Stress damage early prediction system and method based on flexible plantar pressure insole

Abstract

The invention provides a stress damage early prediction system and method based on a flexible plantar pressure insole, and belongs to the technical field of sports medicine and intelligent wearable, and the system comprises a sensing acquisition module used for controlling a sensor array in the flexible plantar pressure insole to acquire plantar pressure signals, a signal conditioning module used for conditioning the plantar pressure signals, and a data processing module used for processing the plantar pressure signals. The system comprises a data processing module used for carrying out signal preprocessing on collected plantar pressure signals, a data processing module used for carrying out plantar continuous pressure field reconstruction on the preprocessed signals, and a risk assessment module used for extracting 52-dimensional feature vectors from pressure field reconstruction results and inputting the 52-dimensional feature vectors into a constructed depth-integration double-layer model to carry out quantitative assessment on damage risks. And outputting a multi-mode early warning. The problems that an existing plantar pressure monitoring device is low in precision, poor in scene adaptability and lack of intelligent early warning capacity are solved.

Description

Technical Field

[0001] This invention relates to the fields of sports medicine and smart wearable technology, and in particular to an early prediction system and method for stress injuries based on flexible plantar pressure insoles. Background Technology

[0002] Plantar pressure is a core indicator reflecting the biomechanical function of the foot, and its abnormal distribution is a significant cause of stress injuries. Current plantar pressure monitoring and injury early warning technologies have the following key shortcomings:

[0003] Limited monitoring scenarios: Traditional laboratory force measurement platforms (such as the Kistler force table) have high accuracy (error of about 5%), but they are bulky (occupying an area of ​​≥1m²). 2 They are expensive (≥500,000 yuan per unit), can only perform low-frequency detection in fixed scenarios, cannot capture dynamic stress changes during daily exercise or training, and cannot meet the needs of continuous monitoring.

[0004] Poor dynamic anti-interference ability: Existing flexible pressure insoles (such as smart insoles on the market) use simple Kalman filtering or mean filtering, which cannot effectively suppress motion artifacts (foot sliding) and temperature and humidity drift in dynamic scenarios such as running and jumping. The data distortion rate is ≥15%, which makes the pressure signal unable to reflect the real force state.

[0005] Low accuracy of pressure field reconstruction: Most products can only output the peak / average pressure values ​​of discrete sensing points, lack the ability to reconstruct continuous pressure fields, and cannot reveal the spatiotemporal evolution of pressure hotspots on the sole of the foot (such as metatarsal heads), making it difficult to identify local pressure anomalies corresponding to early micro-damage.

[0006] The lack of quantitative models for injury early warning: Existing systems can only passively display pressure data and have not established a correlation model between pressure characteristics and injury risk. This makes it impossible to achieve a closed loop of data → risk → intervention, resulting in delayed early warning (usually diagnosis can only be made after clinical symptoms such as pain appear), and missing the best time for intervention.

[0007] Therefore, this invention proposes an early prediction system and method for stress injuries based on flexible plantar pressure insoles. Summary of the Invention

[0008] This invention provides a system and method for early prediction of stress injuries based on flexible plantar pressure insoles, in order to solve the aforementioned technical problems.

[0009] This invention provides an early prediction system for stress injuries based on flexible plantar pressure insoles, comprising:

[0010] The sensing and acquisition module is used to control the sensor array in the flexible plantar pressure insole to acquire plantar pressure signals;

[0011] The signal conditioning module is used to preprocess the acquired plantar pressure signals. The signal preprocessing includes motion artifact filtering, baseline correction, and temperature and humidity compensation.

[0012] The data processing module is used to reconstruct the continuous pressure field of the foot on the preprocessed signal. The pressure field reconstruction includes: preliminary interpolation, finite element constraint optimization and physiological and physical constraints.

[0013] The risk assessment module is used to extract 52-dimensional feature vectors from the pressure field reconstruction results and input them into the constructed deep-integrated two-layer model to quantify the damage risk assessment and output multimodal early warning.

[0014] Preferably, the sensor array is based on an elastic polymer and covers the heel, forefoot, metatarsal heads and arch area, with 16 to 128 sensing points. The center-to-center distance of the sensors in key areas is ≤2mm, and the sampling rate is 500Hz. The collected analog signals are amplified by a preamplifier and then converted into digital signals by an ADC, wherein the plantar pressure signal is the digital signal.

[0015] Preferably, the signal conditioning module includes:

[0016] The motion artifact filtering unit is used to process the digital signal using an improved RLS filtering algorithm. The filtering formula is as follows:

[0017] h[n+1]=h[n]+K[n](d[n]-h T [n]x[n]);

[0018] Where h[n] is the filter coefficient vector at time n; h[n+1] is the filter coefficient vector at time n+1; K[n] is the gain vector at time n; d[n] is the desired pressure signal vector at time n; x[n] is the input signal vector at time n; h T [n] is the transpose of the filter coefficient vector at time n;

[0019] The baseline correction unit is used to calculate the original sampled value of each sensing point in the k-th frame based on a multi-scale sliding window family. sliding window mean μ k According to the formula Achieve baseline drift correction;

[0020] The temperature and humidity compensation unit is used to correct the sensor resistance value using a fourth-order polynomial compensation model to address temperature and humidity interference. The compensation formula is: ρ corr =ρ raw +cp1(∈)+cp2(∈ 2 )+cp3(∈ 3 )+cp4(∈4 ), where ∈ is the normalized temperature difference; ρ corr The resistance value after compensation; ρ raw The original resistance value is the uncompensated value; cp1, cp2, cp3, and cp4 are compensation coefficients.

[0021] Preferably, the data processing module includes:

[0022] The initial interpolation unit is used to reconstruct the initial continuous pressure field using RBF interpolation based on the coordinates (xi, yi) of the preprocessed discrete sensing points and the corresponding pressure values ​​Pi. The interpolation formula is as follows: Where N is the number of sensing points; α i σ is the interpolation coefficient; σ is the region adaptive scaling parameter, σ = 2mm for the metatarsal head region and σ = 5mm for the heel region;

[0023] Finite element constrained optimization elements are used to solve the objective function min based on a three-dimensional finite element model of the foot, using the initial continuous pressure field as the initial value. u AP-V 2 +γR(P), where A is the sensor sensitivity matrix; V is the vector composed of digital signals; P is the pressure field distribution; R(P) is the regularization term based on foot anatomy; and γ is the regularization parameter.

[0024] The physiological and physical constraint unit is used to apply total force conservation constraints and low-pressure constraints on the arch of the foot, and outputs a smooth and continuous pressure field across the entire plantar surface.

[0025] Preferably, the risk assessment module includes:

[0026] The vector extraction unit is used to extract a 52-dimensional feature vector from the pressure field reconstruction results, based on time-domain features, frequency-domain features, morphological features, and complexity features.

[0027] The model analysis unit is used to input the 52-dimensional feature vector into a deep-ensemble two-layer model, output 64-dimensional high-level features based on the first layer, and process the high-level features using Soft-GradientBoost ensemble to output the damage risk probability. Where ε() is the Sigmoid function; h r0 (o) is the r0th weak classifier for analyzing high-level feature o, and R0 = 100. The first layer is a feature transformation layer, which is composed of a double hidden layer fully connected network. The second layer is a risk prediction layer, which is integrated by Soft-GradientBoost.

[0028] The deep-ensemble two-layer model employs Focal-Tversky composite loss to suppress positive and negative sample imbalance, with the loss formula L being:

[0029]

[0030] Where γ0 is the focusing parameter, with a value of 2.0; α1 is... The weighting coefficient is 0.35; y is the damage label. Used for quantization Similarity between them;

[0031] Risks are categorized into low-risk, medium-risk, and high-risk based on probability, and multimodal early warning is implemented.

