynamic benchmark based on probability distribution function and drift adaptive compensation can accurately capture small outlier fluctuations that deviate from the individualized steady-state trajectory, greatly reducing false positive technical alarms while ensuring the sensitivity of capturing systemic risk signals.

[0037] (4) The hierarchical adaptive early warning algorithm of this application achieves multi-dimensional quantitative evaluation of the stability of the monitored object's state through deep coupling of graph neural networks (GNN) and attention mechanisms. Unlike traditional single-point exceedance alarms, this scheme uses GNN to capture the topological constraint relationships and abnormal propagation patterns between various physical quantities in the sensing network, and identifies system-level instability of the physiological network through the dynamic evolution of the adjacency matrix. Combined with multi-dimensional technical context metadata fused by the attention mechanism, the system can automatically implement weight gating based on external physical intervention and operating context, effectively avoiding non-technical early warning triggers caused by external device access or manual operation. The final output hierarchical early warning command signal provides the monitoring system with predictive automated feedback basis, realizing a technical closed loop from data sensing to intelligent decision-making. Attached Figure Description

[0038] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Many anticipated advantages of the embodiments and other embodiments of the invention will be readily recognized as they become better understood through reference to the following detailed description. Elements in the drawings are not necessarily to scale. The same reference numerals refer to corresponding similar parts.

[0039] Figure 1 This is a flowchart of a multi-parameter data processing method for flexible sensor arrays and adaptive early warning according to an embodiment of this application. Figure 2 This is a framework diagram of a flexible sensor array and adaptive early warning multi-parameter data processing system according to an embodiment of this application. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0041] refer to Figure 1 , Figure 1 A flowchart illustrating a multi-parameter data processing method for flexible sensor arrays and adaptive early warning according to an embodiment of this application is shown. As shown in the figure, the method includes the following steps: S1. Acquire multi-dimensional raw physiological signals continuously collected by a flexible multimodal sensor array.

[0042] In some specific embodiments, flexible multimodal sensor arrays utilize heterogeneously integrated functional units to transform the multidimensional physiological fluctuations of the monitored object into raw electrical signals that reflect state characteristics, thereby providing an objective physical data basis for long-term technical monitoring of human physiological state without involving disease diagnosis conclusions.

[0043] Specifically, the flexible multimodal sensor array constructs a collaborative sensing architecture covering biochemical indicators of the body surface, subcutaneous tissue fluid composition, deep vascular dynamics, and biopotential information. This array achieves the capture of physical quantities of multi-source information from the human body through the integration of heterogeneous sensors, including: a metabolite sensor based on microfluidic sweat collection, an interstitial fluid biochemical sensor based on microneedle electrodes, a hemodynamic sensor based on a flexible ultrasonic transducer, and an electrocardiogram / electromyogram sensor based on conductive polymers. At the hardware integration level, each sensing unit is integrated into a wearable patch or wristband via flexible printed circuits, employing a layered spatial layout. Specifically, arranged sequentially from the skin contact surface to the outer layer are: a microneedle interstitial fluid sensor array in the central region, microfluidic sweat sensors distributed around the center, flexible ultrasonic transducers in the side regions, and polymer electrocardiogram electrodes arranged in a ring at the edges; stable signal interconnection between the sensing units is achieved through embedded flexible silver nanowires. Among them, the metabolite sensor based on microchannel sweat collection, as a component of the flexible multimodal sensing array, adopts a microscale flow channel system integrating functionalized hydrogels. It utilizes capillary driving forces generated by differences in the wettability of the flow channel surface to achieve autonomous collection and directional transport of sweat, thereby guiding the analyte to the sensing active area. In terms of fabrication, the preparation and integration of the above sensing units can be achieved with reference to existing flexible electronics manufacturing technologies. The core lies in providing continuous and stable multidimensional physiological raw signal input for subsequent hierarchical adaptive early warning algorithms, enabling real-time technical monitoring and evaluation of physiological state characteristics.

[0044] In some specific embodiments, the flexible sensing array is realized through a multilayer functionalized composite structure. The flexible ultrasonic transducer uses a PVDF-TrFE piezoelectric film as the core functional material, and molecular dipoles are oriented and aligned through a thermo-pressure polarization process at 80 MV / m (temperature 120°C). Laser cutting is then used to array the sensing units, thereby improving the sensing sensitivity to ultrasonic echo signals. For biochemical feature acquisition, the microneedle electrode uses hyaluronic acid as the structural substrate, and a needle height of 500-800 μm is fabricated using a micromolding process. The radius of curvature of the needle tip is less than 10. The microneedle array was then used; subsequently, a 200nm thick Au / Cr conductive coating was deposited on the surface of the microneedles using magnetron sputtering, ensuring impedance stability while achieving extremely low puncture pain.

[0045] S2. A joint algorithm based on variational mode decomposition (VMD) and adaptive noise complete set empirical mode decomposition (CEEMDAN) is used for multi-dimensional physiological raw signals. VMD is used to separate the trend term and oscillation term in the signal, and CEEMDAN is applied to extract the intrinsic mode function of the oscillation term. Spectral purity analysis is performed on each component to map the dynamic reconstruction weights. Online drift detection is performed using redundant measurements between multi-dimensional physiological raw signals. Bayesian filtering is applied to estimate sensor bias and gain error, and selective calibration update is performed according to the online confidence index. Finally, the denoised physiological feature signal is reconstructed by combining the dynamic reconstruction weights and calibration parameters.

