Positioning anti-falling method and anti-falling system based on multimode signal processing, and wearable device

Abstract

The invention relates to a positioning anti-falling method and anti-falling system based on multimode signal processing and wearable equipment. In order to solve the problems that in the prior art, multi-source signal noise influence is large, time sequence alignment occurs frequently, and judgment errors occur frequently, two-dimensional Gaussian kernel smoothing and normalization are conducted on plantar pressure array data, a pressure distribution matrix with continuous space is generated, then a pressure center track and motion manifold characteristics of the pressure center track are extracted, and a gait phase coordinate system is established. The method comprises the following steps: calculating the dominant phase and offset of each signal through nonlinear time-phase mapping in combination with acceleration, a gyroscope and a gravity center offset signal, completing multi-modal signal confidence scoring by using an offset penalty and stability weight mechanism, and finally realizing risk grade judgment and providing physical alarm by using a lightweight decision tree model. According to the method, the time sequence consistency and data robustness of different sensing source signals are improved, accurate and rapid early warning of the falling risk is achieved, meanwhile, the dynamic self-adaptive calibration capability is achieved, and the false alarm and leakage alarm probability is reduced.

Description

Technical Field

[0001] This invention relates to the field of multimodal wearable detection and early warning systems, and in particular to a positioning and drop prevention method, drop prevention system and wearable device based on multimodal signal processing. Background Technology

[0002] Multimodal wearable positioning and fall prevention technology is currently gaining widespread attention in applications such as elderly health monitoring, portable fall warning, and gait recognition. Existing integrated wearable devices on the market, such as smart insoles, wristbands, and multi-parameter health monitoring watches, generally integrate accelerometers, gyroscopes, and pressure sensors, utilizing multimodal fusion algorithms to identify user movement behavior and provide fall warnings. The mainstream technological trends mainly manifest in three approaches: first, multi-signal fusion decision mechanisms based on signal strength weighting or Dempster-Shafer evidence theory; second, end-to-end deep learning modeling paths (such as RNN, CNN fusion networks, and temporal attention models); and third, adaptive feature extraction and hierarchical alarm models. These approaches balance data fusion and real-time response requirements, evolving along a product iteration roadmap of "early detection, early warning, low latency, and high accuracy."

[0003] In typical application scenarios, taking drop-resistant wearable devices as an example, the following process is usually adopted: sudden movement of the user is detected by acceleration and gyroscope signals, and the current gait stage is estimated by combining the plantar pressure transformation. Finally, the risk level is output by the embedded algorithm, triggering sound, light, and vibration level alarms. Some high-level systems also support adaptive recognition in complex terrain walking scenarios such as stairs and slopes.

[0004] However, existing technologies still have the following technical shortcomings and unmet needs: First, the signal confidence estimation mechanism is insufficient. Currently, many systems use simple weighted averaging or fixed threshold strategies to fuse multimodal sensor signals, failing to dynamically reflect the true reliability of sensors under different scenarios, gait states, and sudden conditions. For example, when a user walks on a slope or stairs, the peak of the foot pressure and inertial sensor time series often shifts. Existing systems do not dynamically adjust the confidence levels of each signal, easily leading to signal mismatch and increasing the risk of false alarms or missed alarms.

[0005] Second, black-box modeling methods such as deep learning rely on large-scale sample libraries and high-computing-power environments, making them difficult to deploy in real time on low-power, resource-constrained microcontroller chips (such as Cortex-M4). Furthermore, they have poor interpretability and traceability, making it difficult to gain the trust of medical terminals and regulatory authorities.

[0006] Third, the system lacks robustness in complex gait scenarios. Existing methods do not physically model the correlation between multi-source signals such as acceleration, gyroscope, and pressure and specific phases of the gait cycle. They also lack dynamic evaluation of the temporal consistency, phase shift, and evolutionary stability of multi-mode signals. This makes it difficult to accurately adapt to individual user gait differences and special working conditions (such as carrying heavy objects or wearing different types of footwear), resulting in static and rigid alarm strategies.

[0007] Fourth, there is a lack of root cause misjudgment and system self-adjustment mechanisms. Existing technologies do not provide local logging and adaptive parameter update tools for severe signal mismatch scenarios such as "phase lockout," and cannot support individualized dynamic parameter calibration, affecting the accuracy and user experience over long-term use. Summary of the Invention

[0008] This application provides a positioning and drop protection method, a drop protection system, and a wearable device based on multi-mode signal processing, aiming to solve one of the problems or issues of the prior art mentioned in the background.

[0009] Firstly, the positioning and drop protection method, drop protection system, and wearable device based on multi-mode signal processing provided in this application specifically include: S1: Perform two-dimensional Gaussian kernel smoothing and normalization on the original pressure data to generate a spatially continuous pressure distribution matrix; S2: Extract the pressure center trajectory from the pressure distribution matrix and construct a local motion manifold by combining the pressure center displacement vectors of three consecutive frames; S3: Identify the maximum pressure cluster in the heel region within a single step cycle as the main pressure peak, and establish a gait phase coordinate system with the time corresponding to the main pressure peak as the phase zero point. S4: Map the timestamps of each input signal, including the triaxial acceleration energy envelope, the gyroscope angular velocity absolute value envelope, and the extreme point of the center of gravity shift rate of change, to the gait phase coordinate system, obtain the dominant phase position of the input signal on the phase axis, and calculate its offset relative to the pressure peak phase. S5: Based on the phase offset of each input signal and its standard deviation of offset over five consecutive gait cycles, evaluate phase consistency and evolutionary stability, and generate an initial confidence score. S6: Calculate the final confidence scores of acceleration, gyroscope and center of gravity offset signals based on the initial confidence scores, and input the three scores into the preset lightweight decision tree model to output the corresponding alarm level judgment result.

[0010] Secondly, this application provides a positioning and fall prevention system based on multi-mode signal processing, characterized in that it uses the above-mentioned positioning and fall prevention method based on multi-mode signal processing to provide fall prevention warnings for the elderly.

[0011] Thirdly, this application provides a wearable device equipped with the aforementioned positioning and drop protection system based on multi-mode signal processing.

[0012] The positioning and drop protection method, drop protection system, and wearable device based on multi-mode signal processing provided in this application have the following beneficial effects: (1) By constructing a multimodal physiological signal fusion mechanism based on gait phase alignment, the accuracy and robustness of motion risk identification are significantly improved, effectively overcoming the problem of frequent false alarms caused by signal asynchrony, noise interference or individual gait differences in dynamic walking scenarios of traditional wearable devices; This scheme establishes a unified periodic reference coordinate system with the pressure peak as the phase anchor point, and maps the acceleration energy envelope, angular velocity dominant phase and the rate of change of center of gravity offset to the phase axis in a unified manner, realizing cross-modal signal alignment in time-phase dual domain, avoiding the cumulative error caused by simple time alignment; Further, the phase consistency constraint and stability weighting mechanism are introduced, which can accurately identify and suppress instantaneous abnormal signals caused by sudden movement disturbances or sensor drift, ensuring that the judgment of key biomechanical indicators always focuses on stable and physiologically meaningful functional stages, thereby greatly improving the reliability of the system in complex real environments.

[0013] (2) Lightweight confidence modeling and hierarchical decision logic are adopted to achieve a balance between high sensitivity and high specificity with extremely low computational overhead, effectively meeting the multiple stringent requirements of wearable devices for real-time performance, low power consumption and clinical usability. Unlike traditional intelligent early warning methods that rely on large-scale training data and complex model inference, this solution is based entirely on the design of a scoring system based on physically interpretable biomechanical characteristics. It directly quantifies the credibility of each signal source through indicator functions and threshold rules, and dynamically adjusts the weights in combination with short-term stability statistics. It can adapt to the gait characteristics of different user groups without parameter tuning. The final alarm decision is driven by a minimal decision tree, which supports local fast inference and condition-triggered re-verification mechanism. The entire process is completed within an average delay of less than the preset time. It can run continuously on low-power microcontrollers without excessively consuming battery resources. At the same time, the system has adaptive evolution capabilities. It automatically identifies high-frequency abnormal scenarios and updates personalized tolerance parameters by aggregating "phase lockout" events daily, gradually optimizing the local model matching degree and forming a closed-loop individualized calibration path.

[0014] (3) The overall architecture balances engineering practicality and medical interpretability, not only avoiding compliance obstacles in medical supervision of black-box models such as deep learning, but also providing clear and observable intermediate variable outputs for clinical evaluation, enhancing the trust foundation of doctors and users in the system's judgment results; all core indicators such as COP trajectory, phase offset, stability standard deviation, and confidence score have clear biomechanical meanings, facilitating cross-validation and intervention strategy formulation in conjunction with professional knowledge; in addition, the local multimodal revalidation mechanism and edge log recording function provide structured data support for subsequent remote health analysis and rehabilitation process tracking, enabling the device to not only achieve immediate early warning, but also serve as a long-term gait quality monitoring tool for various application scenarios such as postoperative rehabilitation, elderly fall prevention, and athlete training management. In summary, this solution, while ensuring high-performance detection capabilities, achieves a deep integration of algorithm simplicity, system autonomy, and clinical adaptability, and has strong potential for industrialization and promotion value. Attached Figure Description

[0015] Figure 1 The main flowchart of the positioning and drop protection method, drop protection system and wearable device with multi-mode signal processing is provided for the preferred embodiment of this application.

[0016] Figure 2 This application provides a sub-flowchart of a positioning and drop protection method, a drop protection system, and a wearable device based on multi-mode signal processing for a preferred embodiment.

[0017] Figure 3 Another sub-flowchart is provided for a preferred embodiment of the present application regarding a positioning and drop protection method, a drop protection system, and a wearable device based on multi-mode signal processing. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0020] like Figure 1As shown, this application provides a positioning and drop protection method, a drop protection system, and a wearable device based on multi-mode signal processing, specifically including: S1: Perform two-dimensional Gaussian kernel smoothing and normalization on the original pressure data to generate a spatially continuous pressure distribution matrix; S2: Extract the pressure center trajectory from the pressure distribution matrix and construct a local motion manifold by combining the pressure center displacement vectors of three consecutive frames; S3: Identify the maximum pressure cluster in the heel region within a single step cycle as the main pressure peak, and establish a gait phase coordinate system with the time corresponding to the main pressure peak as the phase zero point. S4: Map the timestamps of each input signal, including the triaxial acceleration energy envelope, the gyroscope angular velocity absolute value envelope, and the extreme point of the center of gravity shift rate of change, to the gait phase coordinate system, obtain the dominant phase position of the input signal on the phase axis, and calculate its offset relative to the pressure peak phase. S5: Based on the phase offset of each input signal and its standard deviation of offset over five consecutive gait cycles, evaluate phase consistency and evolutionary stability, and generate an initial confidence score. S6: Calculate the final confidence scores of acceleration, gyroscope and center of gravity offset signals based on the initial confidence scores, and input the three scores into the preset lightweight decision tree model to output the corresponding alarm level judgment result.

[0021] Specifically, as an example, step S1 involves performing two-dimensional Gaussian kernel smoothing and normalization on the raw pressure data collected by the insole foot pressure sensor array to eliminate local noise interference and generate a spatially continuous pressure distribution matrix. This specifically includes: S1.1: Obtain the raw pressure data frames output by the insole foot pressure sensor array. Each frame contains discrete pressure sampling values ​​from 8 sensing units, including 3 in the forefoot, 2 in the arch, and 3 in the heel, forming an 8-dimensional vector spatial pressure snapshot, which serves as the initial input condition for subsequent spatial filtering and topology reconstruction.