[0032] Preferably, the three-dimensional finite element model of the foot is a multi-layer structure including bone, soft tissue, and insole, realized based on high-resolution CT / MRI images of the foot, using Delaunay-Voronoi hybrid subdivision to generate ≥2000 quadrilateral elements, and assigning elastic modulus field and Poisson's ratio field according to the differences in anatomical regions, with the elastic modulus of the heel region set to 1.2 GPa, the metatarsal head region set to 1.8 GPa, and the soft tissue region set to 0.3 MPa.

[0033] Preferred options also include:

[0034] The anatomy module is used to divide the anatomical surface of the foot into 11 macro-regions and 47 secondary functional units based on the dynamically updated three-dimensional finite element stress knowledge base of the foot, and to reserve magnetic subnet interfaces at the extreme arch morphology positions.

[0035] The scoring calculation module is used to calculate a composite score Ξ_k∈[0,100] within each of the secondary functional units by comprehensively considering the three-dimensional indices of local load increment, historical damage attribution rate, and manufacturing tolerance, and then map it to the sensor node surface density Δ. k :

[0036]

[0037] Where, Δ min =3.4pts / cm 2 ,Δ max =21.7pts / cm 2 The index 1.15 was obtained by fitting the error-cost curve to the multi-objective genetic search.

[0038] The search module is used to set the set of geometric center coordinates of all sensitive units as... Define the centroid error functional Measure the local error of the sensor array relative to the center of the target region; define the manufacturing budget functional. Quantify the manufacturing cost and process complexity of the layout; adjust the target weight ratio The array geometry optimization problem is solved using a hybrid ant colony-SGD search algorithm. The outer layer generates 8872 candidate clusters through 60 iterations of the ant colony algorithm, while the inner layer fine-tunes each cluster for 120 rounds using the Adam-SGD algorithm to obtain the coordinates of sensitive cells that balance accuracy and cost. Nu represents the total number of sensitive cells, and rp0 represents the geometric center coordinates of the p0th sensitive cell.

[0039] Based on the search results, a sheet sensor was deployed to cover an area of ​​113×87mm to obtain a sensor array.

[0040] This invention provides a method for early prediction of stress injuries based on flexible plantar pressure insoles, comprising:

[0041] Step 1: Control the sensor array in the flexible plantar pressure insole to collect plantar pressure signals;

[0042] Step 2: Perform signal preprocessing on the acquired plantar pressure signals, which includes motion artifact filtering, baseline correction, and temperature and humidity compensation.

[0043] Step 3: Reconstruct the continuous pressure field of the foot on the preprocessed signal. The pressure field reconstruction includes: preliminary interpolation, finite element constraint optimization and physiological and physical constraints.

[0044] Step 4: Extract 52-dimensional feature vectors from the pressure field reconstruction results and input them into the constructed deep-integrated two-layer model to quantify the damage risk assessment and output multimodal early warning.

[0045] Compared with the prior art, the beneficial effects of this application are as follows:

[0046] In terms of measurement accuracy, through optimized sensor layout and hybrid reconstruction algorithm, the error of plantar pressure distribution reconstruction is controlled within 4.5%, and the spatial resolution of key areas reaches 2mm. In terms of system performance, a dedicated signal processing chip is used to achieve parallel computing, with a processing latency of less than 30ms and a sampling rate of up to 500Hz. In terms of wearing comfort, the multi-layer flexible structure design makes the insole thickness less than 3mm and the bending stiffness less than 0.1N·m. In terms of clinical application value, it can provide an early warning of stress injury 21 days in the athlete population with an accuracy rate of 89.7%.

[0047] In the prevention of diabetic foot, the system's detection of early neuropathy achieved a sensitivity of 93.4% and a specificity of 87.6%. In training scenarios, real-time feedback reduced peak plantar pressure by 40% and the incidence of stress fractures by 68%. In the field of rehabilitation medicine, the correlation coefficient between the system's evaluation results and standard clinical scales remained above 0.85.

[0048] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0049] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0051] Figure 1 This is a structural diagram of the stress injury early prediction system based on flexible plantar pressure insoles in an embodiment of the present invention;

[0052] Figure 2 This is a flowchart of an early prediction method for stress injuries based on flexible plantar pressure insoles, as described in an embodiment of the present invention. Detailed Implementation

[0053] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0054] This invention provides an early prediction system for stress injuries based on flexible plantar pressure insoles, such as... Figure 1 As shown, it includes:

[0055] The sensing and acquisition module is used to control the sensor array in the flexible plantar pressure insole to acquire plantar pressure signals;

[0056] The signal conditioning module is used to preprocess the acquired plantar pressure signals. The signal preprocessing includes motion artifact filtering, baseline correction, and temperature and humidity compensation.

[0057] The data processing module is used to reconstruct the continuous pressure field of the foot on the preprocessed signal. The pressure field reconstruction includes: preliminary interpolation, finite element constraint optimization and physiological and physical constraints.

[0058] The risk assessment module is used to extract 52-dimensional feature vectors from the pressure field reconstruction results and input them into the constructed deep-integrated two-layer model to quantify the damage risk assessment and output multimodal early warning.

[0059] In this embodiment, the multimodal early warning includes:

[0060] Visual: Real-time rendering of 3D stress-time volume using OpenGL-ES on mobile devices; synchronous display of color-coded risk indicators (four levels: green, blue, orange, and red).

[0061] Hearing: Bone conduction headphones output a three-frequency envelope (400 / 800 / 1600Hz); the sound pressure level increases linearly with risk up to 78dB.

[0062] Tactile feedback: The insole has an embedded LRA (linear resonant actuator) in the forefoot and heel, with an amplitude of 1.5g to 3g and a pulse width of 50 to 200ms, which can be adjusted as needed.

[0063] Automatic matching of intervention strategies includes:

[0064] The system includes 84 parameterized recipes across three main categories: load redistribution, gait pattern, and training volume fine-tuning. Matching rules are as follows:

[0065] when And when the hotspot load ratio is >2.8×BW, immediately push the stride-cadence fine-tuning scheme (interactive metronome + visual comparison); when The system recommends a combination of ice application, dynamic stretching, and insole replacement within 6 hours after the race. Here, BW represents body weight.

[0066] Only standard monitoring prompts are provided at this time.

[0067] Automatically collect historical data for 10 seconds after intervention; if the peak pressure decreases by <12% or If the decrease is less than 0.05, it is marked as failing and automatically upgraded to a higher level for intervention; in the experimental cohort (number of 120), the closed-loop success rate was 88.4%, with an average of 1.3 upgrades to achieve the target.

[0068] In this embodiment, the hardware platform includes: a sensing insole (240×85×2.9mm), a processing unit (dual-core Cortex-M7 MCU), and a mobile terminal. Key performance parameters: measurement range 5-1200kPa, sensitivity 0.3kPa, wireless transmission distance 20m, operating current 18mA@100Hz.

[0069] Clinical validation involved 120 subjects and followed them for 6 months. The area under the ROC curve for predicting stress injury was 0.916 (95% CI: 0.892–0.941). Comparative testing with an optical force plate system showed an intraclass correlation coefficient (ICC) of 0.972 for peak pressure measurement.

[0070] Preferably, the sensor array is based on an elastic polymer and covers the heel, forefoot, metatarsal heads and arch area, with 16 to 128 sensing points. The center-to-center distance of the sensors in key areas is ≤2mm, and the sampling rate is 500Hz. The collected analog signals are amplified by a preamplifier and then converted into digital signals by an ADC, wherein the plantar pressure signal is the digital signal.