[0046] In some specific embodiments, the adaptive enhancement processing steps of the joint algorithm include: S21, Perform Variational Mode Decomposition (VMD) to extract the trend and oscillation terms. First, preset the number of modes for variational mode decomposition. (Value) ), balance parameters (Value) and convergence tolerance (Value) ), to the original physiological signals Precisely resolved into multiple frequencies with specific center frequencies Furthermore, the intrinsic mode function with limited bandwidth Among them, parameters The physical granularity of signal analysis is determined, ensuring that baseline drift, multidimensional physiological rhythms, and high-frequency interference can be completely separated from complex flexible sensing signals; balance parameters Used to constrain the bandwidth compactness of each component, ensuring that narrowband physiological characteristics such as heart rate or respiration are not contaminated by broadband motion noise in the frequency domain; convergence tolerance. As a stopping criterion for numerical optimization, it ensures computational stability and repeatability when extracting weak feature signals. Essentially, this process utilizes the Alternating Direction Multiplier Method (ADMM) to solve a constrained variational minimization problem, its mathematical expression being:

[0047] And the constraints are satisfied.

[0048] In the formula, The raw physiological signals represent data directly acquired by the flexible multimodal sensor array over time. The changing raw physical quantity signal stream contains effective physiological characteristics (such as electrocardiogram, pulse wave, respiration, etc.) as well as complex motion artifacts coupled in it and environmental baseline drift; The intrinsic mode function / modal component refers to the eigenmode function obtained after decomposition. Each of the following is an independent eigenmode component: Typically, it represents a low-frequency trend or baseline drift, while subsequent components are mapped to physiological fluctuation characteristics or high-frequency noise components in different frequency bands. The center frequency represents the number of... Modal components The algorithm iteratively searches for the center frequency surrounding the frequency space. It automatically locks the specific location of each characteristic rhythm in physiological signals (such as the main frequency of heart rate, the interference frequency of cadence, etc.); The total number of modalities, the preset number of decompositions, and reasonable settings are required. The value ensures that effective physiological characteristics and non-stationary disturbances are physically separated in different modal spaces; For Dirac Functions and Hilbert transform kernel Used to calculate the analytical signal of each modal component. For exponential frequency shifting; For time gradient operators; square Norm; For minimization operators.

[0049] In terms of specific decomposition logic, the system utilizes the low-frequency convergence characteristics of VMD to decompose the first mode. center frequency Automatically locks on proximity The Hz region, thus reducing frequencies below Hz. The Hz component is defined as a trend term, used to adaptively remove non-stationary baseline shifts caused by charge accumulation on the flexible sensor contact surface or slow body positional displacement. (Except for...) external subsequent modes These are called oscillation terms, and their center frequencies cover ultra-low frequencies ( Hz), low frequency ( Hz), intermediate frequency ( Hz) to high frequency ( By segmenting the signal into the frequency domain within the physiological characteristic range of Hz (1 Hz), a set of adaptive Wiener filters is used to accurately capture microscale features such as pulse waves, electrophysiological waveforms, or hemodynamic oscillations. This narrowband processing ensures that the reconstructed signal retains high-fidelity amplitude and phase information, completing the technical conversion from chaotic raw physical quantities to structured physiological characteristic signals without involving pathological diagnosis of the monitored object.

[0050] S22, perform CEEMDAN to further refine the oscillation term decomposition. When refining the oscillation terms obtained from VMD decomposition, the additional noise standard deviation ratio of the adaptive noise complete set empirical mode decomposition (CEEMDAN) is adjusted. (Value) ) and noise set number (Value) It can effectively suppress mode aliasing and extract eigenmode functions with high signal-to-noise ratio. .

[0051] Specifically, step S22 includes: S221: Receive the oscillation term signal separated in step S21 and define it as the initial residual signal. Then to Add the first amplitude controlled by Nstd to the middle A Gaussian white noise sequence Using empirical mode decomposition operators Extract its first-order component and use the noise set number After averaging, the first eigenmode function is obtained. .in, This represents the initial oscillation component extracted by VMD decomposition, which contains weak physiological fluctuations and high-frequency dynamic artifacts of the monitored object; The first generated by the system A sequence of Gaussian white noise is used to identify overlapping frequency components with the aid of random perturbation; This represents an empirical mode decomposition operator for extracting first-order time-scale features from synthetic signals. This is the first-order noise weighting factor, used to adjust the energy proportion of auxiliary noise in the initial decomposition stage; while This represents the total number of noise sets. By performing a statistical average on the results of multiple decompositions, the randomness of the white noise is ensured to cancel each other out, thereby enabling the effective physiological and physical characteristics hidden in the oscillation term to be accurately restored on a specific time scale.

[0052] S222, automatically updates the residual signal to ensure energy conservation and lossless information conversion during the decomposition process by calculating the current residual signal. This leads to the recursive decomposition phase. Subsequent stages (...) ) eigenmode functions All decomposed based on the residual terms of the previous order, and their formulas are expressed as follows: In the formula, For the stripped-out first An intrinsic mode function component with a single characteristic frequency range; The remaining energy sequence after the previous decomposition serves as the physical input for the current decomposition. The mechanism of the noise weighting factors set according to Nstd is to introduce specific auxiliary frequencies for each decomposition layer, thereby forcibly separating motion interference and physiological signals with similar frequencies, and fundamentally eliminating the mode mixing phenomenon. Extracting the first from the synthesized signal An empirical mode decomposition operator with time-scale characteristics. Finally, a convergence determination is performed according to the preset stopping criterion stopCr, when the residual signal... The number of extreme points is less than two, or its amplitude energy is lower than the preset tolerance. (like When ), the recursive loop stops. Here... As a stopping criterion tolerance, it is used to determine whether the residual terms no longer possess physical characteristics that can be further analyzed. Ultimately, the oscillation term is completely analyzed into a set of intrinsic mode functions with well-defined time-scale characteristics, providing a high-resolution physical feature source for subsequent signal purity verification and adaptive weight reconstruction using a spectrum library. This ensures high-fidelity enhancement of physiological signals without involving any medical diagnostic judgments.