[0022] When acquiring the raw pressure data frames from the insole's plantar pressure sensor array, a multi-channel synchronous sampling method (sampling frequency: 100Hz, quantization accuracy: 12bit) is used to achieve parallel signal acquisition from eight pressure sensing units: three in the forefoot, two in the arch, and three in the heel. The timestamp alignment accuracy of each channel is maintained at ≤5ms to ensure the timing consistency of the multi-point pressure data. An analog-to-digital converter (ADC) converts the analog signals from each pressure sensing unit into digital pressure values, forming a pressure sampling vector containing eight elements. ,in This represents the digitized pressure value of the i-th pressure sensing unit at the current moment. Further, this is achieved through a calibration coefficient matrix. The sampling vector is calibrated and corrected, that is, each pressure value is subjected to... The calculations are performed to eliminate amplitude deviations introduced by differences in sensor sensitivity. Furthermore, the background pressure baseline value for each sensing unit is calculated using a dynamic zero-point offset calibration algorithm (window length: 2 seconds). And subtract the corresponding baseline from the calibrated pressure value to achieve Zero-point adjustment is performed to suppress long-term drift effects. Eight pressure values ​​collected at the same time are combined into a spatial pressure snapshot vector through a time window buffer and a high-precision timestamp parameter is added to form an indexable set of raw pressure data frames, laying the data foundation for the subsequent two-dimensional spatial grid mapping in S1.2.

[0023] Through algorithms such as synchronous sampling, calibration correction, and zero-point offset calibration, the initial analog signal is transformed into a combination of timestamp and pressure vector data with physiological reference correction, realizing high-precision digitization and standardization of the original pressure data, and providing consistent input conditions for spatial filtering and topology reconstruction.

[0024] For example, during a stable walking scenario, the sampled values ​​of the three sensors in the forefoot are 312, 298, and 305 (unit: digital-to-analog conversion count), the sampled values ​​of the two sensors in the arch are 280 and 275, and the sampled values ​​of the three sensors in the heel are 340, 335, and 342. After ADC quantization, a vector [312, 298, 305, 280, 275, 340, 335, 342] is formed. The corresponding calibration coefficient matrix is ​​[1.02, 0.98, 1.00, 1.01, 0.99, 1.03, 1.02, 1.01]. After calibration, the outputs of each sensor are calculated as 312 × 1.02 = 318.24, 298 × 0.98 = 292.04, etc. The background baseline value is calculated by averaging over a 2-second window. For example, the baseline of the first sensing unit in the forefoot is 15.3, which is subtracted to form a correction value of 302.94. The final output raw pressure data frame is [302.94, 276.74, 290.00, 267.55, 257.25, 335.80, 320.70, 336.42], with an appended sampling timestamp of 1638456123.056. When this data frame enters the subsequent spatial mapping stage, it can significantly improve the amplitude accuracy and spatial distribution continuity of the subsequent two-dimensional Gaussian smoothing.

[0025] S1.2: Based on the original pressure data frame, construct a two-dimensional spatial grid mapping relationship, distribute the 8 discrete sensing units according to their physical layout coordinates (x_i, y_i) on the insole into an 8×8 resolution spatial grid, and generate a sparse pressure space matrix P_sparse(x,y) to establish the geometric correspondence between discrete sampling points and continuous foot area.

[0026] Based on the 8-dimensional spatial pressure snapshot vector output by S1.1, a physical layout mapping method for insoles is adopted (parameters: fixed geometric coordinates x_i, y_i corresponding to each sensing unit, with coordinate units in millimeters) to achieve geometric position matching between discrete sensing units and a two-dimensional spatial grid. By establishing a one-to-one correspondence between the pressure value of each sensing unit and its physical layout coordinates, a sensing point set data structure {(x_i, y_i, p_i)} is generated, providing a basic coordinate system for subsequent spatial rasterization.

[0027] A two-dimensional spatial interpolation algorithm (parameters: 8×8 resolution grid, interpolation method: nearest neighbor fill, initial value of unsampled points: 0) is used to map the sensor point set data structure to a grid matrix. This mapping assigns the pressure value p_i of each sensor unit to the corresponding grid coordinates (x_g, y_g), forming a sparse pressure space matrix P_sparse(x,y). This matrix spatially contains all sampled points of the target foot region, but zero values ​​are present at unsampled locations.

[0028] Furthermore, by using a spatial coordinate normalization method (parameters: normalization range [0,7], step size of 1 grid unit), the actual millimeter coordinates of the sensing unit are mapped to the grid coordinate system, ensuring that a uniform 8×8 spatial resolution is maintained for different sizes of insoles and sensor layouts. This process unifies the spatial reference of different batches of hardware devices, laying the foundation for cross-device data fusion.

[0029] By using the position calibration vector construction method (parameters: 50 calibration samples, force range 0-500N), the sensor position error compensation parameters obtained from the experimental calibration are applied to the sparse matrix generation process to correct the mapping coordinates of each sensing unit, thereby eliminating spatial positioning deviations caused by sensor installation tolerances or insole deformation.

[0030] A structural index mapping table (content: raster coordinate index → ​​sensor unit number) is used to realize the reverse correspondence lookup from the raster matrix to the physical sensor network, providing an index acceleration mechanism for local neighborhood weighting during subsequent Gaussian convolution filtering.

[0031] Using the aforementioned two-dimensional spatial grid mapping algorithm, the pressure values ​​of the eight discrete sampling points are precisely embedded into the 8×8 sparse spatial matrix P_sparse(x,y) of the continuous foot region, realizing the geometric correspondence between discrete sensing points and continuous space, and providing complete spatial input conditions for the next step of two-dimensional Gaussian kernel smoothing.

[0032] For example, in a scenario with a foot length of 265mm and an effective insole width of 90mm, the physical coordinates of the three forefoot sensing units are (25,15), (45,15), and (65,15), the coordinates of the two arch sensing units are (35,45) and (55,45), and the coordinates of the three heel sensing units are (25,75), (45,75), and (65,75), with pressure values ​​of 120N, 140N, 135N, 100N, 95N, 150N, 160N, and 155N, respectively. The millimeter coordinates are mapped to a grid coordinate system according to the ratio of 90mm → 8 grid widths and 265mm → 8 grid lengths to obtain the corresponding grid position indices, such as (2,0), (4,0), and (6,0). During interpolation, the pressure values ​​of each sensing unit are directly assigned to their corresponding grid positions, while unsampled grids are set to zero, forming a sparse pressure matrix P_sparse(x,y). To eliminate installation errors, calibration compensation parameters Δx and Δy are applied, with a maximum correction of 0.2 grid units, to adjust the mapped coordinates. The resulting sparse matrix accurately reflects the plantar load distribution characteristics of the current gait segment under spatial location and pressure value constraints, providing high-fidelity input for subsequent Gaussian spatial smoothing.

[0033] S1.3: Apply a two-dimensional Gaussian kernel convolution operation to the sparse pressure space matrix P_sparse(x,y), and use a Gaussian kernel function G(x,y)=exp(-(x^2+y^2) / (2σ^2)) with a standard deviation σ=1.5 grid units for spatial smoothing to suppress high-frequency noise caused by local contact jitter, and generate a denoised continuous pressure distribution matrix P_smooth(x,y).

[0034] S1.4: Calculate the L2 norm ||P_smooth||_2 of all grid pressure values ​​in the continuous pressure distribution matrix P_smooth(x,y). Based on this norm, perform global amplitude normalization on the matrix to generate a normalized pressure distribution matrix P_norm(x,y)=P_smooth(x,y) / ||P_smooth||_2, so that the pressure distribution during different time periods has comparable amplitude and eliminates the interference caused by differences in individual weight or walking force.

[0035] S1.5: Output the normalized pressure distribution matrix P_norm(x,y) as the input data for the pressure center trajectory extraction module in the next stage. This matrix has spatial continuity and amplitude normalization characteristics, and constitutes the reference field for gait phase zero-point detection and multi-mode signal phase alignment.

[0036] Step S2: Extract the pressure center trajectory based on the normalized pressure distribution matrix, and construct a local motion manifold by combining the pressure center displacement vectors of three consecutive frames, which is used to characterize the dynamic load transfer mode during gait. Specifically, it includes: S2.1: Obtain the pressure distribution matrix sequence after two-dimensional Gaussian kernel smoothing and normalization. Based on the spatial coordinates of each sensor position in each frame matrix and its corresponding pressure intensity value, calculate the weighted average position and generate the pressure center coordinates at each sampling time. Perform vector difference operation on the pressure center coordinate sequence within the continuous time window to obtain the pressure center displacement vector between adjacent frames, so as to characterize the instantaneous migration direction and rate of the foot load in the support phase.

[0037] S2.2: Construct a local velocity field based on the displacement vector of the pressure center in three consecutive frames. Use the displacement vector of the current frame as the main driving vector. Perform normalized projection processing on the displacement vectors of the previous and next frames. Calculate the ratio of the cosine of the angle between the vector and the main vector to the amplitude to evaluate the motion consistency and acceleration change trend during load transfer. If the cosine of the angle is less than 0.85 and the amplitude suddenly increases by more than 1.5 times the standard deviation of the mean, it is marked as a potential gait phase transition point.

[0038] Based on a sequence of pressure center displacement vectors from three consecutive frames, a local velocity field construction method is employed (input conditions: pressure center displacement vectors from the current frame, the previous frame, and the next frame). The displacement vector from the current frame is used as the basis for this method.<v_i> As the main driving vector, a benchmark for local dynamic analysis is established.

[0039] Furthermore, by using the vector normalization projection method (parameter: unit vectorization mode), the displacement vectors of the previous frame and the next frame are projected onto the direction of the main driving vector, forming a projection vector set, which is used to measure the degree of alignment between the load transfer and the main driving direction.

[0040] Furthermore, the cosine similarity calculation method is used to perform cosine similarity calculation on the original vector before and after projection and the main driving vector. The formula for the cosine value of the angle is as follows:

[0041] in To compare vectors, The main driving vector Represents the dot product of vectors. This represents the magnitude of the vector.

[0042] Furthermore, using the amplitude ratio calculation method (parameter: amplitude normalization reference), the ratio of the displacement vector to the magnitude of the main driving vector in the previous frame and the next frame is calculated. The amplitude ratio formula is as follows:

[0043] in It is the primary driving vector.

[0044] Furthermore, by combining the cosine of the included angle with the amplitude ratio through statistical analysis, the formulas for calculating the mean and standard deviation of the amplitude are as follows:

[0045]

[0046] in For sample amplitude, This represents the number of samples.

[0047] By comparing with preset threshold conditions (cosine of included angle < 0.85 and amplitude ratio > μ + 1.5σ), potential gait phase transition point markers are generated as structural segmentation references for subsequent local motion manifold construction.

[0048] By using a combined analysis method of the included angle cosine and amplitude ratio, the displacement vector evaluation results of the previous step are transformed into criteria for gait change consistency and acceleration change trend, thereby achieving the expected technical effect of rapid identification of dynamic load transfer patterns and anomaly point capture.

[0049] For example, for pressure center displacement data with a sampling frequency of 100Hz, the main driving vector magnitude is set to 12mm, and the magnitudes of the previous and next frames are 10mm and 16mm respectively. Using the cosine formula, the cosθ value between the previous frame and the main driving vector is 0.88, and 0.82 for the next frame. The mean amplitude μ = 13mm, the standard deviation σ = 2mm, and the amplitude ratio of the next frame is 1.33 (i.e., 16 / 12), which is higher than the threshold of 1.33 corresponding to the ratio μ + 1.5σ = 13 + 3 = 16mm. Furthermore, the condition that the cosine value of the angle is less than 0.85 is met. Therefore, the next frame is identified as a potential gait phase transition point. In this scenario, the detection module can accurately capture gait phase changes within a single step cycle, improving the sensitivity and accuracy of load migration situation analysis.