[0071] Preferably, the signal conditioning module includes:

[0072] The motion artifact filtering unit is used to process the digital signal using an improved RLS filtering algorithm. The filtering formula is as follows:

[0073] h[n+1]=h[n]+K[n](d[n]-h T [n]x[n]);

[0074] Where h[n] is the filter coefficient vector at time n; h[n+1] is the filter coefficient vector at time n+1; K[n] is the gain vector at time n; d[n] is the desired pressure signal vector at time n; x[n] is the input signal vector at time n; h T [n] is the transpose of the filter coefficient vector at time n;

[0075] The baseline correction unit is used to calculate the original sampled value of each sensing point in the k-th frame based on a multi-scale sliding window family. sliding window mean μ k According to the formula Achieve baseline drift correction;

[0076] The temperature and humidity compensation unit is used to correct the sensor resistance value using a fourth-order polynomial compensation model to address temperature and humidity interference. The compensation formula is: ρ corr =ρ raw +cp1(∈)+cp2(∈ 2 )+cp3(∈ 3 )+cp4(∈ 4 ), where ∈ is the normalized temperature difference; ρ corr The resistance value after compensation; ρ raw The original resistance value is the uncompensated value; cp1, cp2, cp3, and cp4 are compensation coefficients.

[0077] In this embodiment, the analog front-end of the signal conditioning module adopts a cascaded structure of chopper-stabilized preamplifier + programmable resistor array gain unit + third-order Σ-Δ modulator + fifth-order CIC digital demodulation. The overall bandwidth is determined by an external passive anti-aliasing filter. Among them, the restriction is on f c =510Hz, ensuring twice the Nyquist margin. Zero bias suppression: utilizing alternating polarity biasing technology, every T x=5ms for input signal v in (t) Invert once and modulate to carrier frequency w x Nearby; after passing through the phase-locked tracking demodulator, the residual 1 / f noise power is reduced to: Where N0 is the noise floor density, w0 is the noise inflection point, and the exponent λ≈0.9 is determined by the on-chip substrate noise model. c The channel cutoff frequency is 510Hz. Physically, this represents the upper limit of the signal bandwidth allowed through the analog link, ensuring complete acquisition of the plantar pressure signal's spectral components while filtering out high-frequency noise. As a hardware design parameter, f_c is set to a value slightly higher than half the sampling rate (510Hz corresponds to twice the Nyquist frequency margin), fixed through a passive anti-aliasing filter. During calibration, this value is selected based on the system's sampling rate and signal spectrum requirements to avoid signal aliasing while preserving the main frequency components of foot dynamics. T x The chopper bias period is the parameter that determines how the low-frequency offset and 1 / f noise are modulated onto the carrier frequency w. x Nearby, thus utilizing subsequent synchronous demodulation for filtering. As a design parameter, T x The amplifier architecture must be chosen to be small enough to push flicker noise outside the signal bandwidth (5ms is chosen in this system to balance noise suppression and power consumption). In implementation, this cycle is generated by a clock circuit to trigger the inverting input of the chopper switch. x =2π / T x And w x This is the angular frequency of the chopper carrier wave. in (t) becomes -v after inversion. in (t). N0 is obtained through chip noise characteristic measurement or datasheet. w0 is given by the manufacturer depending on the chip manufacturing process or obtained through noise spectrum measurement. λ is obtained by measuring the spectral curve of the chip substrate noise and performing characteristic fitting.

[0078] In this embodiment, the programmable gain array discretizes the amplification factor into G = {2}. 5 ,2 6 ,…,2 12 Let the instantaneous input peak-to-peak amplitude be App, and the target full-scale AFS = 1.9V, then the adaptive tracing algorithm is selected.

[0079] G*=max{g∈Gg·App<0.9AFS}, where G* is the optimal gain; App is the instantaneous peak-to-peak amplitude;

[0080] Gain self-consistent convergence speed < 4 frames, insert half-frame fade-in window ω during switching. blend(n) = (1-cos(πn / M)) / 2 suppresses jumps.

[0081] Σ-Δ modulation: modulator noise transfer function Tn (z)=(1-z -1 ) 3 Signal transmission function T s (z)=z -3 .

[0082] Preferably, the data processing module includes:

[0083] The initial interpolation unit is used to reconstruct the initial continuous pressure field using RBF interpolation based on the coordinates (xi, yi) of the preprocessed discrete sensing points and the corresponding pressure values ​​Pi. The interpolation formula is as follows: Where N is the number of sensing points; α i σ is the interpolation coefficient; σ is the region adaptive scaling parameter, σ = 2mm for the metatarsal head region and σ = 5mm for the heel region;

[0084] Finite element constrained optimization elements are used to solve the objective function min based on a three-dimensional finite element model of the foot, using the initial continuous pressure field as the initial value. u AP-V 2 +γR(P), where A is the sensor sensitivity matrix; V is the vector composed of digital signals; P is the pressure field distribution; R(P) is the regularization term based on foot anatomy; and γ is the regularization parameter.

[0085] The physiological and physical constraint unit is used to apply total force conservation constraints and low-pressure constraints on the arch of the foot, and outputs a smooth and continuous pressure field across the entire plantar surface.

[0086] In this embodiment, the pressure field reconstruction includes: normalizing the spatial coordinates of the sensitive unit; defining a kernel family consisting of L scales and adaptively fitting a multi-scale radial kernel; decomposing the pressure field within the normalization domain Ω; and performing a physically consistent projection (to constrain the reconstruction results to conform to the foot contact boundary conditions, a potential function S(u,v) is introduced, making...). S(u,v)=—η(1—ω(u,v)), where ω(u,v) is the foot geometry mask and η is the lateral offset that is adaptively updated based on the integral of the negative values ​​across the entire field. The result is the decomposition result; Π(u,v) represents the boundary conditions.

[0087] In this embodiment, the adaptive scale radial kernel fitting is:

[0088] λ_l=λ_minθ l-1 ;

[0089] Where d is the spatial distance; α_l is the smoothness control parameter, ranging from 0 to 2; θ is the scale expansion factor, with a value greater than 1; and λ_min is the minimum kernel scale parameter.

[0090] In this embodiment, the pressure field reconstruction solution equation is:

[0091]

[0092] Where ε1 is the kernel ridge regularization parameter; b1 is a matrix composed of radial kernels; b2 is the pressure field reconstruction coefficient vector; The linearized pressure measurement vector δ_lj is the Kronecker delta function; x_r is the original coordinate value of the r-th sensor in the l-direction; x_s is the original coordinate value of the s-th sensor in the j-direction.

[0093] Preferably, the risk assessment module includes:

[0094] The vector extraction unit is used to extract a 52-dimensional feature vector from the pressure field reconstruction results, based on time-domain features, frequency-domain features, morphological features, and complexity features.

[0095] In this embodiment, the time-domain characteristics include: peak pressure P max Pressure-time integral Gait period T0, contact time CT; frequency domain characteristics include: fast Fourier transform dominant frequency fm, spectral energy distribution E(f); morphological characteristics include: pressure center trajectory COP(t), forefoot-to-hindfoot pressure ratio Fr, left-right foot symmetry index. Complexity features include: sample entropy and multi-scale entropy; P(t) is the pressure at time t, t∈[t1,t2]; P left The pressure value for the left foot; P right The pressure value for the right foot is shown. It should be noted that the time domain features have 8 dimensions, the frequency domain features have 12 dimensions, the morphological features have 10 dimensions, and the complexity features have 22 dimensions.