[0053] S23, Identify motion artifact energy based on spectral features. Receive the set of eigenmode functions obtained from the decomposition in previous step S22, and obtain their spectral distribution by performing a frequency domain transformation on each component. And combined with a pre-set motion artifact feature spectrum library (including walking / jogging) Hz, arm swing (characteristic frequency bands such as Hz) to achieve real-time identification and power spectral density integration of characteristic frequencies introduced by dynamic behavior.

[0054] Specifically, through the first The frequency domain response is obtained by performing a Fourier transform on each intrinsic mode function (IMF). The energy belonging to the physiological signal frequency band in this component is calculated using the following formula. And the energy belonging to the motion artifact frequency band :

[0055]

[0056] In the formula, Representing the The energy component belonging to the physiological characteristic frequency band in each intrinsic mode function, the magnitude of its integral value reflects the extraction purity of the physiological components in that mode component; This represents the energy of the frequency band belonging to the motion artifact characteristic band in the modal component, and is used to quantitatively measure the interference intensity caused by the dynamic behavior of the monitored object to the signal at a specific time scale; The preset effective frequency band for physiological signals is set according to specific monitoring needs (such as the frequency band corresponding to heart rate characteristics). Hz or respiratory characteristics corresponding (Hz, etc.) as a frequency scale for anchoring effective biophysical information; Based on the interference characteristic frequency band preset by the motion artifact characteristic spectrum library, it is used to accurately locate the noise energy distribution range introduced by physical actions such as walking, arm swinging or body position changes. This represents the first Individual eigenmode functions The complex spectrum obtained after performing a Fourier transform is obtained by calculating the square of its magnitude. Obtain the power spectral density distribution of this component. This step realizes the technical transformation from time-domain waveforms to frequency-domain quality evaluation indicators. By quantifying the ratio of physiological energy to artifact energy, it provides an objective physical confidence basis for subsequent adaptive weight reconstruction allocation.

[0057] S24, the ratio of spectral purity is calculated using a logical function (Sigmoid function). Mapped to dynamically reconstructed weights :

[0058] In the formula, Slope control parameter (value) ), used to adjust the sensitivity of the weights to the spectral ratio.

[0059] In some specific embodiments, adaptive performance calibration includes: S21': Construct a multi-sensor heterogeneous redundancy calibration architecture. A collaborative sensing matrix is ​​established using heterogeneous units in a flexible multimodal sensor array, and online redundant measurement of the same physical quantity is achieved by configuring the main sensing channel and redundant reference channels.

[0060] Specifically, for biochemical indicators, a microneedle interstitial fluid sensor is used as the main channel, supplemented by a sweat biochemical sensor as a redundant reference; for hemodynamic parameters, a flexible ultrasonic transducer combined with pulse wave conduction time (PWTT) is used to perform cross-validation; for bioelectrical and optical characteristics, a synergistic layout of ECG electrodes and photoplethysmography pulse waves, as well as reflective and transmissive photoelectric sensors, is used, and online parameter alignment is performed regularly with reference to external technical standards, thereby providing a multi-dimensional physical reference benchmark for online drift detection; S22', Establish a state tracking model for sensor technology based on Bayesian filtering. This is achieved by constructing state transition equations. With observation equation The evolution of the sensor's physical performance is dynamically and recursively tracked. During this process, time... Sensor state vector It includes the measurement bias term and gain error term to be estimated, which are expressed through the state transition matrix. Describe the performance evolution of sensors caused by aging of flexible materials or environmental temperature drift. The process noise vector; the observation vector. Used to quantify the real-time physical residual between the primary sensing channel and redundant channels or a reference standard, and combined with the observation matrix. The covariance matrix composed of the observation variances of each channel This enables a quantitative characterization of the uncertainty in sensor technology status; among which, To observe the noise vector, The variance is measured using the main sensor. The variance is measured for redundant sensors.

[0061] S23' performs online confidence assessment based on standardized deviation distance. This is achieved by calculating confidence assessment indices. The consistency of the current hardware data stream is quantified in real time, and the metrics are... Use time The standardized deviation distance between the measurements of the master and slave sensors is expressed by the following formula: In the formula, The main sensor at time Real-time measurement value, For redundant sensors at any time The real-time measurement value. This index normalizes the observation residual by introducing the measurement noise variance, aiming to eliminate spurious drift caused by physiological transient fluctuations, thereby accurately identifying the true data offset caused by changes in sensor physical performance, and providing a decision-making basis for implementing subsequent selective calibration strategies.

[0062] S24', finally performs adaptive selective calibration updates and parameter corrections. Based on the confidence level assessment index... The real-time distribution of data is used to implement a hierarchical adaptive update strategy to ensure the robustness of calibration behavior. When When the confidence level is less than 2, the sensor is determined to have sufficient confidence, and a standard recursive update is performed to correct the bias in the sensor state vector. With gain parameters ;when When the value is between 2 and 3, it is determined that the data has potential interference. The gain update weight of the Bayesian filter at the current moment is reduced (e.g., the correction gain is multiplied by a down-adjustment factor) to prevent over-correction due to instantaneous sensor fluctuations; while when... If the value is greater than or equal to 3, it is determined that there is a significant risk of data outliers or hardware failure, and the parameter update is automatically rejected.