[0050] S2.3: Construct a two-dimensional motion manifold subspace using three continuous displacement vectors in the local velocity field, generate a set of orthogonal bases using the Gram-Schmidt orthogonalization method, and project the original displacement vectors onto this low-dimensional manifold space to obtain the compressed representation of the manifold projection coefficients. These coefficients reflect the typical evolution mode of the current gait segment in the local dynamic structure and suppress the interference caused by unstructured jitter.

[0051] S2.4: Perform principal component analysis (PCA) on the manifold projection coefficients, extract the eigenvector corresponding to the first principal component as the dominant motion direction, and calculate its cumulative variance contribution rate; if the contribution rate is greater than 85%, it is determined that the current three-frame window is in the stable load transfer stage and can be used to establish a reliable gait phase reference benchmark; otherwise, trigger the local resampling verification mechanism, extend the observation window to five frames and reconstruct the manifold.

[0052] Based on the two-dimensional motion manifold projection coefficient sequence generated in step S2.3, the projection coefficients corresponding to the three-frame windows are extracted using the covariance matrix decomposition method (parameters: sample dimension = 2, data-centric mean vector μ as zero base point) to achieve statistical characterization of low-dimensional dynamic patterns.

[0053] Furthermore, the covariance matrix of the manifold projection coefficients is calculated using the principal component analysis (PCA) algorithm (parameters: the eigenvalue solving method adopts QR decomposition iteration, and the convergence threshold ε=1×10^-6).

[0054] Furthermore, the first principal component direction vector is extracted using the eigenvalue decomposition method (parameter: eigenvalues ​​are sorted in descending order, and the eigenvector corresponding to the first largest eigenvalue is selected). And calculate the cumulative variance contribution rate of the first principal component, as shown in the following formula:

[0055] in, The largest eigenvalue, The th in the entire set of eigenvalues indivual.

[0056] Furthermore, the calculated The judgment criteria are constructed by comparing the value with a preset stability threshold of 0.85. If the conditions are met, then This serves as the dominant motion direction for the current three-frame window, and the window is marked as the stable load transfer phase.

[0057] Furthermore, when the judgment condition is not met, a local resampling verification mechanism is triggered (parameters: window length is extended to 5 frames, and time span is increased to 1.67 times the original gait period), and the manifold construction of S2.3 and the above PCA process are re-executed on the newly added frame data to improve the ability to distinguish instantaneous load migration anomalies.

[0058] Based on the principal component analysis results, the manifold projection coefficients from the previous step are transformed into dominant motion direction data with variance contribution rate scaling, thereby achieving automatic identification of the stable load transfer stage and reliable establishment of gait phase reference benchmark.

[0059] For example, within a three-frame window period containing the gait support phase, the two-dimensional manifold projection coefficient matrix is ​​[[0.12,0.04],[0.15,0.05],[0.11,0.03]], the sample size n=3, the mean vector is calculated as μ=[0.1267,0.04], and substituting it into the covariance formula, we get C=[[2.1667e-04, 5.0000e-05],[5.0000e-05, 0.0001]]. After eigenvalue decomposition, the largest eigenvalue λ1 = 2.5e-04 and the second largest eigenvalue λ2 = 1.0e-04 were obtained. The cumulative variance contribution rate CR = 0.7143 did not reach the threshold of 0.85, triggering local resampling to extend to 5 frames. In the extended data [[0.12,0.04],[0.15,0.05],[0.11,0.03],[0.14,0.05],[0.13,0.04]], the covariance matrix was recalculated and PCA decomposition was performed to obtain λ1 = 3.9e-04 and λ2 = 8.0e-05. The CR = 0.8298 approached the threshold and was marked as a boundary stable state. The first principal component vector in this state was applied to the subsequent stability index calculation of S2.5. In this scenario, the dominant motion direction of the resampling window significantly improved the recognition accuracy of the load transfer mode and reduced the misjudgment of phase zero points caused by instantaneous attitude adjustment under the condition of walking on a slope.

[0060] S2.5: Combine the analysis of the dominant motion direction and the radius of curvature of the pressure center trajectory to calculate the local dynamic stability index: Based on the rate of change of the tangent angle of the pressure center path and the manifold energy concentration, generate a quantitative score between 0 and 1 to indicate whether the current gait segment has sufficient biomechanical continuity to support subsequent phase zero-point locking; this score serves as the pre-confidence gating input of the 'pressure peak' identification module and participates in the conditional judgment of phase coordinate system establishment in S3.

[0061] like Figure 2 As shown, step S3 involves identifying the maximum pressure cluster in the heel region within a single-step cycle as the main pressure peak, establishing a gait phase coordinate system with its peak time as the phase zero point, and discretizing this phase axis into 64 equally divided phase slots to form a unified time-phase mapping reference. Specifically, this includes: S3.1: Acquire the raw pressure time series data of the plantar pressure sensor array within a single gait cycle. Based on the local pressure signal sequence output by the three pressure sensing units arranged in the heel region, perform sliding window mean filtering to eliminate high-frequency physiological noise and motion artifact interference, and generate a smoothed synthetic pressure waveform in the heel region as the input signal for subsequent pressure peak detection.

[0062] S3.1: Obtain the original pressure time-series matrix of the plantar pressure sensor array within a single gait cycle, extract the local pressure signal sequence based on the three pressure sensor subsets arranged in the heel region, perform sliding window mean filtering on the signal sequence to eliminate high-frequency physiological noise and motion artifact interference, and generate a smoothed synthetic pressure waveform in the heel region.

[0063] S3.2: Perform a peak detection algorithm on the smoothed heel region synthetic pressure waveform to locate the main pressure peak with the largest amplitude and duration that meets the physiological constraints of the ground contact phase (>80ms) in each gait cycle, and mark its corresponding timestamp t0 as the phase zero point of the current gait cycle, as the starting reference point for constructing the gait phase coordinate system.

[0064] S3.2: Perform a peak detection algorithm on the smoothed heel region synthetic pressure waveform to locate the pressure main peak with the largest amplitude and duration that conforms to the physiological constraints of the ground contact phase in each gait cycle. Mark the corresponding timestamp as the phase zero point of the current gait cycle and generate a starting reference point with biomechanical significance to construct the gait phase coordinate system.

[0065] S3.3: Based on the phase zero-point sequence extracted from multiple consecutive gait cycles, the average gait cycle duration T_cycle_avg is calculated. This statistic is used to model individualized gait rhythms, and a single gait cycle is mapped to a normalized phase interval [0, 2π). A nonlinear time compression function φ(t) = 2π·(t_avg) is generated to match the user's actual walking rhythm. t0) / T_cycle_current.

[0066] S3.3: Based on the phase zero-point sequence extracted from multiple consecutive gait cycles, the average gait cycle duration is calculated. This statistic is used to model individualized gait rhythms, and a single gait cycle is mapped to a normalized phase interval [0, 2π). A nonlinear time compression function matching the user's actual walking rhythm is generated as the core mapping mechanism for time-phase conversion.

[0067] S3.4: The normalized gait phase interval [0, 2π) is divided into 64 discrete phase slots, each corresponding to a phase span of π / 32 radians, forming a gait phase coding space with a fixed resolution, which is used to map multimodal sensor data collected at different times to a unified biomechanical phase coordinate system.

[0068] S3.4: The normalized gait phase interval [0, 2π) is divided into 64 discrete phase slots, each corresponding to a fixed radian span, forming a gait phase coding space with a fixed resolution. This enables the discretization of phase labeling for data collected at different times, providing a quantitative scale for the alignment analysis of cross-modal signals under a unified phase system.

[0069] S3.5: Output a set of phase mapping parameters containing the phase zero point, the current gait cycle length, and the boundary information of the 64 equally divided phase slots. This serves as a common reference for the conversion of the acceleration energy envelope, gyroscope angular velocity envelope, and the extreme point timestamp of the center of gravity shift rate of change into phase coordinates in subsequent steps, ensuring that the relative positions of multi-source signals are compared and the consistency analysis is performed within the same biomechanical rhythm framework.

[0070] S3.5: Output a phase mapping parameter set containing the phase zero point t0, the current gait cycle length T_cycle_current, and the boundary information of the 64 equally divided phase slots. This parameter set is used in subsequent steps to batch convert the timestamps of the acceleration energy envelope, gyroscope angular velocity envelope, and extreme points of the center of gravity shift rate of change into the corresponding phase coordinates φ_acc, φ_gyro, and φ_com, thereby achieving the alignment and comparison of cross-modal signals in the gait phase space.

[0071] like Figure 3 As shown, step S4 involves mapping the triaxial acceleration energy envelope, the gyroscope angular velocity absolute value envelope, and the timestamps of the extreme points of the center of gravity shift rate of change to the gait phase coordinate system. This obtains the dominant phase position of each signal on the phase axis and calculates its offset relative to the pressure peak phase. Specifically, this includes: S4.1: Obtain the raw acceleration time-domain signal output by the triaxial accelerometer. Based on the local motion manifold subspace and the radius of curvature of the pressure center trajectory constructed in the previous steps, perform envelope demodulation processing on the acceleration vector synthesis signal. Use the Hilbert transform algorithm to extract the acceleration energy envelope to obtain a continuous scalar sequence characterizing the evolution of motion impact intensity over time, and generate an acceleration energy evolution field as the input condition for dominant phase extraction.

[0072] S4.1: Obtain the raw acceleration time-domain signal output by the triaxial accelerometer. Based on the local motion manifold subspace and the radius of curvature of the pressure center trajectory constructed in the previous steps, perform envelope demodulation processing on the acceleration vector synthesis signal. Use the Hilbert transform algorithm to extract the acceleration energy envelope to obtain a continuous scalar sequence characterizing the motion impact intensity over time, and generate an acceleration energy evolution field as the input condition for the dominant phase extraction.

[0073] The raw acceleration time-domain signal output by the triaxial accelerometer is acquired. The sensor sampling buffer mechanism (sampling rate ≥ 100Hz) is used to achieve synchronous capture of the acceleration components of the three axes X, Y and Z, forming a structured acceleration vector sequence A(t) = [a_x(t), a_y(t), a_z(t)].

[0074] Based on the local motion manifold subspace constructed in the previous steps, the acceleration vector sequence is projected onto the dominant basis system of this subspace to filter out non-gait-related high-frequency noise components and retain the principal dynamic components consistent with the curvature radius change pattern of the pressure center trajectory, thereby generating the manifold-constrained acceleration signal A_m(t).

[0075] Furthermore, the manifold constraint acceleration components along the three axes are synthesized into a single scalar norm signal using a vector synthesis algorithm (parameters: square root operation), thereby achieving a time-series characterization of the overall motion impact intensity. The synthesis formula is as follows: in , and These are the acceleration components along the three axes.

[0076] Furthermore, the Hilbert transform algorithm (parameter: analytic signal generation mode) is used to perform envelope demodulation processing on the acceleration scalar norm signal a(t), constructing its corresponding analytic signal Z(t) = a(t) + j·H[a(t)], where H[·] is the Hilbert transform operator. The magnitude formula of the analytic signal is as follows: in It is the acceleration energy envelope sequence.