[0096] The model analysis unit is used to input the 52-dimensional feature vector into a deep-ensemble two-layer model, output 64-dimensional high-level features based on the first layer, and process the high-level features using Soft-GradientBoost ensemble to output the damage risk probability. Where ε() is the Sigmoid function; h r0 (o) is the r0th weak classifier for analyzing high-level feature o, and R0 = 100. The first layer is a feature transformation layer, which is composed of a double hidden layer fully connected network. The second layer is a risk prediction layer, which is integrated by Soft-GradientBoost.

[0097] The deep-ensemble two-layer model employs Focal-Tversky composite loss to suppress positive and negative sample imbalance, with the loss formula L being:

[0098]

[0099] Where γ0 is the focusing parameter, with a value of 2.0; α1 is... The weighting coefficient is 0.35; y is the damage label. Used for quantization Similarity between them;

[0100] Risks are categorized into low-risk, medium-risk, and high-risk based on probability, and multimodal early warning is implemented.

[0101] In this embodiment, the first layer is: g = ELU(W1z+b1), o = ELU(W2g+b2), where g is the intermediate transformation feature of the first layer; W1 and W2 are the weight matrices of the neural network; and b1 and b2 are the bias vectors of the neural network.

[0102] Preferably, the three-dimensional finite element model of the foot is a multi-layer structure including bone, soft tissue, and insole, realized based on high-resolution CT / MRI images of the foot, using Delaunay-Voronoi hybrid subdivision to generate ≥2000 quadrilateral elements, and assigning elastic modulus field and Poisson's ratio field according to the differences in anatomical regions, with the elastic modulus of the heel region set to 1.2 GPa, the metatarsal head region set to 1.8 GPa, and the soft tissue region set to 0.3 MPa.

[0103] Preferred options also include:

[0104] The anatomy module is used to divide the anatomical surface of the foot into 11 macro-regions and 47 secondary functional units based on the dynamically updated three-dimensional finite element stress knowledge base of the foot, and to reserve magnetic subnet interfaces at the extreme arch morphology positions.

[0105] The scoring calculation module is used to calculate a composite score Ξ_k∈[0,100] within each of the secondary functional units by comprehensively considering the three-dimensional indices of local load increment, historical damage attribution rate, and manufacturing tolerance, and then map it to the sensor node surface density Δ. k :

[0106]

[0107] Where, Δ min =3.4pts / cm 2 ,Δ max =21.7pts / cm 2 The index 1.15 was obtained by fitting the error-cost curve to the multi-objective genetic search.

[0108] The search module is used to set the set of geometric center coordinates of all sensitive units as... Define the centroid error functional Measure the local error of the sensor array relative to the center of the target region; define the manufacturing budget functional. Quantify the manufacturing cost and process complexity of the layout; adjust the target weight ratio The array geometry optimization problem is solved using a hybrid ant colony-SGD search algorithm. The outer layer generates 8872 candidate clusters through 60 iterations of the ant colony algorithm, while the inner layer fine-tunes each cluster for 120 rounds using the Adam-SGD algorithm to obtain the coordinates of sensitive cells that balance accuracy and cost. Nu represents the total number of sensitive cells, and rp0 represents the geometric center coordinates of the p0th sensitive cell.

[0109] Based on the search results, a sheet sensor was deployed to cover an area of ​​113×87mm to obtain a sensor array.

[0110] In this embodiment, the sensitive unit is encapsulated in a SEBS-PDMS bilayer elastomer using an AgNW-h-BN interlaced grid framework, and Cu / Ni / Au multilayer films are deposited sequentially on the upper and lower electrodes; the surface is treated with an 80nm fluorosilane self-assembled monolayer (SAM) to suppress sweating.

[0111] The piezoresistive response was obtained under triaxial synchronous loading tests at temperatures ranging from -30℃ to 70℃ and RH from 0% to 95%. Where Ω0 is the no-load resistance value, (varsigma1,varsigma2,varsigma3)=(3.02×10- 3 (0.86, 1.04), after 1,000,000 cycles of compression at 1200 kPa, the resistance drift is within 3.1% and the hysteresis error is no more than 1.9%.

[0112] In this embodiment, the flexible sensor array adopts a seven-layer structure design: the substrate is a 20μm thick liquid crystal polymer film, and a conductive mesh is formed by micro-nano imprinting; the sensing unit adopts a carbon nanotube / polydimethylsiloxane composite material, whose piezoresistive characteristics satisfy: Where α and η are material parameters, determined through calibration experiments as α = 1.47 × 10⁻⁶. -3 η = 0.83. The array contains 96 sensing units, and a differentiated density layout D(x,y) = D0[1+κ1I(x,y)] is used in the heel, metatarsal head and arch regions, where I(x,y) is the regional importance index, κ1 is the density adjustment coefficient, and D0 is the original density layout.

[0113] In this embodiment, the model input data comes from high-precision foot imaging data (such as CT scans, MRI, etc.) to obtain the fine structure of the subject's foot bones and soft tissues. Three-dimensional geometry and anatomical parameters of each component of the foot (including bones, tendons, fat pads, etc.), such as arch height, bone density, and soft tissue thickness, are extracted through image segmentation to achieve digital modeling of the foot's anatomical structure. These biological structural parameters serve as model input, providing a high-resolution anatomical basis for subsequent biomechanical analysis. The model is typically built at a resolution of 0.5 mm to ensure the capture of the influence of fine structures on the biomechanical response.

[0114] Foot plantar region segmentation method: Based on foot anatomy and load characteristics, the foot plantar contact surface is divided into several biomechanical functional zones, such as the heel area, arch area (the medial longitudinal arch area is usually not directly stressed), forefoot metatarsal area, and phalangeal area. These can be further subdivided into multiple discrete contact units to improve spatial resolution. For example, some studies have set multiple discrete contact points on each foot plantar surface to simulate distributed loads. The region segmentation considers differences in bone distribution and plantar tissue thickness, giving each region independently defined material properties and geometric features in the model, thus accurately reflecting the stress characteristics of different parts.

[0115] The principles for determining mechanical sensitivity factors are as follows: Mechanical sensitivity factors are introduced into the model to measure the intensity of local tissue response to mechanical loads. These factors are determined according to the following principles: mechanical indicators closely related to the occurrence of stress-induced injuries are selected, such as peak stress / strain, stress gradient, energy density per unit area, load application time, and loading rate. These indicators are considered precursors to micro-injuries in soft tissue or bone tissue. For example, excessively high peak pressure and pressure-time integrals in the metatarsal region often indicate an increased risk of stress-induced bone injury. The model couples biological structures and mechanical fields at multiple levels, linking foot anatomical parameters with stress distribution to screen for mechanical factors that can sensitively reflect injury risk. Preferred mechanical sensitivity factors should be robust across different individuals and can be experimentally calibrated to correspond to early injury markers (such as bone marrow edema detected by MRI).

[0116] Differences in Regional Coefficient Matrices: The model establishes separate mechanical response coefficient matrices for each divided plantar region to describe the relationship between external load and internal stress / strain in that region. Due to differences in bone structure and soft tissue properties across different regions, the coefficient matrices exhibit significant differences. For example, the heel region, with its thick fat pad and large bone (calcaneus), has a strong capacity for buffering and absorbing loads, and its corresponding coefficient matrix shows a lower proportion of external pressure being converted into bone stress. Conversely, the forefoot metatarsal head region has a higher structural rigidity and a relatively concentrated contact area, potentially generating higher bone stress per unit pressure, thus resulting in a larger sensitivity coefficient matrix coefficient in this region. The model reflects these differences by employing different material models and elastic moduli for each region.