[0063] Finally, the weights are dynamically reconstructed using step S24. The components are weighted and combined, and the deviation estimated at the current time S24' is used as the basis for the calculation. With gain parameters Perform the following reconstruction and restoration:

[0064] In the formula, As a clean data source, the high signal-to-noise ratio physiological feature signal of the final reconstructed output provides physical support for individualized baseline modeling and feature vector extraction in the subsequent S3 step; Discrete update time index for Bayesian filter calibration; For a moment The estimated sensor measurement bias is used to offset the additive error introduced by the drift of the flexible sensor; For a moment The estimated sensor gain parameters are used to compensate for the multiplicative error introduced by sensitivity fluctuations; The time-domain variable corresponding to the physiological signal; This is the upper limit of the number of modes that are empirically verified to be positively correlated with physiological characteristic frequencies and selected from the VMD decomposition results of the preceding step S21, in order to ensure that the reconstruction process contains only effective bioelectric or biophysical fluctuation components. The first one extracted in step S21 Each intrinsic mode component represents a macroscopic physiological rhythm component with a large energy proportion in the signal. The set of valid intrinsic mode function indices retained after the spectral purity verification in step S23 serves as a logical checkpoint, ensuring that only components that pass the "physiological / artifact" energy comparison test can participate in the final reconstruction. This corresponds to the further refined decomposition of the chattering term in step S22, resulting in the first... Each intrinsic mode function represents a microscale component in the signal that reflects the underlying physiological ripples. The weighted coefficient system generated by the Sigmoid logic function mapping in step S23 has a value range of [value range missing]. This weighting coefficient reflects the first... By assigning high weights to components with dominant physiological components and physically suppressing components severely affected by motion artifacts, the system achieves adaptive restoration of physiological state evolution characteristics under complex dynamic disturbances. This reconstruction logic, through selective superposition of components at different scales, eliminates environmental drift and motion noise while preserving the morphological characteristics of the original physiological signals of the monitored object to the greatest extent possible. Furthermore, the entire process involves only technical enhancements at the signal processing level and does not involve the determination of pathological properties.

[0065] S3. Based on physiological characteristic signals, determine a dynamic reference benchmark that adapts to the state of the monitored object, and combine the dynamic reference benchmark to perform multi-scale time series feature extraction to obtain multi-scale time series feature vectors.

[0066] In some specific embodiments, the multi-scale time-series feature vector includes instantaneous features, short-term statistical features, and long-term trend features. Specifically, multi-scale time-series feature extraction is performed based on a dynamic reference benchmark and real-time monitoring sequences to generate high-dimensional feature vectors characterizing the evolution of physiological states. The extraction logic covers signal attributes at different time resolutions: at the instantaneous scale, the signal amplitude and rate of change at the current moment are calculated in real time to characterize the immediate response features of physiological quantities; at the short-term scale, time-domain statistical indicators and frequency-domain energy distribution features are extracted using a sliding window to describe the fluctuation purity and spectral structure of physiological rhythms within short periods; at the long-term scale, the diurnal cyclical rhythm and trend shift of signals are identified by analyzing long-term time series to construct macroscopic features characterizing the long-term evolution of physiological homeostasis. After fusion processing, the above multi-scale time-series features form a standardized feature vector, realizing the mapping of time-domain signals to feature space, providing structured data support for subsequent risk warning decisions, and this process only involves the mathematical analysis of physical signal distribution characteristics, without involving any diagnostic conclusions.

[0067] In some specific embodiments, step S3 includes: S31, Preprocess the physiological feature signal based on the physical logic threshold to remove outliers and non-physiological abnormalities that exceed the reasonable physical range; S32, apply Gaussian mixture model or kernel density estimation to fit the historical data distribution characteristics of each physiological parameter in order to establish a probability distribution function that characterizes the random evolution of the signal and delineate the dynamic distribution range; S33 identifies the real-time physical state labels of the monitored object, including resting, active, and sleeping states. Differential distribution modeling of physiological signal characteristics under different physical state labels is then performed to generate a dynamic reference benchmark adaptively aligned with the real-time physiological behavior state of the monitored object. The physical state labels are obtained by extracting macroscopic envelope features from the physiological feature signals output in step S2. The macroscopic envelope features include the instantaneous amplitude sequence obtained by performing a Hilbert transform on the trend component, and the time-varying energy proportion distribution of each component. Specifically, the macroscopic envelope features are constructed by performing a Hilbert transform on the extracted trend term to obtain its instantaneous amplitude sequence (reflecting the "DC bias" fluctuation profile caused by large body movements or sensor pressure changes), combined with the energy proportion distribution of each component (including physiological energy). With artifact energy The time-varying distribution of the monitored object. When the monitored object switches from static mode to dynamic mode, the physical energy envelope of the relevant data channel undergoes a sudden jump in a specific frequency band; by monitoring the fluctuation variance of the amplitude sequence and the center frequency shift of the energy distribution, adaptive identification of the physical state label of the monitored object can be achieved.