[0077] The acceleration energy envelope sequence is continuously resampled in the time domain using a time-normalized interpolation algorithm (parameter: gait cycle length T_step) to generate a high-resolution acceleration energy evolution field that matches the gait rhythm, ensuring time alignment with the radius of curvature information of the pressure center trajectory.

[0078] The above-mentioned algorithm or processing method transforms the acceleration vector synthesis signal into an acceleration energy evolution field with continuous time evolution characteristics, realizing a distortion-free transition from input conditions to dominant phase extraction, and preserving the motion impact characteristics related to the spatiotemporal topology of plantar pressure.

[0079] For example, in the operating scenario of an embedded anti-fall monitoring shoe, the sampling rate of the triaxial accelerometer is set to 200Hz, and the dominant basis of the local motion manifold subspace is generated by training the load transfer curvature radius from the forefoot to the heel. When synthesizing the acceleration scalar norm signal, the acceleration components of each axis are a_x(t)=0.3g, a_y(t)=0.1g, and a_z(t)=1.2g, respectively. After square root operation, a(t)=1.239g is obtained. The modulus of the analytic signal is generated by Hilbert transform, and the peak value of the acceleration energy envelope E_acc(t_peak)=1.305g is obtained. Using the gait period length T_step=1.2s, the energy envelope field is resampled to a 64-point phase resolution through time normalization interpolation to achieve precise alignment with the gait phase coordinate system established in S3. In this scenario, the energy evolution field exhibits stable peak characteristics near the phase zero point, providing a high-confidence basis for the subsequent calculation of the acceleration-dominated phase angle in S4.2.

[0080] S4.1: Perform envelope demodulation processing on the triaxial acceleration signal, and use the Hilbert transform algorithm to extract the acceleration energy envelope E_acc(t) to obtain a continuous scalar sequence characterizing the evolution of motion impact intensity over time, which serves as the dominant feature input of the acceleration signal.

[0081] S4.2: Based on the acceleration energy envelope E_acc(t) generated in S4.1, locate the timestamp t_acc_max corresponding to its global maximum value within a single-step cycle, and according to the gait phase coordinate system φ∈[0,2π) established in the previous step S3, use the linear mapping relationship φ_acc = 2π·(t_acc_max) The dominant phase position φ_acc corresponding to the extreme point is calculated using t0) / T_step, where t0 is the moment of the main pressure peak and T_step is the current step size estimate.

[0082] S4.2: Based on the global maximum point in the acceleration energy evolution field, locate its corresponding timestamp, combine it with the phase mapping parameter set output by the previous steps (including the pressure peak time and the current step size estimate), and use the linear normalization mapping mechanism to convert the timestamp into the corresponding gait phase coordinates, generate the acceleration dominant phase angle, and realize the projection transformation of the time domain extreme point to the biomechanical phase space.

[0083] S4.2: Based on the global maximum point in the acceleration energy evolution field, locate its corresponding timestamp t_acc_max. Combined with the phase mapping parameter set output from the previous step S3 (including the pressure peak time t0 and the current step size estimate T_step), use a linear normalization mapping mechanism to convert t_acc_max into the corresponding gait phase coordinates φ_acc = 2π·(t_acc_max) t0) / T_step generates the acceleration-dominant phase angle, realizing the projection transformation of the time domain extremum point to the biomechanical phase space.

[0084] S4.3: Obtain the time-domain signal of angular velocity output from the Z-axis of the gyroscope. Based on the established zero point of the pressure peak phase and the nonlinear compression function of the gait period, perform absolute value conversion and sliding window root mean square filtering on the angular velocity signal to generate the absolute value envelope of the angular velocity. Extract the timestamps corresponding to its extreme points and convert them into the dominant phase angle of the gyroscope according to the same phase mapping rule to generate the dominant phase parameter of rotational dynamics, which is used to characterize the relative temporal position of the attitude adjustment behavior in the gait period.

[0085] The angular velocity time-domain signal output from the Z-axis of the gyroscope is obtained as the input object of this sub-step. The input conditions include the sensor's sampling rate, range, and gait phase mapping parameter set to ensure that the data information has a corresponding relationship between the time domain and the phase domain.

[0086] The absolute value processing method is used to convert the original angular velocity signal. Each sampling point is mapped to its absolute value sequence to eliminate the influence of rotation direction on amplitude analysis and enhance the symmetry characteristics of attitude adjustment intensity.

[0087] Furthermore, a sliding window root mean square filtering algorithm (window length N_samp, sliding step size Δ_samp) is used to smooth the angular velocity absolute value sequence and generate the angular velocity absolute value envelope. The calculation formula is as follows: ,in For the index of sampling points within the window, This represents the number of samples in the window.

[0088] Furthermore, a peak detection algorithm is used to analyze the absolute value envelope of the angular velocity. Extract extreme points and locate the main peak timestamp. This ensures that the peak value is characteristic of a posture adjustment event within the physiological gait cycle.

[0089] Furthermore, the phase mapping parameter set output in the preceding step S3.5 (including the pressure peak time) is used. With step size estimate ), through the linear normalization function Mapped to gait phase coordinates The formula is: This enables the projection of time-domain event points onto the biomechanical phase space.

[0090] Through the aforementioned signal processing chain, the gyroscope Z-axis angular velocity signal is converted into a rotational dynamics dominant phase parameter with gait phase coordinates, which is used to reflect the relative temporal position of attitude adjustment behavior in the gait cycle, thereby achieving the technical effect of dynamic alignment of multimodal signals.

[0091] For example, in a gyroscope Z-axis signal with a sampling rate of 200Hz and a range of ±500° / s, the sliding window length N_samp is set to 20 sampling points and the step size Δ_samp to 5 sampling points. The absolute value of the input angular velocity sequence is used to calculate the root mean square value within the sliding window to generate a smooth envelope. The main peak timestamp was detected within one gait cycle (T_step = 1.2 seconds). ; in, For t_com_peak, For the phase zero-point timestamp, This represents the duration of the gait cycle.

[0092] This processing method transforms the position of the extreme point of the rate of change in the time domain into the dominant phase angle of the center of gravity shift in a biomechanical sense, thereby achieving unified phase labeling for cross-modal signals.

[0093] Through the above algorithm chain, the center of gravity offset signal obtained in the previous step is transformed into dominant phase angle data with time-phase mapping characteristics, realizing accurate phase positioning and rhythm alignment of key events in human balance regulation.

[0094] For example, in a certain implementation scenario, the sampling rate of the attitude sensor is set to 100Hz, and the process noise covariance matrix Q of the Kalman filter parameter is taken as... The measurement noise covariance matrix R is taken as... The peak value of the rate of change of the smoothed center of gravity offset signal within a certain gait cycle is obtained as follows: m / s, corresponding to the timestamp is seconds. When the gait period T_step is seconds, phase zero point t0 is The mapping formula is calculated as follows:

[0095] Calculations show that φ_com is approximately The phase angle is labeled in a unified gait phase coordinate system and stored as the dominant phase feature of the center of gravity shift in the current gait cycle. It is then used as an initial dependent variable in the subsequent phase shift and confidence assessment stages to participate in the fusion decision, thereby significantly improving the localization accuracy of balance events and the robustness of system response in complex scenarios.

[0096] S4.4: Perform a first-order difference operation on the centroid offset ΔCoM(t) estimated jointly by attitude angle integral and Kalman filtering to generate the rate of change signal dΔCoM / dt, and detect its local extremum set {t_com_extremum}; select the timestamp t_com_max corresponding to the extremum point with the largest amplitude, according to φ_com = 2π·(t_com_max) t0) / T_step maps it to the gait phase coordinate system to obtain the dominant phase φ_com of the center of gravity offset.

[0097] S4.4: Obtain the centroid offset ΔCoM(t) jointly estimated by attitude angle integral and Kalman filtering. Based on the dynamic stability index and local motion manifold energy concentration defined in the previous step, perform a first-order difference operation on this signal to generate the rate of change signal dΔCoM / dt. Detect the timestamp t_com_max corresponding to the extreme point of its maximum amplitude, and then use the phase mapping function φ_com = 2π·(t_com_max) t0) / T_step converts it into the dominant phase angle of the center of gravity offset, generating phase annotation results for key events in human balance regulation.

[0098] S4.5: Based on the dominant phases φ_acc, φ_gyro, and φ_com obtained in S4.2, S4.3, and S4.4 respectively, and combined with the zero point φ0 (i.e., φ=0) of the pressure peak phase defined in S3, calculate the absolute offset of each signal phase relative to φ0, δφ_acc = |φ_acc φ0|、δφ_gyro = |φ_gyro φ0|、δφ_com = |φ_com φ0| outputs a phase deviation vector with physiological rhythm reference significance, which is used for subsequent phase consistency assessment and confidence penalty mechanism construction.

[0099] S4.5: Based on the acceleration-dominant phase angle, gyroscope-dominant phase angle, and center of gravity offset-dominant phase angle, combined with the pressure peak phase zero point, calculate the absolute phase offset of the three relative to the phase zero point, and output a phase deviation vector set with physiological rhythm reference significance. This serves as the core input for evaluating the temporal consistency of multimodal signals and forms the initial dependent variable set for the subsequent confidence penalty mechanism construction.

[0100] Based on the acceleration-dominant phase angle, gyroscope-dominant phase angle, and center of gravity offset-dominant phase angle output from steps S4.2, S4.3, and S4.4, and combined with the zero point of the pressure peak phase φ0=0, the absolute phase offset calculation method is used to achieve quantitative characterization of the phase difference of multi-mode signals.

[0101] Furthermore, by taking the absolute value of the difference, the dominant acceleration phase angle φ_acc is compared with the zero point of the pressure peak phase φ0 to obtain the absolute phase offset of the acceleration signal:

[0102] Furthermore, the absolute offset of the gyroscope's dominant phase angle φ_gyro is calculated using the same method to obtain the offset of the gyroscope signal:

[0103] Furthermore, an absolute offset calculation is performed on the dominant phase angle φ_com of the center of gravity offset to generate the offset amount of the center of gravity offset signal:

[0104] Furthermore, the absolute phase shifts of the three types of signals are combined into a phase deviation vector group, which is represented as:

[0105] Furthermore, a phase deviation set with physiological rhythm reference significance is formed through the output of the vector group. This set is used as the core input for evaluating the temporal consistency of multimodal signals, providing an initial set of dependent variables for subsequent offset penalty factor calculation and stability weight allocation.

[0106] By using the above calculation method, a unified quantitative relationship is established between the dominant phase of different sensors and the plantar pressure reference phase, realizing the standardized comparison of multimodal data in the time-phase domain, which helps to suppress false alarms caused by gait rhythm differences.

[0107] For example, in a fall risk monitoring scenario for the elderly with a gait cycle of T_step = 1.2 seconds, the zero point of the pressure peak phase φ0 is locked by the heel landing event, the acceleration-dominant phase angle φ_acc = 0.52π radians, the gyroscope-dominant phase angle φ_gyro = 0.61π radians, and the center of gravity offset-dominant phase angle φ_com = 0.47π radians. Substituting these values ​​into the above formula, we calculate δφ_acc = |0.52π 0|=0.52π radians, δφ_gyro=|0.61π 0|=0.61π radians, δφ_com=|0.47π 0|=0.47π radians, and these three values ​​form a phase deviation vector set {0.52π, 0.61π, 0.47π}. The system uses this vector set to determine if there is a deviation between the acceleration and the gyroscope signal close to the physiological tolerance threshold of π / 4. It then generates a deviation penalty factor, which is increased in the stability calculation to reduce the probability of false triggering. The above process, with a computation latency of no more than 5ms on the ARM Cortex-M4 platform, achieves real-time processing of multimodal signal consistency discrimination, significantly improving the reliability of early warning in complex gait scenarios.