[0117] In numerical implementation, the stiffness and damping matrices of the finite element mesh vary in different regions, forming a spatially non-uniform coefficient distribution matrix. This ensures that the model can accurately capture the differences in the sensitivity of different plantar regions to pressure stimuli. A multi-layered modeling approach is introduced: the plantar surface is considered to be composed of a multi-layered structure of bone, soft tissue, and insole. Deformation of the upper layer is transmitted to the lower layer through a coupling matrix, thus more realistically simulating the mechanical behavior of the plantar surface. This multi-layered coupling model, while maintaining spatial resolution, improves the fitting and prediction performance of complex mechanical fields, enabling detailed prediction of stress hotspots within the foot and providing a solid foundation for stress injury risk assessment.

[0118] Compared to traditional simplified models, the plantar biomechanical sensitive model of this invention significantly improves spatial resolution and predictive performance. On one hand, by utilizing high-precision individual anatomical data and regional subdivision, the accuracy of pressure distribution simulation and the ability to depict local details are enhanced. On the other hand, the introduction of a biomechanical sensitivity factor matrix enables the model to detect potential overload signs in various foot structures in advance, thus providing a basis for early injury prediction. Overall, this model provides a high-fidelity biomechanical foundation for subsequent pressure reconstruction and risk analysis.

[0119] Flexible array data acquisition: A flexible pressure sensor array embedded in the insole is used to acquire pressure data from a limited number of discrete points on the sole of the foot in real time. After calibration, the sensors can provide pressure values ​​at key points within the dynamic gait cycle. Due to the limited and discrete distribution of sensing points in wearable insole systems, the directly obtained raw pressure maps often have low spatial resolution and cannot fully reflect the continuous pressure field of the sole. Therefore, models and algorithms are needed to reconstruct the discrete measurements. For data preprocessing, the sensor data is first filtered and gait phase segmented to ensure the accuracy and synchronization of the input data (e.g., distinguishing pressure at different stages within the support phase).

[0120] Interpolation and reconstruction algorithms are integrated: Multi-scale radial basis function (RBF) interpolation or Kriging interpolation algorithms are employed to reconstruct the pressure values ​​at discrete sensing points into a continuous pressure distribution surface across the entire plantar region. Specifically, prior structural information provided by the plantar biomechanical sensitive model is used to determine the overall shape of the pressure distribution at a coarse scale, while at a fine scale, radial basis functions are used to smooth the neighborhood, capturing local pressure peaks and gradient changes. This multi-scale approach ensures that both global trends are depicted without losing local details. Furthermore, the finite element projection method can be combined: the collected discrete pressure is applied as boundary conditions to the foot finite element model, solving the contact problem to obtain the pressure distribution on the plantar mesh. This is equivalent to using the physical constraints of the model for inversion calculations, ensuring that the reconstruction results conform to the laws of foot mechanics. To improve computational efficiency, the influence coefficient matrix of the model can be pre-calculated, and an approximate distribution can be quickly obtained using linear projection, followed by iterative refinement.

[0121] Integration of physical constraints and prior knowledge: Prior knowledge of foot structure and mechanics is integrated during the reconstruction process to constrain and correct the interpolation results, thereby improving the reliability and fidelity of the reconstruction.

[0122] Total force conservation: The pressure distribution after correction interpolation should meet the force balance requirement, that is, the integral of the entire plantar pressure should be close to the body weight (or the load of the current support box). This can be achieved by adding Lagrange multipliers or normalization factors during interpolation to ensure that the total force of the reconstructed pressure map is consistent with the measured value.

[0123] Zero-pressure constraint in the arch region: Based on anatomical knowledge, the normal arch region (medial longitudinal arch) is usually suspended and bears no pressure, or very little pressure, when standing. Therefore, zero-pressure or minimum-pressure constraints are set for the arch region in the interpolation algorithm to prevent the interpolation from generating unrealistic "false pressure". For special cases such as flat feet, the constraint strength in this region can be adjusted according to the differences in foot type.

[0124] Smooth boundary transition: In contact / non-contact boundary areas such as the edge of the toe, physical constraints are introduced to smooth the change of pressure gradient, avoiding sharp edge fluctuations caused by sparse sensors, which do not conform to real contact behavior.

[0125] Structural knowledge and sensor data fusion: Local sensitivity coefficients provided by a plantar biomechanical sensitivity model are used to correct biases in the interpolation results. For example, if the model indicates that a certain area is particularly sensitive to pressure changes, the weight of the interpolation for that area can be increased, making the reconstruction result closer to reality in that area. Conversely, for areas that the model indicates are not sensitive or have high hardness (such as below the calcaneus), their pressure values ​​can be appropriately regularized and smoothed. Through this model-data fusion, the reconstruction algorithm not only relies on mathematical interpolation but also incorporates foot biomechanics, ensuring that the estimated pressure distribution respects both the measured data and the physiological structural characteristics and real physical laws of the foot. The result is a high-fidelity, high-resolution pressure distribution estimation map across the entire plantar surface, clearly showing the trajectory of the pressure center, the magnitude and distribution range of pressure peaks in each area, providing a reliable data foundation for subsequent feature extraction and risk analysis.

[0126] Multidimensional plantar pressure feature selection: Multiple characteristic parameters characterizing gait loading are extracted from the reconstructed plantar pressure distribution and its temporal evolution to quantify its correlation with stress-related injuries. Feature selection covers multiple aspects including the time domain, frequency domain, and morphology.

[0127] Temporal characteristics include peak pressure per step, pressure-time integral (PTI), load escalation rate, contact time, and gait cycle. These characteristics directly reflect the cumulative load and impact intensity on the tissues. Peak pressure and longer pressure duration are often associated with increased injury risk. In particular, repeated high pressure in certain areas (such as the bases of the 2nd to 5th metatarsals and the heads of the first and third metatarsals in the forefoot) has been proven to be an important early warning indicator of lower limb overload.

[0128] Frequency domain characteristics: Spectral analysis of stress time series is performed to extract parameters such as gait frequency components, spectral energy distribution, and FFT coefficient phase angle. Frequency domain characteristics can reveal the patterns of gait load variation over time; for example, high-frequency components reflect sudden impact forces, while low-frequency components reflect steady-state load patterns. Studies have found that some frequency domain indicators (such as the angular characteristics of FFT coefficients in stress time series) have significant effects in predicting lower limb overuse injuries.

[0129] Morphological features: Foot load morphological features are extracted based on pressure distribution patterns, such as the offset of the center of pressure (COP) trajectory, the eccentricity of the pressure distribution, the ratio of forefoot to hindfoot pressure, and the ratio of medial to lateral pressure. These features reflect gait posture and force patterns (e.g., excessive pronation leading to medial pressure concentration). Abnormal distribution patterns may correspond to poor force application habits and potential injury risks.

[0130] Complexity and stability characteristics: Quantitative indicators such as entropy and complex networks, including sample entropy and multi-scale entropy, are introduced to measure the complexity and stability of gait stress signals. Higher entropy values ​​indicate large fluctuations and unpredictability in gait stress patterns, which may imply unstable foot load distribution and increase the risk of chronic injury.

[0131] Injury Labeling and Data Labeling: Establishing a mapping relationship between plantar pressure characteristics and stress injuries using clinical diagnostic data. Specifically, this involves collecting plantar pressure data and corresponding injury labels (e.g., whether tibial stress fracture or plantar fasciitis has occurred) from athletes. Injury label determination relies on the medical gold standard, such as MRI imaging results. MRI can sensitively detect early signs of bone stress injury (e.g., bone marrow edema), providing an indication before fractures occur. The injury level determined by MRI is used as a label and correlated with previously extracted pressure characteristics. One approach is to use statistical analysis and feature importance ranking to screen for a subset of features significantly associated with injury occurrence (e.g., excessive forefoot pressure, medial-lateral pressure imbalance, etc.). For example, studies have found that localized high pressure in the forefoot and load characteristics considering both medial and lateral aspects are strong indicators for predicting lower limb overload injuries.