[0068] S34 uses exponentially weighted moving average or recursive filtering algorithms to continuously track the statistical center shift of the probability distribution function (specifically, the position of the peak probability density in the mean vector of each component in the Gaussian mixture model or the kernel density estimate) to achieve adaptive compensation for baseline drift. S35, based on a dynamic reference benchmark, extracts multi-scale time-series feature vectors. The multi-scale time-series feature vectors include instantaneous features for characterizing immediate response features, short-term statistical features for describing short-period fluctuation attributes, and long-term trend features for characterizing the long-term evolution of physiological homeostasis.

[0069] S4. An adaptive early warning algorithm based on a hierarchical decision architecture is used to detect anomalies in multi-scale time-series feature vectors. Anomaly propagation patterns among multiple parameters are identified through parameter correlation topology. An attention mechanism is used to fuse multi-dimensional technical context metadata to perform dynamic weight correction and time-series trend prediction. Finally, differentiated hierarchical early warning command signals are output.

[0070] In some specific embodiments, the adaptive early warning algorithm adopts a hierarchical decision architecture. It generates graded risk data and outputs differentiated early warning command signals by calculating the deviation of the measured value from the dynamic normal range data, constructing a parameter correlation model based on graph neural network (GNN) to identify abnormal propagation patterns in the physiological network, using the attention mechanism to fuse structured multidimensional technical contextual metadata to adjust the risk factor weights, and using a time-series prediction architecture to predict the risk evolution trend.

[0071] Specifically, the adaptive early warning algorithm adopts a hierarchical decision-making architecture: The first layer is a single-parameter risk contribution assessment, which calculates the statistical deviation of real-time monitoring values ​​from a dynamic reference baseline (e.g., ...). -score or percentile), and generate an independent risk contribution score based on the preset sensor channel weights.

[0072] The second layer is multi-parameter collaborative anomaly identification, which identifies anomaly propagation patterns among multiple parameters through parameter association topology. Specifically, it involves constructing a parameter association model based on a graph neural network (GNN). Identify and monitor abnormal propagation patterns in the network. This includes the set of nodes. A set consisting of different types of sensor nodes , This serves as an index for parameter categories. The specific technical details and associated feature dimensions of each sensing node are shown in Table 1; in this embodiment, the node set... It covers frequency characteristic sources (such as heart rate and respiratory rate), pressure characteristic sources (such as systolic blood pressure and diastolic blood pressure), concentration characteristic sources (such as blood oxygen saturation, blood glucose, and lactate), and temperature characteristic sources. Feature vector For each node The associated high-dimensional technical indicators dynamically aggregate multi-scale temporal features extracted from the preceding S3 step. As shown in Table 2, the feature vectors... Specifically, it covers real-time sampled values, rates of change, and nonlinear dynamic exponents (such as the nonlinear exponent of heart rate and the exponential component of blood oxygen saturation), used to characterize the evolutionary properties of specific physical quantities from multiple criteria in both the time and frequency domains. Edge set This is a logical set characterizing the technological coupling relationships between different sensing nodes, describing the strength of topological constraints within the monitoring network. Specific edge categories, connected node pairs, and their associated weights are also included. The calculation mechanism is detailed in Table 2; edge set It covers output power coupling edges (connecting frequency characteristic sources and pressure characteristic sources), respiratory-circulatory compensation correlation edges, and metabolic response activation correlation edges, etc. Adjacency matrix The core matrix describing the strength of the correlation between parameters is composed of the edge set. Dynamic weights in Real-time computation structure. As shown in Table 2, the adjacency matrix is ​​dynamically updated through the product of the Pearson correlation coefficient and the physical dynamics estimate, or based on the mapping relationship between the activation function and the standardized reference benchmark. The element values ​​are used to quantitatively characterize the overall steady state of the monitored object at the system level.

[0073] Table 1. Multimodal sensing node set and node feature definition

[0074] Table 2. Edge set definition and association weight calculation mechanism

[0075] in, This refers to the Pearson correlation coefficient, which measures the degree of linear correlation between two physiological parameters over time series. The Sigmoid function is used to map the difference between physical quantities to the (0,1) interval, thereby achieving non-linear normalization of the weights. This is the time-domain rate of change operator, corresponding to the signal evolution slope defined in the eigenvector; It is a negative exponential kernel function (RBF kernel), used to calculate the consistency probability of parameter distribution based on Euclidean distance; It is a normalized exponential function used to achieve a probabilistic distribution of weights in a multi-path competition state; They are from the sensing nodes to Real-time sampling data, covering heart rate, systolic blood pressure, mean blood pressure, blood oxygen, respiratory rate, blood glucose, lactate, and perfusion index; and These are the synthetic eigenvalues ​​(such as estimated output values) estimated in real time by the system based on a multi-parameter matrix, and their corresponding standardized reference benchmarks. and The boundary between the real-time monitoring value of the average pressure and the system's preset technical alarm threshold; The baseline proportion of the respiratory-circulatory compensatory change rate of the monitored subjects under steady state; and : The statistical center (mean) and dispersion (standard deviation) of the joint distribution of metabolic characteristics; The initial reference level for biochemical characteristics determined in the individualized baseline model of the monitored subjects; the formula above includes norm, baseline, or Parameters with suffixes such as "current" are all dynamically provided by the dynamic reference baseline that is adaptively adjusted according to the state of the monitored object in step S3; that is, based on the current physical state label of the monitored object (such as resting, active, or sleeping), the statistical center or distribution characteristics are extracted in real time from the corresponding probability distribution function as the reference value. The parameters of the identification are acquired in real time from the self-sensing array and kept synchronized with the instantaneous features and short-term statistical features in the multi-scale time-series feature vector. By nonlinearly combining the dynamic features reflecting the real-time physiological fluctuations of the monitored object with the dynamic reference benchmark reflecting the individualized steady-state law, the physical attributes of the sensing nodes (such as pressure, frequency, concentration, etc.) are transformed into graph neural network topological attributes (edge ​​weights). The dimensional transformation of the model. This modeling method based on dynamic comparison of "benchmark-feature" provides precise mathematical support for the subsequent identification of abnormal propagation patterns that deviate from the individualized steady-state trajectory in the perception network, ensuring that the early warning decision can adaptively exclude the interference of physiological baseline drift caused by the switching of normal behavioral states.