[0108] S4.5: Based on the acceleration-dominant phase angle φ_acc, the gyroscope-dominant phase angle φ_gyro, and the center of gravity offset-dominant phase angle φ_com, combined with the pressure peak phase zero point φ0=0, calculate the absolute phase offset δφ_acc = |φ_acc| relative to φ0 for each of the three. φ0|、δφ_gyro = |φ_gyro φ0|、δφ_com = |φ_com φ0| outputs a set of phase deviation vectors with physiological rhythm reference significance, which serves as the core input for evaluating the temporal consistency of multimodal signals and forms the initial set of dependent variables for the subsequent construction of the confidence penalty mechanism.

[0109] Step S5: Based on the phase shift of each signal and its standard deviation over five consecutive gait cycles, assess phase consistency and evolutionary stability, and generate an initial confidence score including a shift penalty factor and stability weights. Specifically, this includes: S5.1: Obtain the acceleration-dominant phase φ_acc, gyroscope-dominant phase φ_gyro, and center of gravity shift rate-dominant phase φ_com output from step S4, and calculate the phase shift δφ_acc=|φ_acc| of each signal within the current gait cycle based on the zero point of the pressure peak phase φ0=0. φ0|、δφ_gyro=|φ_gyro φ0|、δφ_com=|φ_com φ0| is used to quantify the temporal misalignment of multi-source signals relative to the spatiotemporal topological reference of plantar pressure, and to generate a preliminary offset parameter set {δφ_acc, δφ_gyro, δφ_com}.

[0110] Acceleration-dominant phase based on the output of step S4 gyroscope dominant phase and the dominant phase of the rate of change of the center of gravity offset At the gait phase zero point Under these conditions, the absolute offset calculation method is used to quantify the difference between the three types of signals and the main pressure peak phase.

[0111] Applying the absolute value calculation formula to the dominant phase signal of acceleration This allows for the calculation of the phase misalignment degree within the current gait cycle, and the resulting scalar value. As one of the offset parameters.

[0112] Furthermore, using the same absolute value formula Process the gyroscope's dominant phase signal, eliminate sign effects, extract the amplitude difference between it and the phase zero, and generate the offset. Used for multimodal phase comparison analysis.

[0113] Furthermore, the absolute value offset is calculated for the dominant phase signal of the rate of change of the center of gravity offset. This yields a temporal deviation measure within the current gait cycle, forming the offset. .

[0114] By using three types of absolute offset calculation formulas, the temporal misalignment of acceleration, gyroscope, and center of gravity offset signals relative to the main pressure peak is transformed into a set of offset parameters with unified dimensions. This provides a quantitative basis for subsequent stability assessment and initial confidence value calculation, and enables the physical comparability of the offset characteristics of multi-source signals.

[0115] For example, in the scenario of a 70-year-old subject, the dominant phase of acceleration was obtained as follows: radians, gyroscope dominant phase The dominant phases are radians and centroid shift. Radius, pressure peak phase zero point set Radians. Substitute into the formula. achievable radians; substitute into the formula get radians; substitute into the formula get Radius. Application results show that the phase misalignment of the three types of signals is significantly different within this gait cycle, providing a high-resolution quantitative basis for adjusting the confidence score based on stability weights in the next step. Among them, the larger offset of the acceleration signal may indicate the presence of gait impact events.

[0116] S5.2: Based on the phase offset sequences of each signal stored in the historical five-gait cycle buffer queue, the continuous five-cycle δφ sequence corresponding to the current signal is extracted using a sliding time window. The phase offset standard deviations σ_δφ_acc, σ_δφ_gyro, and σ_δφ_com are calculated to evaluate the evolution stability of each sensor signal within a short time window. A stability feature index set {σ_δφ_acc, σ_δφ_gyro, σ_δφ_com} is generated as a basis for judging whether the signal is affected by environmental interference or individual gait variations.

[0117] Based on the phase offset sequences of each signal stored in the historical five-step cycle buffer queue, a sliding time window is used to extract five consecutive cycles of data corresponding to the current signal, thus preparing the input for short-term stability analysis. Using a time window index mapping method (window length parameter: 5, step size parameter: 1), the latest five-cycle phase offsets of three types of signals—accelerometer, gyroscope, and center of gravity offset—are located sequentially, forming the δφ_acc, δφ_gyro, and δφ_com sequences within the sliding window. Furthermore, a variance calculation algorithm is used to perform mean removal within each of the three sequences, and the sum of squares of the phase offset sample set is calculated, then divided by the sample size minus one to calculate the standard deviation of the phase offset. The standard deviation calculation formula is as follows:

[0118] in, This represents the phase shift of the acceleration signal within a single gait cycle. This is the arithmetic mean of the sequence. This represents the number of samples within the window.

[0119] Using the same formula structure, replace the input signal types with gyroscope and center of gravity offset data to generate... and This allows for the quantitative characterization of the evolutionary stability of each channel. Furthermore, the standard deviation results of the three types of signals are combined into a stability feature index set {σ_δφ_acc, σ_δφ_gyro, σ_δφ_com}, which serves as the core input set for subsequent phase consistency and stability weight determination. Through structured calculation of the variance function, the original phase offset sequence is transformed into a stability index with comparability and quantitative standards, enabling precise characterization of the degree to which each sensor signal is affected by environmental interference or individual gait variations.

[0120] For example, in a test embodiment involving a mixed scenario of walking up stairs and walking on flat ground, the δφ sequence of the gait control chip's cached acceleration channel is {0.41π, 0.38π, 0.40π, 0.42π, 0.39π}, the gyroscope channel is {0.35π, 0.37π, 0.36π, 0.38π, 0.36π}, and the center of gravity offset channel is {0.29π, 0.34π, 0.33π, 0.31π, 0.30π}. Substituting the acceleration sequence into the standard deviation formula, we obtain... gyroscope Center of gravity shift In the stability weight mapping table, acceleration and gyroscope both belong to the high stability range (σ < π / 12), and are assigned a weight of 0.9, while the center of gravity offset is in the medium stability range (π / 12 ≤ σ < π / 6), and is assigned a weight of 0.6. The output stability feature index set is directly used for the stability level classification and offset penalty logic in the next sub-step, providing a quantitative basis for the dynamic adjustment of the confidence of multimodal warning signals during stair walking. In this scenario, the system significantly improves the adaptability and robustness of the alarm strategy under complex gait conditions.

[0121] S5.3: Based on the comparison between the phase offset δφ and the preset physiological tolerance threshold π / 4, perform phase consistency judgment: if δφ>π / 4 and lasts for more than 2 phase slots (corresponding time>30ms), then trigger the phase drift flag and generate an offset penalty factor P_drift∈{0, 0.2}, where P_drift=0.2 indicates that there is a significant temporal mismatch, and the initial confidence needs to be reduced to suppress the risk of false triggering caused by non-physiological mutations.

[0122] S5.4: Based on the interval to which the phase offset standard deviation σ_δφ belongs, the stability level is divided: if σ_δφ < π / 12, it is determined to be a high stability state, and a high stability weight w_stab = 0.9 is generated; if π / 12 ≤ σ_δφ < π / 6, it is determined to be a medium stability state, and a medium stability weight w_stab = 0.6 is generated; if σ_δφ ≥ π / 6, it is determined to be a low stability state, and a low stability weight w_stab = 0.3 is generated. A stability weight mapping table is constructed to differentiate the signal reliability under different operating conditions.

[0123] Based on the phase offset standard deviation index set {σ_δφ_acc, σ_δφ_gyro, σ_δφ_com} output in step S5.3, the stability level is determined by using an interval division method (thresholds: π / 12, π / 6).

[0124] A comparison operation method is used to compare the phase offset standard deviation σ_δφ of each signal with a high stability threshold. When a comparison is performed, The state is determined to be highly stable, and stability weights are generated through assignment operations. This is to reflect the consistency of signal evolution within a short time window.

[0125] Furthermore, using the same comparison operation method, With moderate stability threshold range to Perform interval matching, when ≤ σ_δφ < π 6 When the state is determined to be moderately stable, stability weights are generated. It is used to reflect the availability of a signal under certain variations.

[0126] Furthermore, using the interval discrimination method, when When this occurs, the signal is identified as a low-stability state and assigned a stability weight. This reduces the trust level in the signal during subsequent alarm decisions.

[0127] By differentiating the stability levels, the σ_δφ value of each signal is mapped to the corresponding w_stab value, forming a stability weight mapping table. This table records the signal channel type and stability weight in key-value pairs, providing direct weight input for the initial confidence score formula in step S5.5, and realizing differentiated adjustment of the signal confidence level under different operating conditions.

[0128] For example, in a scenario where an elderly user is descending stairs, the standard deviation of the acceleration phase shift measured over five consecutive gait cycles is: Radius, standard deviation of gyroscope phase offset is Radius, centroid offset, phase offset standard deviation is Radians. The acceleration signal σ_δφ_acc = 0.21 radians is compared with a threshold. The signal σ_δφ_gyro = 0.28 radians, compared to the interval [0.262, 0.524), meets the high stability condition, and is mapped to w_stab_acc = 0.9. The gyroscope signal σ_δφ_gyro = 0.28 radians, compared to the interval [0.262, 0.524), meets the medium stability condition, and is mapped to w_stab_gyro = 0.6. The centroid offset signal σ_δφ_com = 0.35 radians also falls within this interval, and is mapped to w_stab_com = 0.6. After mapping, a stability weight table {acc: 0.9, gyro: 0.6, com: 0.6} is generated. In the subsequent calculation of the initial confidence level in S5.5, these weights directly participate in the formula calculation, significantly increasing the contribution of the high-stability channel to the alarm decision, while maintaining a moderate contribution of the medium-stability channel, thereby improving the robustness of the overall alarm strategy in stairwell scenarios.

[0129] S5.5: Combine the offset penalty factor P_drift and the stability weight w_stab to generate the initial confidence score for each signal Conf_initial = (1 P_drift) × w_stab, outputting the initial confidence score group {Conf_initial_acc, Conf_initial_gyro, Conf_initial_com} for the three channels of acceleration, gyroscope and center of gravity offset, which serves as the input basis for the final confidence score fusion and hierarchical alarm decision in the next step.

[0130] Based on the offset penalty factor P_drift and stability weight w_stab output from steps S5.3 and S5.4, initial confidence scores are calculated for the three types of signals: acceleration, gyroscope, and center of gravity shift, respectively, to establish a unified quantization scale for multi-source signal fusion determination.

[0131] A numerical substitution method is used, inputting the offset penalty factor P_drift and the stability weight w_stab for each channel into the initial confidence score formula to achieve joint quantification of phase drift deduction and stability weight. The formula is as follows:

[0132] Furthermore, by substituting P_drift_acc and w_stab_acc into the acceleration signal channel using this formula, an initial confidence score of acceleration, Conf_initial_acc, is generated, ensuring that the product of the deduction term and the weight term reflects the dynamic consistency and topological stability within the current gait cycle.

[0133] Furthermore, by substituting P_drift_gyro and w_stab_gyro into the gyroscope signal channel using this formula, the initial confidence score Conf_initial_gyro of the gyroscope is generated, so that the rotational dynamics information, after deducting the influence of abnormal drift, yields a stability-weighted result.

[0134] Furthermore, by substituting P_drift_com and w_stab_com into the center of gravity offset signal channel using this formula, an initial confidence score Conf_initial_com for center of gravity offset is generated to quantify the temporal reliability of human balance regulation behavior.