[0132] Modeling Methods: A quantitative model of plantar pressure and stress injury risk is constructed by combining machine learning and deep learning methods. Traditional methods may consider ensemble learning, such as random forests and gradient boosting trees, to train multidimensional features and output a score or probability of injury risk. These methods improve model stability by integrating multiple decision trees and can provide an intuitive interpretation of feature importance. Simultaneously, a dual-tower structure model based on deep learning is introduced to fuse multimodal features: two sub-networks (towers) are constructed, such as one tower accepting plantar pressure distribution image sequences as input (using CNN / LSTM to extract spatial-temporal features), and the other tower accepting extracted statistical feature vectors or other biological signals (such as gait frequency, acceleration, and biomechanical parameters) as input.

[0133] The two towers extract high-level feature representations separately, then merge them in a fusion layer, and output the predicted value of damage risk through a fully connected network. This dual-tower structure ensures that different modalities independently learn their optimal representations while achieving information complementarity, thus improving the model's discriminative performance. During training, a multi-task learning strategy can be employed, simultaneously incorporating damage classification (presence or absence of damage) and damage severity regression into the optimization objective, thus approximating the medical significance of the model's output.

[0134] Multimodal feature fusion and generalization enhancement: To improve the robustness and generalization ability of the model, the model fuses data from multiple sources as input, including:

[0135] Dynamic plantar pressure characteristics (from insole sensors) and foot anatomy / structural characteristics (from imaging, such as arch height and bone mineral density).

[0136] Combining actively monitored kinematic data (such as gait speed, cadence, and lower limb joint angles) with stress characteristics forms a more comprehensive characterization of the injury mechanism.

[0137] By integrating environmental and training load data (such as running mileage and terrain hardness) with biomechanical characteristics, the impact of external load factors on injury risk can be quantified.

[0138] During the fusion process, attention mechanisms or weighting coefficients can be employed to allow the model to automatically learn the importance weights of different modal features. The fused model exhibits better stability when facing different individuals and scenarios, and is less susceptible to fluctuations in single features. For example, although complex features such as frequency domain coefficients and AR autoregressive coefficients are prominent in the model, their interaction with easily interpretable spatiotemporal features can prevent the model's decisions from over-reliance on difficult-to-explain factors, thereby improving the reliability of clinical applications. In summary, the constructed quantitative model can map multidimensional plantar pressure features to the probability or risk level of stress injury, achieving an objective quantitative assessment of injury risk.

[0139] Real-time gait assessment per gait cycle: Utilizing the aforementioned quantitative model, real-time risk assessment is performed for each gait cycle. When the wearer completes a gait cycle (from heel strike to forefoot lift-off), the system instantly calculates the plantar pressure characteristics of that cycle and inputs them into the model, outputting the risk value or probability of stress injury caused by that gait. Through professional calibration, this risk output can correspond to the probability of injury occurrence; for example, an output of 5% indicates low risk, while an output of 30% or higher falls into the high-risk range. Real-time assessment provides immediate feedback after each step: if a poor posture in a step leads to a sharp increase in risk, the system can immediately alert the user via wearable devices (such as mobile phones or wristbands), reminding them to adjust their gait posture or reduce impact. This progressive, real-time monitoring ensures a timely warning mechanism, allowing users to take corrective measures before danger accumulates.

[0140] Multi-period cumulative load monitoring: In addition to single-step assessment, the system tracks the cumulative effect of chronic load over longer time scales. Using a sliding time window (e.g., one week or one month), it aggregates risk indicators from multiple gait cycles to assess chronic stress load levels. Specifically, it calculates the cumulative or weighted average of risk indices over a period to reflect the cumulative stress borne by the tissue. For example, moderate risk loads lasting several days may gradually accumulate to form a high-risk state. The system defines load accumulation indices for different time scales, distinguishing between acute overload (short-term high-risk peak) and chronic fatigue (long-term moderate-risk accumulation). Through trend analysis, once a multi-period risk index shows an upward trend or even exceeds the safety threshold, the system issues an early warning, indicating a potential chronic stress injury hazard. This monitoring method is similar to a fatigue injury accumulation model, integrating the instantaneous risk of a single step over time to help identify behavioral patterns that, while not extreme in a single gait, may cause injury in the long run.

[0141] Model Output Probability Calibration: To ensure the interpretability and decision-making value of the model's predicted risk probabilities, probability calibration of the model output is necessary. Based on the actual injury incidence rates in historical datasets, the numerical probabilities in the model output are adjusted to align with the actual probability distribution. For example, Platt scaling or confidence-based calibration curves can be used to map the original model output to the corrected risk probabilities. After calibration, the probability thresholds corresponding to different risk levels become more explicit and reliable (e.g., high risk is defined as a probability >30% corresponding to the observed high injury incidence range). Simultaneously, evaluation metrics such as the model's AUC and Brier score are used to measure the discriminative power and accuracy of the calibrated model. The calibrated probability output can be used for personalized health decision-making: for example, a gait risk probability of 20% corresponds to a moderate risk in the general population, allowing users to refer to this probability to assess the preventative measures they need to take.

[0142] Risk level classification and feedback: Based on the risk probability or score output by the model, stress injury risks are classified into different levels (e.g., low risk, medium risk, high risk, very high risk), establishing a graded early warning mechanism. Each level corresponds to a specific feedback and intervention strategy.

[0143] Low risk (e.g., probability <10%): Provide daily encouraging feedback to remind you to maintain good habits, while the system continues routine monitoring.

[0144] Medium risk (e.g., probability 10%–30%): Provide warning feedback, advising users to pay attention to training volume and posture, and insert popular science information such as "Recently heavy load, pay attention to rest and stretching".

[0145] High risk (e.g., probability > 30%): Provides a strong alert, advising users to take immediate action to reduce load, such as decreasing the number of high-intensity training days, changing footwear, or strengthening specific muscle groups to improve force distribution. If the risk continues to rise, it is recommended to seek medical attention promptly.

[0146] Extremely high risk (if defined): The system may directly recommend stopping the relevant activity and taking protective measures to prevent serious injury.

[0147] Feedback is pushed in real time via mobile app or wearable device, in the form of text alerts, graphic risk reports, and, when necessary, sound or vibration reminders. Importantly, the feedback is personalized, providing tailored suggestions based on the user's foot type and historical data. For example, if someone with high arches is at risk of medial superior lateral arch compression, the feedback will specifically suggest intervention measures such as using medial support insoles.

[0148] Closed-loop feedback and dynamic model updates: This system is designed with a closed-loop structure of monitoring, evaluation, and intervention. After outputting risk assessments and intervention suggestions, the system continues to monitor the user's response to the suggestions and subsequent gait changes, feeding new stress data back into the model. On one hand, if a user takes intervention (such as adjusting running posture or resting for several days), the system will observe whether the risk index decreases, thus verifying the intervention effect; if the risk decreases, the feedback loop converges positively, and the model can also record this successful case for continuous learning. On the other hand, if a user does not adopt the suggestions and later experiences a stress injury event, the system can use this result for online retraining of the model, adjusting model parameters to improve the prediction accuracy for similar individuals. Dynamic model updates can also adapt to changes in stress distribution caused by changes in individual status (such as changes in weight or increased muscle strength), achieving personalized evolution by periodically retraining with new data or fine-tuning model weights. The entire closed-loop mechanism ensures that the model is not used statically but is continuously improved through user interaction, and its advancement lies in its ability to adaptively improve predictive performance and reliability. For example, newly collected large-sample gait data can be used to expand the training set, and incremental learning can be used to update the model, enabling it to generalize to a wider range of people and sports scenarios. Ultimately, the system forms a closed-loop intelligent management system integrating real-time monitoring, advanced analysis, risk assessment, feedback intervention, and model evolution, demonstrating outstanding technological advancement and innovation in the early prediction and prevention of stress injuries.