[0076] The third layer involves technology context fusion and weight adjustment. An attention mechanism is used to non-linearly fuse the multi-dimensional technology context metadata (as shown in Table 3) with physiological feature vectors. The technology context metadata includes historical state descriptors encoded by embedded vectors, external device intervention parameters associated with time sequences, and artificial event markers encoded by discrete events. Based on the currently identified technology operation mode, preset context gating rules are dynamically applied to the weight coefficients of the early warning algorithm. With technical early warning threshold Perform real-time online corrections.

[0077] Table 3 shows the nonlinear fusion of multidimensional technological context metadata and physiological feature vectors.

[0078] Specifically, the execution logic of the context gating rules includes: when a periodic extracorporeal circulation intervention mode is detected where the external auxiliary circulation device is in operation, a logic mutual exclusion and feature switching strategy is implemented by resetting the weights of susceptible parameters to zero (such as setting the serum creatinine weight). This is used to implement signal interference suppression, while simultaneously activating intervention feature monitoring logic to compensate the weights of relevant operational indicators (such as ultrafiltration rate) to the effective monitoring level (such as setting...). This enables precise monitoring under external physical intervention; when a high-sensitivity monitoring mode with an activation cycle of less than 24 hours is detected or a specific trigger condition is identified, weight compensation and threshold correction logic is executed to increase the risk contribution weight of cyclic feature components (such as increasing the heart rate weight). And adaptively relax the technical tolerance boundaries of specific parameters (such as setting the blood oxygenation threshold). This enables sensitive detection of initial, minute physical fluctuations. When a low metabolic response pattern is detected, where the activity index of the monitored object is below a preset threshold, a sensitivity suppression and feature enhancement strategy is implemented. This involves reducing the weight of respiratory feature components (e.g., ...). To reduce non-technical false alarms caused by variations in signal rhythm and to increase the technical weight of carbon and oxygen monitoring indicators (such as...). In power support intervention modes where the intensity of external power support equipment intervention exceeds a critical value, benchmark bias adjustment logic is executed by increasing the offset of the target physical quantity benchmark (e.g., setting the mean arterial pressure target boundary). Simultaneously, auxiliary feature monitoring sequences are activated to compensate for physiological type shifts introduced by external intervention. Furthermore, when the system identifies a physical displacement pattern triggered by a movement event marker or a significant physical displacement, it enters an anti-interference movement monitoring mode. Through logical judgment, selective channel suppression is performed, suspending the early warning response of non-critical parameter channels and retaining only components with strong periodic rhythmic characteristics for risk synthesis, ensuring the stability of the early warning command signal in complex dynamic physical environments. Additionally, symbols... This refers to the risk contribution weight of each feature source in the initial state, while the symbol... Refers to the effective warning weight after dynamic correction by the context gating rules; subscript and These are respectively mapped to intervention parameter nodes (such as serum creatinine and ultrafiltration rate) defined in the multidimensional technical context metadata.

[0079] The fourth layer predicts the temporal evolution trend, using Long Short-Term Memory (LSTM) networks or Transformer architectures to predict the evolution trajectory of the risk score. When the predicted value reaches the physical safety boundary within a preset future period... At that time, the system outputs an advance warning command.

[0080] Finally, a tiered early warning command signal with differentiated technical attributes is output: low-risk level triggers an information recording level command; medium-risk level triggers a mobile terminal prompt level command; and high-risk level triggers the highest-level technical response sequence command. The entire process involves only mathematical analysis of physical feature distribution and equipment operating status, and the resulting early warning command signal is used to drive the automated feedback of the monitoring system, without involving the diagnosis and determination of the pathological nature of the monitored object.

[0081] Continue to refer to Figure 2 As an implementation of the above method, in a second aspect, this application provides an embodiment of the framework diagram 200 of a multi-parameter data processing system for flexible sensor arrays and adaptive early warning. This system embodiment is similar to... Figure 1 Corresponding to the illustrated method embodiment, this system can be specifically applied to various electronic devices. The system 200 includes a signal acquisition module 201, a signal collaborative reconstruction module 202, a dynamic benchmark modeling module 203, and a hierarchical early warning decision module 204, all interconnected, wherein: The signal acquisition module 201 is configured to acquire multi-dimensional raw physiological signals continuously collected by a flexible multimodal sensor array; The signal collaborative reconstruction module 202 is configured to use a joint algorithm based on variational mode decomposition and adaptive noise complete set empirical mode decomposition on multi-dimensional physiological raw signals. It performs spectral purity analysis on each component to map dynamic reconstruction weights; utilizes redundant measurements between multi-dimensional physiological raw signals for online drift detection; applies Bayesian filtering to estimate sensor bias and gain error; performs selective calibration update based on online confidence index; and finally combines dynamic reconstruction weights and calibration parameters to reconstruct the denoised physiological feature signal. The dynamic benchmark modeling module 203 is configured to determine a dynamic reference benchmark that adaptively adjusts with the state of the monitored object based on physiological feature signals, and performs multi-scale time series feature extraction in combination with the dynamic reference benchmark to obtain multi-scale time series feature vectors. The hierarchical early warning decision module 204 is configured to use an adaptive early warning algorithm based on a hierarchical decision architecture to perform anomaly detection on multi-scale time-series feature vectors, use an attention mechanism to fuse multi-dimensional technical context metadata to perform dynamic weight correction and time-series trend prediction, and output differentiated hierarchical early warning command signals.