[0135] Through the above mathematical operations, the offset penalty and stability weight are effectively fused into a confidence scalar group {Conf_initial_acc, Conf_initial_gyro, Conf_initial_com} in the interval [0,1], realizing the reliability quantification of different sensor signals under a unified scale, and providing a highly consistent input for subsequent final confidence score fusion and hierarchical alarm decision-making.

[0136] For example, in a real-world stair-descent scenario, the average acceleration phase shift δφ_acc over five consecutive gait cycles is 0.72 radians, and the standard deviation σ_δφ_acc is 0.22 radians. This exceeds the physiological tolerance threshold of π / 4, and the shift persists for more than two phase slots, resulting in P_drift_acc = 0.2. The stability level is determined to be moderately stable, with w_stab_acc = 0.6. Substituting into the formula: The calculated Conf_initial_acc = 0.48. In the same scenario, the gyroscope channel's δφ_gyro is 0.5 radians and σ_δφ_gyro is 0.10 radians, which is considered within limits. P_drift_gyro = 0, indicating high stability. w_stab_gyro = 0.9, substituting these values ​​into the formula yields Conf_initial_gyro = 0.9. The center of gravity offset channel's δφ_com is 0.85 radians and σ_δφ_com is 0.20 radians, indicating an offset exceeding limits. P_drift_com = 0.2, indicating medium stability. w_stab_com = 0.6, substituting these values ​​into the formula yields Conf_initial_com = 0.48. These results show that in this stair-down scenario, the initial confidence levels of the acceleration and center of gravity offset channels are relatively low, while the gyroscope channel maintains a high confidence level. This provides differentiated input conditions for the subsequent decision tree to execute precise hierarchical alarms under complex conditions.

[0137] Step S6: Based on the initial confidence score, calculate the final confidence scores for acceleration, gyroscope, and center of gravity shift signals, and input the three scores into a preset lightweight decision tree model to output the corresponding alarm level determination result. Specifically, this includes: S6.1: Based on the phase offsets δφ_acc, δφ_gyro, and δφ_com of each sensor signal output from S5, and their phase offset standard deviations σ_δφ over five consecutive gait cycles, an offset penalty calculation is performed—an indicator function I (δφ>π / 4) is applied to each signal to determine whether there is a significant phase drift. If so, a fixed deduction of 0.2 is introduced. At the same time, the stability weight w_stab is mapped according to the interval to which σ_δφ belongs (high stability: σ_δφ<π / 12→0.9; medium stability: π / 12≤σ_δφ<π / 6→0.6; low stability: σ_δφ≥π / 6→0.3) to quantify the topological stability of the signal evolution process and generate an intermediate scoring variable containing penalty and weight factors.

[0138] Based on the phase offset set {δφ_acc, δφ_gyro, δφ_com} output in step S5 and the corresponding five-step periodic phase offset standard deviation set {σ_δφ_acc, σ_δφ_gyro, σ_δφ_com}, an interval discrimination method (parameter: phase tolerance threshold π / 4) is used to perform timing consistency detection on each signal to determine whether there is significant phase drift.

[0139] Furthermore, by defining an indicator function I (δφ>π / 4), a fixed deduction coefficient of 0.2 is applied to signals that meet the significant phase out-of-bounds condition, and a corresponding penalty factor variable is generated for subsequent confidence score deduction calculation.

[0140] Furthermore, based on the stability interval division criteria (high stability: σ_δφ<π / 12; medium stability: π / 12≤σ_δφ<π / 6; low stability: σ_δφ≥π / 6), the σ_δφ of each signal is mapped to a stability weight w_stab, where high stability corresponds to a weight of 0.9, medium stability corresponds to a weight of 0.6, and low stability corresponds to a weight of 0.3, in order to quantify the topological stability of the signal evolution process.

[0141] Furthermore, an intermediate scoring variable set {P_drift_acc, w_stab_acc; P_drift_gyro, w_stab_gyro; P_drift_com, w_stab_com} is constructed, which includes the penalty factor P_drift and the stability weight w_stab, to provide input for the final confidence comprehensive calculation in the next sub-step.

[0142] Through the aforementioned phase offset penalty calculation and stability weight mapping, the original offset parameters and standard deviation features are transformed into intermediate scoring data with dual representations of temporal consistency and dynamic stability, thereby realizing pre-quantization modeling of signal credibility.

[0143] For example, in a user's gait monitoring scenario, the phase shift δφ_acc of the acceleration signal is... In radians, the standard deviation σ_δφ_acc over five consecutive periods is radians; the δφ_gyro of the gyroscope signal is radians, σ_δφ_gyro is radians; the centroid offset signal δφ_com is radians, σ_δφ_com is Radius. A discriminant formula is used. The acceleration and gyroscope signals satisfy the significant drift condition and generate respectively. The P_drift_acc and P_drift_gyro, the centroid offset signal does not satisfy the drift condition P_drift_com is According to the interval mapping, σ_δφ_acc ( The value falls within the moderate stability range, so w_stab_acc is assigned a value. ;σ_δφ_gyro( If the value falls within the low stability range, then w_stab_gyro is assigned a value. ;σ_δφ_com ( The value of w_stab_com falls within the high stability range. The final set of intermediate scoring variables is: acceleration channel {P_drift_acc= , w_stab_acc= }; Gyroscope channel {P_drift_gyro= , w_stab_gyro= }; Center of gravity offset channel {P_drift_com= , w_stab_com= The dataset was substituted into the formula in S6.2 to calculate the corresponding final confidence score, which significantly improved the accuracy and robustness of the warning signal in complex gait scenarios.

[0144] S6.2: Calculate the final confidence score for the acceleration signal—using the intermediate scoring variables generated in S6.1, substitute them into the formula Conf_acc = max(0, 1). 0.2·I(δφ_acc>π / 4) 0.1·I(σ_δφ_acc≥π / 6)) ×w_stab_acc, where the first term represents the phase out-of-bounds penalty and the second term is the stability degradation penalty. After multiplying by the stability weight, the output is the final confidence score Conf_acc, which is in the range of [0,1] and serves as one of the input dimensions of the lightweight decision tree.

[0145] S6.3: Independent confidence modeling is performed on the gyroscope and center of gravity offset signals based on the same mechanism—applying the formula Conf_gyro = max(0, 1) to the gyroscope signal. 0.2·I(δφ_gyro>π / 4) 0.1·I(σ_δφ_gyro≥π / 6)) × w_stab_gyro, applying the formula Conf_com = max(0, 1) to the rate of change of the center of gravity offset signal. 0.2·I(δφ_com>π / 4) 0.1·I(σ_δφ_com≥π / 6)) × w_stab_com, ensuring that all three types of heterogeneous signals are converted into confidence scalars with a unified semantic scale, forming a structured input vector [Conf_acc, Conf_gyro, Conf_com].

[0146] Based on the gyroscope signal phase offset generated in step S6.1 With phase offset standard deviation A confidence calculation method combining phase out-of-bounds penalty and stability degradation penalty (parameters: phase out-of-bounds threshold π / 4, stability degradation threshold π / 6) is adopted to quantify the confidence of the gyroscope signal in the current gait cycle.

[0147] Furthermore, operations are performed using indicator functions. The system determines whether there is significant phase drift in the gyroscope signal and introduces a fixed deduction term of 0.2 when the determination is successful, thereby quantifying the penalty for non-physiological phase drift.

[0148] Furthermore, operations are performed using indicator functions. The system identifies whether the gyroscope signal exhibits stability degradation within a continuous five-step cycle, and introduces a fixed deduction term of 0.1 when the determination is successful, thereby suppressing and regulating short-term instability.

[0149] Furthermore, the deduction results of phase boundary violation penalty and stability degradation penalty are substituted into the final confidence formula. and with stability weights Multiply the results to generate a final confidence score for the gyroscope signal, which is limited to the range [0,1].

[0150] Phase offset based on the centroid shift rate of change signal generated in step S6.1 With phase offset standard deviation The same penalty-weight combination mechanism is used to quantify the confidence of the centroid offset signal.

[0151] Furthermore, operations are performed using indicator functions. Identify whether the centroid offset signal has phase drift, and deduct 0.2 from the initial confidence value if it is true.

[0152] Furthermore, operations are performed using indicator functions. Determine whether the centroid offset signal exhibits stability degradation within a short time window, and deduct 0.1 confidence points if the degradation is confirmed.

[0153] Furthermore, substituting the above deduction results into the formula... and with stability weights Multiplying these results in a final confidence score for the centroid offset signal, which is limited to the range [0,1].

[0154] By using the same independent confidence modeling mechanism described above, the final scores from acceleration, gyroscope, and center of gravity offset signals are combined to form a structured input vector. As input to a lightweight decision tree model, it enables unified semantic scale conversion of multimodal signals.

[0155] For example, in a real-world gait monitoring scenario, the phase shift of the gyroscope signal... Radius, less than the threshold π / 4, no phase drift deduction; phase offset standard deviation Radius, greater than π / 6, stability weight Final gyroscope confidence Phase offset of the center of gravity offset signal Radius, less than π / 4, no phase drift deduction; phase offset standard deviation radians, stability weight The final center of gravity shift confidence level The three-channel input vector is [ , , When input into a lightweight decision tree model, the ability of multimodal alarm strategies to suppress low-stability single channels can be significantly improved.

[0156] S6.4: Input the structured input vector generated in S6.3 into the preset lightweight decision tree model. This model is built based on biomechanical prior knowledge and contains only 3 layers of non-leaf nodes and 8 leaf nodes. It adopts hard threshold rules for partitioning: when Conf_acc≥0.85 and Conf_gyro≥0.75 and Conf_com≥0.7, the 'risk warning' command is output; when any two Conf values ​​are <0.6 and the third Conf value is ≥0.8, the 'start re-verification' command is output; when all three Conf values ​​are <0.5, the 'suppress alarm + record phase loss' command is output, realizing multimodal fusion decision-making with no training required and strong interpretability.

[0157] Based on the structured input vector [Conf_acc, Conf_gyro, Conf_com] generated by S6.3, a lightweight decision tree initialization loading method is used (parameters: number of node layers = 3, number of leaf nodes = 8, rule type = hard threshold) to realize tree branching and condition filtering of multimodal confidence input.

[0158] Furthermore, by constructing biomechanical prior rules (parameters: risk warning threshold combination = [0.85, 0.75, 0.7], re-verification trigger combination = [<0.6, <0.6, ≥0.8], suppression alarm combination = [<0.5, <0.5, <0.5]), the node condition judgment function is initialized and loaded, and the rule node mapping table is obtained as the basis for engine control.

[0159] Furthermore, an input vector threshold matching algorithm is used (parameters: multi-dimensional comparison mode = simultaneous satisfaction, global comparison precision = 10). -6 This function performs matching and discrimination between each element Conf_acc, Conf_gyro, and Conf_com in the structured input and the rule node, and generates a set of conditional trigger signals {Trigger_Risk, Trigger_Reverify, Trigger_Inhibit}.

[0160] Furthermore, by using a branch selection and state output method (parameter: priority sequence = Trigger_Risk>Trigger_Reverify>Trigger_Inhibit), priority parsing of the conditional trigger signal set is achieved, and corresponding alarm level judgment results are generated, forming a multimodal fusion decision output that requires no training and has strong interpretability.

[0161] , , Output risk warning instructions under certain conditions.

[0162] Any two items and the third item Output the start re-verification command under the given conditions.

[0163] All three conditions are met. Under certain conditions, output an alarm suppression command and record the phase lockout event.