[0149] The beneficial effects of the above technical solution are as follows: In terms of measurement accuracy, through optimized sensor layout and hybrid reconstruction algorithm, the error of plantar pressure distribution reconstruction is controlled within 4.5%, and the spatial resolution of key areas reaches 2mm; in terms of system performance, a dedicated signal processing chip is used to achieve parallel computing, with a processing delay of less than 30ms and a sampling rate of up to 500Hz; in terms of wearing comfort, the multi-layer flexible structure design makes the insole thickness less than 3mm and the bending stiffness less than 0.1N·m; in terms of clinical application value, it can achieve an early warning of stress injury 21 days in the athlete population, with an accuracy rate of 89.7%.

[0150] In the prevention of diabetic foot, the system's detection of early neuropathy achieved a sensitivity of 93.4% and a specificity of 87.6%. In training scenarios, real-time feedback reduced peak plantar pressure by 40% and the incidence of stress fractures by 68%. In the field of rehabilitation medicine, the correlation coefficient between the system's evaluation results and standard clinical scales remained above 0.85.

[0151] In this invention, after obtaining a multimodal warning, the result needs to be transmitted back to the user. However, transmission delays may occur during this transmission process, preventing the user from promptly and effectively understanding any anomalies. Therefore, the invention also includes:

[0152] Lock the output and receiver of the multimodal warning, determine the first historical pointing set from the output to the receiver and the second historical pointing set from the receiver to the output, and determine the transmission deviation value er;

[0153] Simultaneously, the real-time bandwidth B, transmission delay Td, and packet loss rate Lr of the transmission link from the output to the receiver are collected.

[0154] Acquire physiological parameters related to multimodal early warning, including: peak pressure Pp of the plantar pressure field and injury risk level Lrs;

[0155] Calculate the overall reliability strength Rcom of the early warning transmission;

[0156]

[0157] Where Br is the preset reference bandwidth; Td_th is the transmission delay threshold; eth is the azimuth deviation threshold; k8 is the azimuth deviation adjustment coefficient; and Pef is the plantar pressure reference peak value. This is the normalized coefficient for damage risk;

[0158] If Rcom falls below the preset comprehensive reliability threshold Rcth, a hierarchical early warning and dynamic transmission protection mechanism is triggered:

[0159] Determine the type of warning for multimodal conditions:

[0160] If it is a tactile warning, then based on Calculate the tactile interval time Tvi;

[0161] For visual or auditory warnings, a 5G communication channel is selected. If the 5G signal strength is lower than the preset signal strength, the system switches to the LoRa wide-area communication model and operates according to the formula. The alert interval Trd is calculated, with Tbase as the base alert interval. At the same time, the redundancy of the warning information is adjusted according to the damage risk level Lrs.

[0162] In this embodiment, the output is a device that generates multimodal warning information, such as a signal processing terminal (a device that integrates plantar pressure data and generates warnings) that is used with flexible plantar pressure insoles.

[0163] The recipient is the user terminal that receives the warning information, such as a smartwatch or a health management app on a mobile phone.

[0164] The first historical pointer set is the set of historical records of the output party's direction of transmission to the receiver when the output party previously transmitted data to the receiver (such as the direction data of the output party transmitting signals to the receiver at different times and in different scenarios).

[0165] The second historical pointing set is a historical record of the receiver's orientation towards the output when the receiver previously received data from the output (e.g., the orientation data of the receiver receiving signals from the output in different scenarios).

[0166] The transmission deviation value er is a quantitative value of the azimuth pointing deviation calculated by comparing the statistical characteristics (such as mean and variance) of the first and second historical pointing sets (for example, the root mean square of the difference between the two sets of data, which reflects the stability deviation of the transmission azimuth).

[0167]

[0168] Where θref is the azimuth mean reference threshold; ω is the reference threshold for azimuth variance; ω6 is the variance fusion weight, with a value range of 0.3 to 0.7 determined through experimental data fitting. μ01 is the variance of the first history-pointing set; μ01 is the mean of the first history-pointing set. is the variance of the second history-pointing set; μ02 is the mean of the second history-pointing set;

[0169] The peak pressure Pp of the plantar pressure field is the pressure value (unit: kPa) at the location of maximum pressure in the plantar pressure distribution collected by the flexible plantar pressure insole, reflecting the intensity of local stress on the sole.

[0170] The injury risk level Lrs is a stress injury risk probability assessed through a deep-integrated two-layer model based on plantar pressure data, user movement habits, etc., and then the risk level is determined. For example, level 1-2 is low risk, level 3-4 is medium risk, and level 5 is high risk. That is, the value of Lrs is 1, 2, 3, 4, 5.

[0171] In this embodiment, Br is calibrated to 10Mbps, which is determined based on common communication scenarios, and Td_th is calibrated to 200ms.

[0172] In this embodiment, the value of k8 is 3, which was obtained by fitting experimental data.

[0173] In this embodiment, Pef is calibrated to 500 kPa, based on the normal plantar force range.

[0174] In this embodiment, tactile warnings include, for example, vibration of shoe insoles or vibration of watches.

[0175] In this embodiment, visual / auditory warnings are provided, such as mobile phone pop-ups, sound and light alarms, etc.

[0176] In this embodiment, Tvib is set to 10 seconds, which is a preset value, and Tbase is set to 30 seconds, which is a preset value.

[0177] The beneficial effects of the above technical solution are: by quantifying the transmission link, orientation deviation, and physiological damage risk in multiple dimensions, the early warning transmission strategy and reminder frequency are dynamically adjusted. This not only solves the problem of delayed early warning caused by transmission delay and data anomalies, but also optimizes information transmission in combination with damage risk, ensuring that users receive and understand multimodal early warnings in a timely and sufficient manner, and providing time guarantee for early intervention of stress injuries.

[0178] This invention provides a method for early prediction of stress injuries based on flexible plantar pressure insoles, such as... Figure 2 As shown, it includes:

[0179] Step 1: Control the sensor array in the flexible plantar pressure insole to collect plantar pressure signals;

[0180] Step 2: Perform signal preprocessing on the acquired plantar pressure signals, which includes motion artifact filtering, baseline correction, and temperature and humidity compensation.

[0181] Step 3: Reconstruct the continuous pressure field of the foot on the preprocessed signal. The pressure field reconstruction includes: preliminary interpolation, finite element constraint optimization and physiological and physical constraints.

[0182] Step 4: Extract 52-dimensional feature vectors from the pressure field reconstruction results and input them into the constructed deep-integrated two-layer model to quantify the damage risk assessment and output multimodal early warning.

[0183] The beneficial effects of the above technical solution are as follows: by optimizing the sensor layout and hybrid reconstruction algorithm, the error of the reconstruction of plantar pressure distribution can be controlled within 4.5%, and the spatial resolution of key areas can reach 2mm; a dedicated signal processing chip is used to realize parallel computing, with a processing delay of less than 30ms and a sampling rate of up to 500Hz; the multi-layer flexible structure design makes the insole thickness less than 3mm and the bending stiffness less than 0.1N·m, thus improving comfort.