[0082] Although the principles of the present invention have been described in detail above with reference to preferred embodiments, those skilled in the art should understand that the above embodiments are merely illustrative explanations of the implementation of the present invention and are not intended to limit the scope of the present invention. The details in the embodiments do not constitute a limitation on the scope of the present invention. Any obvious changes, such as equivalent transformations or simple substitutions, based on the technical solutions of the present invention without departing from the spirit and scope of the present invention fall within the protection scope of the present invention.

Claims

1. A multi-parameter data processing method for flexible sensor arrays and adaptive early warning, characterized in that, The method includes: S1. Acquire multi-dimensional raw physiological signals continuously collected by a flexible multimodal sensor array; S2. A joint algorithm based on variational mode decomposition and adaptive noise complete set empirical mode decomposition is used for the multi-dimensional physiological raw signal. Spectral purity analysis is performed on each component to map dynamic reconstruction weights. Online drift detection is performed using redundant measurements among the multi-dimensional physiological raw signals. Bayesian filtering is applied to estimate sensor bias and gain error. Selective calibration updates are performed based on online confidence indices. Finally, the denoised physiological feature signal is reconstructed by combining the dynamic reconstruction weights and calibration parameters. S3. Based on the physiological characteristic signal, determine a dynamic reference benchmark that adaptively adjusts with the state of the monitored object, and combine the dynamic reference benchmark to perform multi-scale temporal feature extraction to obtain a multi-scale temporal feature vector; S4. The adaptive early warning algorithm of the hierarchical decision architecture is used to detect anomalies in the multi-scale time-series feature vector. The anomaly propagation mode between multiple parameters is identified through the parameter correlation topology. The attention mechanism is used to fuse multi-dimensional technical context metadata to perform dynamic weight correction and time-series trend prediction. Finally, differentiated hierarchical early warning command signals are output.

2. The multi-parameter data processing method for flexible sensor array and adaptive early warning according to claim 1, characterized in that, The flexible multimodal sensing array includes: a metabolite sensor based on microchannel sweat collection, an interstitial fluid biochemical sensor based on microneedle electrodes, a hemodynamic sensor based on a flexible ultrasonic transducer, and an electrocardiogram / electromyogram sensor based on a conductive polymer.

3. The multi-parameter data processing method for flexible sensor array and adaptive early warning according to claim 1, characterized in that, The S2 step, which uses the joint algorithm to separate the trend term and the oscillation term, specifically includes: By presetting the number of modes, balance parameters, and convergence tolerance, the alternating direction multiplier method is used to analyze the original physiological signal into multiple bandwidth-limited intrinsic mode components, among which the center frequency is lower than that of the intrinsic mode components. The first mode is defined as the trend term representing the baseline drift, and the remaining modes are defined as oscillation terms covering the ultra-low frequency to high frequency range; For the aforementioned oscillation term, a set of intrinsic mode functions with distinct time scale characteristics is extracted by introducing controlled Gaussian white noise to perform recursive decomposition, and by using the ratio of the noise set number to the standard deviation of the additional noise for mean processing, so as to eliminate the mode aliasing between physiological components and motion noise.

4. The multi-parameter data processing method for flexible sensor array and adaptive early warning according to claim 3, characterized in that, The expression for the set of intrinsic mode functions is: In the formula, For the stripped-out first An intrinsic mode function component with a single characteristic frequency range; The total number of noise sets; The remaining energy sequence after the previous decomposition serves as the physical input for the current decomposition. These are the noise weighting factors for each order; For the first A sequence of Gaussian white noise; Extracting the first from the synthesized signal Empirical mode decomposition operator for features at each time scale.

5. The multi-parameter data processing method for flexible sensor array and adaptive early warning according to claim 1, characterized in that, Step S2, which involves performing spectral purity analysis on each component to map dynamic reconstruction weights, specifically includes: performing a frequency domain transformation on each intrinsic mode function to obtain its spectral distribution; combining this with a pre-defined motion artifact feature spectral library; calculating the energy belonging to the physiological signal band and the energy belonging to the motion artifact band by integrating the power spectral density; and using the Sigmoid logic function to map the spectral purity ratio to generate the dynamic reconstruction weights for each component, the calculation formula of which is: In the formula, For dynamic reconfiguration of weights; This is the slope control parameter; The energy carried by physiological signals, This represents the energy of the motion artifact frequency band.