[0164] By using rule matching and branch output algorithms, the confidence results from the previous step are transformed into standardized alarm level data, enabling the system to respond accurately under different confidence combinations.

[0165] For example, during the gait cycle of a real user, the confidence level of the acceleration signal Conf_acc is... The gyroscope signal confidence level Conf_gyro is The confidence level of the center of gravity offset signal is Conf_com Substitute the input vector into the decision tree and use the input vector threshold matching algorithm to perform a judgment, satisfying... , , The risk warning rules are combined to generate a Trigger_Risk signal and output a risk warning command directly according to the priority sequence. This command will be appended with vibration frequency and LED color parameters in subsequent S6.5, triggering the warning module to execute a physical prompt. Field test results show that the response delay under this high-confidence combination is less than 8ms, and no false alarms occurred, verifying the real-time performance and accuracy of this sub-step under multimodal fusion conditions.

[0166] S6.5: The alarm level commands output by S6.4 are formatted and status-marked, generating standardized control messages and transmitting them to the early warning module execution unit. For the 'Risk Warning' command, vibration frequency parameters (50Hz) and LED color codes (yellow light) are added. For the 'Start Re-verification' command, a delay window duration (150ms) and a high-frequency sampling trigger flag are added. For the 'Suppress Alarm' command, the event log label 'Phase Loss' and a snapshot of the current operating condition context are added to ensure that subsequent actions have complete control basis.

[0167] Following step S6, the following is also included: S7: Determine whether the alarm level determination result meets the risk warning conditions. If it does, trigger the vibration motor and LED yellow light prompt. If it is determined to be a low confidence combination, start the multimodal re-verification process or suppress the alarm and record the phase loss event. S8: Aggregates historical phase loss events, statistically analyzes the operating characteristics of high-frequency loss scenarios, and dynamically updates the individualized phase offset tolerance threshold and stability weight mapping table to achieve adaptive calibration of the alarm strategy.

[0168] As an example, step S7 specifically includes: S7.1: Obtain the confidence scores (Conf_acc, Conf_gyro, Conf_com) of the three signals output by the lightweight decision tree model and the corresponding alarm level determination results. Based on the preset risk warning triggering rule, the execution condition judgment is performed: if Conf_acc≥0.85 and Conf_gyro≥0.75 and Conf_com≥0.7, a high-confidence joint event flag is generated as the input condition for triggering the graded warning.

[0169] Based on the structured confidence vector [Conf_acc, Conf_gyro, Conf_com] output by the lightweight decision tree model, a numerical threshold comparison method (parameters: Conf_acc threshold 0.85, Conf_gyro threshold 0.75, Conf_com threshold 0.7) is used to realize the joint condition determination of high confidence.

[0170] Furthermore, a high-confidence joint event flag is generated through a logic AND operation unit (parameters: the three input terminals correspond to the threshold comparison results of acceleration, gyroscope, and center of gravity offset, respectively), and a Boolean trigger signal is obtained.

[0171] Furthermore, an event flag encoding method (parameters: event type encoding = 0x01, high confidence joint event priority = highest level) is adopted to realize the typed encapsulation of event flags and generate serializable alarm trigger data packets.

[0172] Furthermore, by using a timestamp appending algorithm (parameters: precision 1ms, source = system real-time clock RTC), the period number and precise time mark of the event trigger are realized, and a complete event flag record entry is generated.

[0173] Through the above-mentioned judgment and encoding process, the numerical output of the lightweight decision tree is transformed into a high-confidence joint event flag with interface callability, thereby ensuring the preconditions for triggering graded early warnings.

[0174] For example, during a certain gait cycle, the system captures three confidence scores: Conf_acc=0.88, Conf_gyro=0.78, and Conf_com=0.71. After inputting these scores into the threshold comparison module, the following formulas are executed: Output 1; Output 1; Output 1.

[0175] After processing by the three-input logic and arithmetic unit, a high-confidence joint event flag bit is generated (=1), and the event type code 0x01 is assigned using the event flag encoding method, with the priority flag set to the highest level. The timestamp t=345678ms and gait cycle number N=256 collected by the RTC module are added to generate the event record {type:0x01,priority:high,time:345678ms,cycle:256}. This flag can be directly invoked in the early warning module to trigger vibration and LED yellow light alerts. Verification results show that the system maintains high stability in complex slope scenarios and significantly improves the timeliness and accuracy of alarm response.

[0176] S7.2: Based on the generation result of the high-confidence joint event flag, the execution unit in the warning module is triggered to apply a 50Hz pulse drive signal to the vibration motor and activate the LED yellow light indicator to emit a continuous warning light to provide users with a non-intrusive fall risk warning, output multimodal local warning signals, and complete the closed-loop response from judgment to physical feedback.

[0177] S7.3: When the alarm level determination result is any two Conf < 0.6 and the third Conf ≥ 0.8, a low-confidence heterogeneous combination event flag is generated as the trigger for starting the multimodal re-verification process; based on this flag, the fine-tuning sampling mechanism of the arch region in the plantar pressure sensor array is activated, and the gyroscope sampling rate is increased to 200Hz to perform high-frequency data supplementation for 150ms, generating an enhanced sensor data stream for secondary discrimination.

[0178] The confidence values ​​of each sensor in the alarm level determination results output by the lightweight decision tree model are used as the input conditions for this sub-step.

[0179] A conditional discrimination algorithm is used (rule: any two items). And the third item This enables the generation of low-confidence heterogeneous combination event flags and attaches an event trigger identifier code to them for subsequent multimodal re-verification processes.

[0180] Furthermore, by using the event identifier parser (parameter: trigger source is a low-confidence heterogeneous combination), the switch status of the multimodal reverification process inside the early warning module is set, and the sensor sampling control strategy is issued.

[0181] Furthermore, through a sensor control scheduling algorithm (mode: region selection + sampling rate overload), regional fine-tuning sampling is initiated for the plantar pressure sensor array, where two sensing units in the arch region are selected and the sampling sensitivity gain is set to its original value. This will double the accuracy of load change detection in the area.

[0182] Furthermore, by configuring the gyroscope driver instruction set (parameter: target sampling rate 200Hz), the gyroscope sampling rate is increased from the default value of 100Hz to 200Hz, and a dual-buffer structure for angular velocity data caching is enabled to balance the real-time performance and data integrity of high-frequency sampling.

[0183] Furthermore, by using the high-frequency sampling time window control module (set: duration 150ms), a synchronous sampling window is opened when the action is triggered, and the pressure data of the arch area and the high-frequency angular velocity data of the gyroscope are periodically and synchronously collected to ensure that an enhanced sensing data stream is generated under the same time reference.

[0184] By using a multimodal data aggregator, the foot arch region fine-tuning pressure data and gyroscope high-frequency angular velocity data are combined into a structured enhanced sensor dataset by weighting by timestamp, thus providing input conditions with high sensitivity and fine timing for secondary discrimination output.

[0185] For example, in a scenario of descending stairs, the confidence score output by the lightweight decision tree model is: =0.55、 =0.52、 =0.82, triggering the low-confidence heterogeneous combination event flag. The sampling control strategy boosts the arch pressure sensor gain to its original value. The gyroscope sampling rate was changed to 200Hz, and double buffering was enabled. The high-frequency sampling time window was 150ms, within which 300 points of arch pressure data and 30 cycles of gyroscope data were collected. After timestamp-weighted fusion, an enhanced dataset containing pressure load curves and angular velocity transition curves was generated. This dataset significantly improved the accuracy of phase offset calculation in the secondary discrimination stage, reduced non-physiological drift interference, and improved the reliability of early warning judgment in complex gait scenarios.

[0186] S7.4: Perform local data reprocessing on the enhanced sensor data stream: Use the constructed gait phase coordinate system φ∈[0,2π) to perform time-phase mapping on the newly added sampling points, recalculate its dominant phase offset δφ and stability weight w_stab, generate the updated three confidence scores, and input them again into the lightweight decision tree model to output the secondary alarm level judgment result, forming a dynamic verification causal chain.

[0187] The input includes an enhanced sensor data stream generated by a foot arch pressure region fine-tuning sampling mechanism and high-frequency sampling from a gyroscope. This data stream has a mappable relationship between the timestamp domain and the original gait phase coordinate system φ∈[0,2π).

[0188] Using the phase mapping function f(t) = 2π·(t) The algorithm t0) / T_step (where t0 is the peak pressure time and T_step is the current step size estimate) realizes the unified conversion from the time domain to the gait phase space of the enhanced data sampling points, and generates a phase label value φ_i for each new sampling point.

[0189] Furthermore, based on the acceleration energy envelope, the gyroscope angular velocity absolute value envelope, and the center of gravity shift rate of change signal, the phase angle of the dominant event is extracted in the phase coordinate system. By calculating the absolute phase offsets δφ_acc, δφ_gyro, and δφ_com of the three types of signals, an updated phase deviation vector set is formed relative to the zero point φ0 of the pressure main peak phase.

[0190] Furthermore, the standard deviation of the updated phase offset is calculated using a five-cycle sliding window mechanism to obtain σ_δφ_acc, σ_δφ_gyro, and σ_δφ_com, which are used to characterize the phase stability of the enhanced data stream in a short time series. High, medium, and low level stability weights w_stab_acc, w_stab_gyro, and w_stab_com are then mapped according to the stability interval threshold.

[0191] Furthermore, substituting the updated phase offset and stability weights into the confidence calculation formula:

[0192] Where I(·) is the indicator function and π is the ratio of π to π, ensuring that all three types of signals obtain a new confidence scalar after dynamic calibration.

[0193] The above calculations generate a new structured confidence input vector [Conf_acc, Conf_gyro, Conf_com]. This vector is then input into the lightweight decision tree model, and the secondary alarm level determination result is re-output based on the updated hard threshold rule, forming a complete dynamic verification causal chain.

[0194] By reprocessing and re-discriminating local data, the phase consistency index of the enhanced sensor data stream is transformed into a structured confidence vector required for secondary judgment, enabling real-time re-judgment and false alarm suppression of potential risk events in low-confidence combined early warning scenarios.

[0195] For example, in a stair-down scenario, the time resolution for fine-tuning the sampling of the arch pressure region is set to 5 ms, the gyroscope sampling rate is increased to 200 Hz, and the enhanced data stream covers additional detail sampling of two phase slots. During phase mapping, assuming the pressure peak time t0 = 1.250 s and the step size estimate T_step = 1.000 s, a certain acceleration extreme point occurs at t = 1.325 s, and its phase mapping value is calculated as follows:

[0196] The calculated value is φ≈0.471rad, and the offset relative to φ0 is δφ_acc=0.471rad. Within a five-cycle window, the signal's σ_δφ_acc=0.200rad, mapped to a moderate stability weight w_stab_acc=0.6, yields Conf_acc≈0.88 when substituted into the formula. The gyroscope and center-of-gravity offset channels are calculated using the same steps, resulting in Conf_gyro=0.74 and Conf_com=0.71. These are input into a decision tree for secondary discrimination, outputting a 'risk warning' command and updating the corresponding control message. This example verifies the significant improvement in warning accuracy and response speed achieved by enhanced sampling and phase recalculation in complex gait environments.

[0197] S7.5: When all three Conf values ​​are <0.5, a phase loss event flag is generated as the basis for suppressing the current alarm command. Based on this flag, the vibration motor and the audible and visual alarm are prohibited from starting, and the diagnostic log, which includes the current gait cycle number, the dominant phases of each signal φ_acc, φ_gyro, φ_com and their offsets δφ_acc, δφ_gyro, δφ_com, is written to the edge storage area for subsequent online calibration of the individualized model.