[0184] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A system for early prediction of stress-related injuries based on flexible plantar pressure insoles, characterized in that, include: The sensing and acquisition module is used to control the sensor array in the flexible plantar pressure insole to acquire plantar pressure signals; The signal conditioning module is used to preprocess the acquired plantar pressure signals. The signal preprocessing includes motion artifact filtering, baseline correction, and temperature and humidity compensation. The data processing module is used to reconstruct the continuous pressure field of the foot on the preprocessed signal. The pressure field reconstruction includes: preliminary interpolation, finite element constraint optimization and physiological and physical constraints. The risk assessment module is used to extract 52-dimensional feature vectors from the pressure field reconstruction results and input them into the constructed deep-integrated two-layer model to quantify the damage risk assessment and output multimodal early warning.

2. The early prediction system for stress injuries based on flexible plantar pressure insoles according to claim 1, characterized in that, The sensor array is based on an elastic polymer and covers the heel, forefoot, metatarsal heads and arch area. The number of sensing points is 16 to 128. The center distance of the sensors in key areas is ≤2mm and the sampling rate is 500Hz. The collected analog signals are amplified by a preamplifier and then converted into digital signals by an ADC. Among them, the plantar pressure signal is the digital signal.

3. The early prediction system for stress injuries based on flexible plantar pressure insoles according to claim 1, characterized in that, The signal conditioning module includes: The motion artifact filtering unit is used to process the digital signal using an improved RLS filtering algorithm. The filtering formula is as follows: h[n+1]=h[n]+K[n](d[n]-h T [n]x[n]); Where h[n] is the filter coefficient vector at time n; h[n+1] is the filter coefficient vector at time n+1; K[n] is the gain vector at time n; d[n] is the desired pressure signal vector at time n; x[n] is the input signal vector at time n; h T [n] is the transpose of the filter coefficient vector at time n; The baseline correction unit is used to calculate the original sampled value of each sensing point in the k-th frame based on a multi-scale sliding window family. sliding window mean μ k According to the formula Achieve baseline drift correction; The temperature and humidity compensation unit is used to correct the sensor resistance value using a fourth-order polynomial compensation model to address temperature and humidity interference. The compensation formula is: ρ corr =ρ raw +cp1(∈)+cp2(∈ 2 )+cp3(∈ 3 )+cp4(∈ 4 ), where ∈ is the normalized temperature difference; ρ corr The resistance value after compensation; ρ raw The original resistance value is the uncompensated value; cp1, cp2, cp3, and cp4 are compensation coefficients.

4. The early prediction system for stress injuries based on flexible plantar pressure insoles according to claim 1, characterized in that, The data processing module includes: The initial interpolation unit is used to reconstruct the initial continuous pressure field using RBF interpolation based on the coordinates (xi, yi) of the preprocessed discrete sensing points and the corresponding pressure values ​​Pi. The interpolation formula is as follows: Where N is the number of sensing points; α i σ is the interpolation coefficient; σ is the region adaptive scaling parameter, σ = 2mm for the metatarsal head region and σ = 5mm for the heel region; Finite element constrained optimization elements are used to solve the objective function min based on a three-dimensional finite element model of the foot, using the initial continuous pressure field as the initial value. u AP-V 2 +γR(P), where A is the sensor sensitivity matrix; V is the vector composed of digital signals; P is the pressure field distribution; R(P) is the regularization term based on foot anatomy; and γ is the regularization parameter. The physiological and physical constraint unit is used to apply total force conservation constraints and low-pressure constraints on the arch of the foot, and outputs a smooth and continuous pressure field across the entire plantar surface.

5. The early prediction system for stress injuries based on flexible plantar pressure insoles according to claim 1, characterized in that, The risk assessment module includes: The vector extraction unit is used to extract a 52-dimensional feature vector from the pressure field reconstruction results, based on time-domain features, frequency-domain features, morphological features, and complexity features. The model analysis unit is used to input the 52-dimensional feature vector into a deep-ensemble two-layer model, output 64-dimensional high-level features based on the first layer, and process the high-level features using Soft-GradientBoost ensemble to output the damage risk probability. Where ε() is the Sigmoid function; h r0 (o) is the r0th weak classifier for analyzing high-level feature o, and R0 = 100. The first layer is a feature transformation layer, which is composed of a double hidden layer fully connected network. The second layer is a risk prediction layer, which is integrated by Soft-GradientBoost. The deep-ensemble two-layer model employs Focal-Tversky composite loss to suppress positive and negative sample imbalance, with the loss formula L being: Where γ0 is the focusing parameter, with a value of 2.0; α1 is... The weighting coefficient is 0.35; y is the damage label. Used for quantization Similarity between them; Risks are categorized into low-risk, medium-risk, and high-risk based on probability, and multimodal early warning is implemented.

6. The early prediction system for stress injuries based on flexible plantar pressure insoles according to claim 1, characterized in that, The three-dimensional finite element model of the foot is a multi-layered structure including bone, soft tissue, and insole. It is based on high-resolution CT / MRI images of the foot and uses Delaunay-Voronoi hybrid subdivision to generate ≥2000 quadrilateral elements. The elastic modulus field and Poisson's ratio field are assigned according to the differences in anatomical regions. The elastic modulus of the heel region is set to 1.2 GPa, the metatarsal head region is set to 1.8 GPa, and the soft tissue region is set to 0.3 MPa.

7. The early prediction system for stress injuries based on flexible plantar pressure insoles according to claim 1, characterized in that, Also includes: The anatomy module is used to divide the anatomical surface of the foot into 11 macro-regions and 47 secondary functional units based on the dynamically updated three-dimensional finite element stress knowledge base of the foot, and to reserve magnetic subnet interfaces at the extreme arch morphology positions. The scoring calculation module is used to calculate a composite score Ξ_k∈[0,100] within each of the secondary functional units by comprehensively considering the three-dimensional indices of local load increment, historical damage attribution rate, and manufacturing tolerance, and then map it to the sensor node surface density Δ. k : Where, Δ min =3.4pts / cm 2 ,Δ max =21.7pts / cm 2 The index 1.15 was obtained by fitting the error-cost curve to the multi-objective genetic search. The search module is used to set the set of geometric center coordinates of all sensitive units as... Define the centroid error functional Measure the local error of the sensor array relative to the center of the target region; define the manufacturing budget functional. Quantify the manufacturing cost and process complexity of the layout; adjust the target weight ratio The array geometry optimization problem is solved using a hybrid ant colony-SGD search algorithm. The outer layer generates 8872 candidate clusters through 60 iterations of the ant colony algorithm, while the inner layer fine-tunes each cluster for 120 rounds using the Adam-SGD algorithm to obtain the coordinates of sensitive cells that balance accuracy and cost. Nu represents the total number of sensitive cells, and rp0 represents the geometric center coordinates of the p0th sensitive cell. Based on the search results, a sheet sensor was deployed to cover an area of ​​113×87mm to obtain a sensor array.

8. A method for early prediction of stress injuries based on flexible plantar pressure insoles, characterized in that, include: Step 1: Control the sensor array in the flexible plantar pressure insole to collect plantar pressure signals; Step 2: Perform signal preprocessing on the acquired plantar pressure signals, which includes motion artifact filtering, baseline correction, and temperature and humidity compensation. Step 3: Reconstruct the continuous pressure field of the foot on the preprocessed signal. The pressure field reconstruction includes: preliminary interpolation, finite element constraint optimization and physiological and physical constraints. Step 4: Extract 52-dimensional feature vectors from the pressure field reconstruction results and input them into the constructed deep-integrated two-layer model to quantify the damage risk assessment and output multimodal early warning.

Images