6. The multi-parameter data processing method for flexible sensor array and adaptive early warning according to claim 1, characterized in that, Step S2 includes: A heterogeneous redundancy calibration architecture consisting of a main sensing channel and redundant reference channels is established, and a state transition equation is constructed. With observation equation The evolution of the sensor's physical performance is dynamically and recursively tracked; where, at time... state vector Includes the measurement bias term and gain error term to be estimated, and the observation vector. Reflects the real-time physical residual between the main sensing channel and the redundant reference channel; This is the state transition matrix, which describes the performance evolution of the sensor caused by the aging of flexible materials or environmental temperature drift. This is the process noise vector; The observation matrix; For the observed noise vector; Performing selective calibration updates based on online confidence metrics includes: Calculation time The standardized deviation distance between the measurements of the master and slave sensors is used as an online confidence index, and the calculation formula is as follows: In the formula, As a confidence level indicator; The main sensor at time Real-time measurement values; For redundant sensors at any time Real-time measurement values; The variance is measured by the main sensor. Variance is measured for redundant sensors; According to the confidence index Real-time distributed implementation of hierarchical adaptive updates: when Perform standard recursive updates to correct sensor biases. With gain parameters ;when When, restricted calibration is implemented by reducing the update weights; when The system will reject the parameter update and trigger a sensor self-test. The formula for the physiological characteristic signal is as follows: In the formula, The final reconstructed output is a high signal-to-noise ratio physiological feature signal; The discrete update time index for Bayesian filter calibration; For a moment Estimated sensor measurement bias; For a moment Estimated sensor gain parameters; The time-domain variable corresponding to the physiological signal; This represents the upper limit of the number of modes; For the first One intrinsic mode component; For the set of valid intrinsic mode function indices; For the first One intrinsic mode function; For dynamic reconfiguration of weights; The denoised physiological feature signal is input as a clean data source into step S3 to perform dynamic reference benchmark modeling and feature extraction.

7. The multi-parameter data processing method for flexible sensor array and adaptive early warning according to claim 1, characterized in that, Step S3 includes: S31, Preprocess the physiological feature signal based on the physical logic threshold; S32, apply Gaussian mixture model or kernel density estimation to fit the historical data distribution characteristics of each physiological parameter in order to establish a probability distribution function that characterizes the random evolution of the signal and delineate the dynamic distribution range; S33, identify the real-time physical state label of the monitored object, the physical state label includes resting, active and sleeping, and perform differentiated distribution modeling of physiological signal characteristics under different physical state labels to generate a dynamic reference benchmark that adaptively aligns with the real-time physiological behavior state of the monitored object; wherein, the physical state label is obtained by extracting macroscopic envelope features from the physiological feature signals output in step S2, the macroscopic envelope features include the instantaneous amplitude sequence obtained by performing Hilbert transform on the trend component, and the time-varying distribution of the energy proportion of each component; S34, using an exponentially weighted moving average or recursive filtering algorithm to continuously track the statistical center shift of the probability distribution function in order to achieve adaptive compensation for baseline drift; S35, extract multi-scale time-series feature vectors based on the dynamic reference benchmark. The multi-scale time-series feature vectors include instantaneous features for characterizing immediate response features, short-term statistical features for describing short-period fluctuation attributes, and long-term trend features for characterizing the long-term evolution of physiological homeostasis.

8. The multi-parameter data processing method for flexible sensor array and adaptive early warning according to claim 1, characterized in that, The adaptive early warning algorithm adopts a hierarchical decision-making architecture, specifically including: The single-parameter risk assessment layer calculates the statistical deviation of the real-time monitoring value relative to the dynamic reference benchmark and generates an independent-dimensional risk contribution score. A multi-parameter collaborative identification layer is constructed to build a parameter association model based on a graph neural network as the parameter association topology, which is used to identify abnormal propagation patterns in the monitoring network and calculate the system-level risk index. The technology context fusion layer uses an attention mechanism to fuse multi-dimensional technology context metadata and calls context gating rules to perform real-time online correction of warning weights and technology warning thresholds. The temporal evolution prediction layer predicts the trajectory of risk evolution and executes advance warnings when the predicted value reaches the physical safety boundary.

9. The multi-parameter data processing method for flexible sensor array and adaptive early warning according to claim 8, characterized in that, The multi-parameter collaborative recognition layer uses a parameter association model. Assess the overall steady state; among which, It is a set of nodes composed of different categories of physiological parameters, including frequency feature sources, pressure feature sources, concentration feature sources and temperature feature sources; The set of edges characterizing the biological coupling relationships between parameters includes output power coupling edges, pressure-rhythm correlation constraint edges, respiratory-circulatory compensation correlation edges, and metabolic response activation correlation edges; The feature vector includes the instantaneous sampled values, time-domain rate of change, and nonlinear dynamic exponent of each sensing node; The adjacency matrix, which describes the strength of the correlation between parameters, is calculated in real time by the dynamic correlation weights in the edge set and is used to quantitatively characterize the system-level risk index of the overall steady state of the monitored object. The calculation logic for the correlation weights includes: calculating the weight of the output power coupling edge based on the product of the Pearson correlation coefficient and the physical dynamics estimate between the signal sequences of the two sensing nodes; calculating the weight of the respiratory-circulatory compensation correlation edge based on the deviation of the ratio of the characteristic change rate between the two sensing nodes relative to the dynamic reference benchmark; applying radial basis functions to calculate the Euclidean distance of the composite metabolic feature from the dynamic reference benchmark, and calculating the weight of the metabolic response activation correlation edge based on the negative exponential mapping.

10. The multi-parameter data processing method for flexible sensor array and adaptive early warning according to claim 8, characterized in that, The execution logic of the scenario gating rule includes: when an operation mode involving external equipment is detected, the warning weight of the interference parameter is suppressed to zero weight to shield the false disturbance introduced by physical intervention; when a specific warning sensitive mode is detected, the risk contribution weight of the corresponding feature component is increased, and the tolerance boundary of the technical warning threshold is reduced simultaneously to enhance sensitivity; when an external power support intervention mode is detected, the offset of the target physical quantity benchmark is increased to compensate for the feature phenotypic offset introduced by external power support.

Images