[0198] Step S8: Aggregate historical phase loss events every 24 hours, statistically analyze the operating characteristics under high-frequency loss scenarios, and dynamically update the individualized phase offset tolerance threshold and stability weight mapping table to achieve adaptive calibration of the alarm strategy. Specifically, this includes: S8.1: Obtain all 'phase lockout' event data recorded in the edge logs over the past 24 hours, and generate a phase lockout event sequence based on the timestamp of the event, the corresponding gait cycle number, and the combination of triggering sensors, as the input basis for the statistical analysis of operating conditions.

[0199] S8.2: Perform condition clustering analysis on the phase loss event sequence. Use the root mean square of acceleration, the rate of change of plantar pressure gradient and the ground tilt angle estimate corresponding to the gait cycle as clustering feature vectors. Use a lightweight K-means algorithm (K=5) to classify high-frequency loss scene categories and generate a set of condition cluster labels with physiological significance.

[0200] S8.3: For each high-frequency operating condition cluster, calculate the mean of the phase offset standard deviation σ_δφ distribution within five consecutive sampling windows, and optimize the individualized phase offset tolerance threshold based on the statistical results: if the mean is consistently higher than π / 6, then the δφ alarm trigger threshold for the user's corresponding scenario will be dynamically relaxed from π / 4 to π / 3.5 to reduce the false alarm rate.

[0201] S8.4: Based on the loss-of-lock frequency of the three types of signals (acceleration, gyroscope, and center of gravity shift) in each working condition cluster, reconstruct the stability weight mapping table: assign lower basic stability weights to signal channels that frequently lose lock in a specific cluster, and generate a dynamic weight configuration file with scene awareness capabilities.

[0202] S8.5: Write the updated individualized phase offset tolerance threshold and dynamic weight configuration file to the local non-volatile storage area, replace the original parameter table, and mark the start time of the next calibration cycle to complete the closed-loop adaptive calibration of the alarm strategy.

[0203] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0204] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0205] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A positioning and drop prevention method based on multi-mode signal processing, specifically including: S1: Perform two-dimensional Gaussian kernel smoothing and normalization on the original pressure data to generate a spatially continuous pressure distribution matrix; S2: Extract the pressure center trajectory from the pressure distribution matrix and construct a local motion manifold by combining the pressure center displacement vectors of three consecutive frames; S3: Identify the maximum pressure cluster in the heel region within a single step cycle as the main pressure peak, and establish a gait phase coordinate system with the time corresponding to the main pressure peak as the phase zero point. S4: Map the timestamps of each input signal, including the triaxial acceleration energy envelope, the gyroscope angular velocity absolute value envelope, and the extreme point of the center of gravity shift rate of change, to the gait phase coordinate system, obtain the dominant phase position of the input signal on the phase axis, and calculate its offset relative to the pressure peak phase. S5: Based on the phase offset of each input signal and its standard deviation of offset over five consecutive gait cycles, evaluate phase consistency and evolutionary stability, and generate an initial confidence score. S6: Calculate the final confidence scores of acceleration, gyroscope and center of gravity offset signals based on the initial confidence scores, and input the three scores into the preset lightweight decision tree model to output the corresponding alarm level judgment result.

2. The positioning and fall prevention method based on multi-mode signal processing according to claim 1, characterized in that, Step S3 further includes: discretizing the phase axis of the gait phase coordinate system into multiple equally divided phase slots to obtain a unified time-phase mapping reference.

3. The positioning and fall prevention method based on multi-mode signal processing according to claim 1, characterized in that, Following step S6, the following is also included: S7: Determine whether the alarm level determination result meets the risk warning conditions. If it does, trigger the vibration motor and LED yellow light prompt. If it is determined to be a low confidence combination, start the multimodal re-verification process or suppress the alarm and record the phase loss event. S8: Aggregates historical phase loss events, statistically analyzes the operating characteristics of high-frequency loss scenarios, and dynamically updates the individualized phase offset tolerance threshold and stability weight mapping table to achieve adaptive calibration of the alarm strategy.

4. The positioning and drop prevention method based on multi-mode signal processing according to claim 1, characterized in that, Step S1 specifically includes: S1.1: Acquire the raw pressure data frame output by the insole foot pressure sensor array; S1.2: Construct a two-dimensional spatial grid mapping relationship based on the original pressure data frame to generate a sparse pressure space matrix, so as to establish a geometric correspondence between discrete sampling points and continuous foot regions; S1.3: Apply a two-dimensional Gaussian kernel convolution operation to the sparse pressure space matrix, and use a Gaussian kernel function with standard deviation grid units for spatial smoothing to generate a denoised continuous pressure distribution matrix. S1.4: Calculate the L2 norm of all grid pressure values ​​in the continuous pressure distribution matrix, and perform global amplitude normalization on the matrix based on the norm to generate a normalized pressure distribution matrix; S1.5: Output the normalized pressure distribution matrix.

5. The positioning and drop prevention method based on multi-mode signal processing according to claim 1, characterized in that, Step S2 specifically includes: S2.1: Obtain the pressure distribution matrix sequence after two-dimensional Gaussian kernel smoothing and normalization. Based on the spatial coordinates of each sensor position in each frame matrix and its corresponding pressure intensity value, calculate the weighted average position and generate the pressure center coordinates at each sampling time. Perform vector difference operation on the pressure center coordinate sequence within the continuous time window to obtain the pressure center displacement vector between adjacent frames. S2.2: Construct a local velocity field based on the pressure center displacement vector of three consecutive frames. Take the displacement vector of the current frame as the main driving vector. Perform normalized projection processing on the displacement vectors of the previous and next frames to determine the cosine value of the angle between them and the main vector and the amplitude ratio. And determine the potential gait phase transition point based on the cosine value of the angle and the amplitude ratio. S2.3: Construct a two-dimensional motion manifold subspace using three continuous displacement vectors in the local velocity field, generate a set of orthogonal bases using the Gram-Schmidt orthogonalization method, and project the original displacement vectors onto this low-dimensional manifold space to obtain the manifold projection coefficients in a compressed representation. S2.4: Perform principal component analysis on the manifold projection coefficients, extract the eigenvector corresponding to the first principal component as the dominant motion direction, and calculate its cumulative variance contribution rate; establish the three-frame window with a contribution rate greater than the preset value as the stable load transfer stage; S2.5: Combine the analysis of the dominant motion direction and the radius of curvature of the pressure center trajectory to calculate the local dynamic stability index: generate a quantitative score based on the rate of change of the tangent angle of the pressure center path and the energy concentration of the manifold.

6. The positioning and fall prevention method based on multi-mode signal processing according to claim 1, characterized in that, Step S3 specifically includes: S3.1: Acquire the raw pressure time series data of the plantar pressure sensor array in a single gait cycle. Based on the local pressure signal sequence output by the three pressure sensing units arranged in the heel region, perform sliding window mean filtering to eliminate high-frequency physiological noise and motion artifact interference, and generate a smoothed synthetic pressure waveform in the heel region as the input signal for subsequent pressure peak detection. S3.2: Perform peak detection algorithm on the smoothed heel region synthetic pressure waveform to locate the pressure main peak with the largest amplitude and duration that meets the physiological constraints of the ground contact phase in each gait cycle, mark its corresponding timestamp as the phase zero point of the current gait cycle, and generate a starting reference point with biomechanical significance to construct the gait phase coordinate system; S3.3: Based on the phase zero-point sequence extracted from multiple consecutive gait cycles, the average gait cycle duration is calculated. This statistic is used to model individualized gait rhythms, and a single gait cycle is mapped to a normalized phase interval [0, 2π). A nonlinear time compression function matching the user's actual walking rhythm is generated as the core mapping mechanism for time-phase conversion. S3.4: The normalized gait phase interval is divided into multiple discrete phase slots, each phase slot corresponding to a fixed arc span, forming a gait phase coding space with a fixed resolution, realizing the discrete phase labeling of data collected at different times, and providing a quantitative scale for the alignment analysis of cross-modal signals under a unified phase system. S3.5: Output a set of phase mapping parameters containing information on the phase zero point, the current gait cycle length, and the boundary of the equally divided phase slots. This set serves as a common reference for the conversion of the acceleration energy envelope, gyroscope angular velocity envelope, and the extremum timestamp of the center of gravity shift rate of change into phase coordinates in subsequent steps.

7. The positioning and drop prevention method based on multi-mode signal processing according to claim 1, characterized in that, Step S4 specifically includes: S4.1: Obtain the raw acceleration time-domain signal output by the triaxial accelerometer. Based on the curvature radius information of the local motion manifold subspace and the trajectory of the pressure center, perform envelope demodulation processing on the acceleration vector synthesis signal. Use the Hilbert transform algorithm to extract the acceleration energy envelope to obtain a continuous scalar sequence characterizing the motion impact intensity over time, and generate the acceleration energy evolution field. S4.2: Based on the global maximum point in the acceleration energy evolution field, locate its corresponding timestamp, combine it with the phase mapping parameter set output from the previous step, and use the linear normalization mapping mechanism to convert the timestamp into the corresponding gait phase coordinates to generate the acceleration dominant phase angle; S4.3: Obtain the time-domain signal of angular velocity. Based on the zero point of the pressure main peak phase and the nonlinear compression function of the gait period, perform absolute value conversion and sliding window root mean square filtering on the angular velocity signal to generate the absolute value envelope of angular velocity. Extract the timestamps corresponding to the extreme points and convert them into the gyroscope dominant phase angle according to the same phase mapping rule to generate the rotation dynamics dominant phase parameter. S4.4: Obtain the center of gravity offset signal jointly estimated by attitude angle integral and Kalman filter. Based on the dynamic stability index and the energy concentration of the local motion manifold, perform a first-order difference operation on the signal to generate a rate of change signal. Detect the timestamp corresponding to the extreme point with the largest amplitude and convert it into the dominant phase angle of the center of gravity offset through the phase mapping function to generate the phase labeling result. S4.5: Based on the acceleration-dominant phase angle, gyroscope-dominant phase angle, and center of gravity offset-dominant phase angle, combined with the pressure peak phase zero point, calculate the absolute phase offset of the three relative to the phase zero point, and output the phase deviation vector group.

8. The positioning and fall prevention method based on multi-mode signal processing according to claim 1, characterized in that, Step S5 specifically includes: S5.1: Obtain the acceleration-dominant phase, gyroscope-dominant phase, and center of gravity shift rate-dominant phase, and calculate the phase shift of each signal in the current gait cycle based on the zero point of the pressure peak phase, and generate a preliminary offset parameter set; S5.2: Based on the phase offset sequences of each signal stored in the historical five-step periodic buffer queue, extract the continuous five-cycle sequence corresponding to the current signal using a sliding time window, and calculate its phase offset standard deviation; S5.3: Based on the comparison between the phase offset and the preset physiological tolerance threshold, if the phase offset is greater than the preset physiological tolerance threshold and continues to exceed a preset number of phase slots, the phase drift flag is triggered to generate an offset penalty factor. S5.4: Determine the stability level based on the interval to which the standard deviation of the phase offset belongs; S5.5: Combine the offset penalty factor and stability weight to generate the initial confidence score of each signal, and output the initial confidence score group of the three channels of acceleration, gyroscope and center of gravity offset.

9. A positioning and fall protection system based on multi-mode signal processing, characterized in that, The positioning and fall prevention method based on multi-mode signal processing described in claim 1 is used for fall prevention early warning for the elderly.

10. A wearable device, characterized in that, The equipment is a positioning and drop protection system based on multi-mode signal processing as described in claim 9.

Images