Gait walking analysis method and device for stroke patient

Abstract

The invention provides a stroke patient gait walking analysis method and device. The stroke patient gait walking analysis method comprises the steps that pressure data and inertial data generated in the gait walking process of a patient are collected based on a plantar pressure sensing module and an inertial measurement module; performing data preprocessing on the pressure data and the inertial data; constructing a double-flow neural network to respectively extract pressure features and inertial features of the pressure data and the inertial data, performing feature fusion, and calculating step length, stride frequency, gait asymmetry index and gait stability index key indexes based on the fused features so as to realize gait evaluation; an evaluation result is transmitted to the flexible OLED touch screen, and gait parameters, rehabilitation progress, abnormal early warning and training guidance information are displayed in real time through the flexible OLED touch screen. According to the method, the end-to-end process from original sensor data to clinical application is realized, manual monitoring is avoided, the accuracy and clinical practicability of evaluation are improved, and the method is particularly suitable for family rehabilitation monitoring of stroke patients and precise rehabilitation evaluation of medical institutions.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical rehabilitation equipment and gait analysis, and particularly relates to a flexible and portable gait monitoring device and analysis method for stroke patients. Background Art

[0002] Stroke is a common central nervous system disease that can cause functional impairments such as movement, sensation, cognition, and speech in patients. According to statistics, about 2 million new stroke patients occur in China every year, and the number of existing patients exceeds 13 million. For stroke sequelae, rehabilitation training is an important means to restore motor function, especially gait rehabilitation is crucial for improving the patient's self-care ability and social participation.

[0003] Traditional gait analysis systems are mainly based on three-dimensional motion capture systems, force platforms, and electromyography devices in hospital laboratories. These devices are not only expensive and complex to operate, but also require professional personnel to operate, making it difficult to achieve daily home rehabilitation monitoring. In recent years, the development of wearable sensing technology has provided new solutions for gait monitoring, but the main challenges currently faced include: (1) it is difficult for a single sensing modality to comprehensively capture gait characteristics; (2) sensing data is vulnerable to noise interference; (3) traditional analysis methods are sensitive to data loss and anomalies; (4) there is a lack of evaluation criteria for specific gait characteristics of stroke patients.

[0004] The prior art document CN112426149A discloses a gait analysis method based on inertial sensors, but only uses a single sensing modality, and the information is incomplete. CN115944291A introduces a stroke health assessment method based on multiple inertial measurement devices, but its data processing method mainly relies on traditional algorithms and has limited processing ability for data loss problems.

[0005] Therefore, there is an urgent need to develop a gait analysis system that combines multi-source sensing technology and advanced deep learning algorithms, can be used in a home environment, has the ability to process data loss and noise, and provides accurate gait assessment and rehabilitation guidance for stroke patients. Summary of the Invention

[0006] The purpose of the present invention is to provide a gait walking analysis method and device for stroke patients, which solves the problems in the prior art such as incomplete information of a single sensing modality, weak data loss processing ability, and low gait assessment accuracy.

[0007] To achieve the above purpose, the technical solution provided by the present invention is: A gait walking analysis method for stroke patients, comprising:

[0008] Collecting pressure data and inertial data generated during the patient's gait walking based on a plantar pressure sensing module and an inertial measurement module;

[0009] Perform data preprocessing on pressure data and inertial data;

[0010] Construct a two-stream neural network to extract pressure features and inertial features from pressure data and inertial data respectively, perform feature fusion, and calculate key indicators such as step length, step frequency, gait asymmetry index, and gait stability index based on the fused features, thereby realizing gait assessment;

[0011] Transmit the evaluation results to the display device to display gait parameters, rehabilitation progress, abnormal warnings, and training guidance information in real time.

[0012] Furthermore, data preprocessing includes time-domain synchronization, frequency-domain filtering, and multi-scale feature decomposition;

[0013] In time-domain synchronization, the cubic spline interpolation method is used to solve the problems of inconsistent sampling times and missing data of different sensors;

[0014] Frequency-domain filtering is achieved through Butterworth low-pass filters and adaptive Kalman filters;

[0015] Multi-scale feature decomposition uses wavelet transform technology, combines empirical mode decomposition technology to decompose the signal into intrinsic mode functions and residual terms, calculates non-linear features, and constructs a multi-dimensional feature set.

[0016] Furthermore, first, multi-scale decomposition of the signal is realized based on discrete wavelet transform:

[0017]

[0018] Among them, x(t) is the original signal, which is a function of time t; W ψ (j,k) is the wavelet transform coefficient, representing the signal component intensity at scale j and translation k; ψ(·) is the wavelet basis function, ψ * (·) is the complex conjugate of the wavelet basis function; combining empirical mode decomposition technology, the signal x(t) is decomposed into:

[0019]

[0020] Among them, c i is the i-th intrinsic mode function, r n (t) is the residual term, representing the remaining part after decomposition, and n is the total number of mode functions; in addition, by calculating approximate entropy, a comprehensive feature set is constructed to provide a data basis, and the expression of approximate entropy is:

[0021] ApEn(m,r,N) = φ m (r) - φ m+1 (r)

[0022] Among them, ApEn(m, r, N) is the approximate entropy, m is the embedding dimension; r is the similarity threshold, used to judge whether two patterns are similar; N is the length of the time series; φ m (r) and φ m+1 respectively represent the logarithmic average values of the similarity measures in the m and m+1 dimensions.

[0023] Furthermore, a spatio-temporal fractal convolutional network ST-FractalNet is used to learn the preprocessed pressure data. For the set of pressure signals after empirical mode decomposition where, T is the sequence length, S is the number of sensors, F is the feature dimension, including all components the residual term r n (t) and the approximate entropy ApEn value of each component, c i is the i-th intrinsic mode function, r n (t) is the residual term, representing the remaining part after decomposition, n is the total number of mode functions; the basic computational unit of ST-FractalNet can be expressed as:

[0024]

[0025] Among them, represents the feature map output of the l-th layer of the fractal convolutional network; and respectively represent the convolutional weight matrix and bias vector of the l-th layer; * represents the convolutional operation; α l ∈[0,1] is the path selection parameter for adaptive learning, allowing the network to dynamically switch between feature extraction paths of different complexities;

[0026] ST-FractalNet also introduces a plantar pressure point topological perception mechanism:

[0027] A topo =softmax(φ(G adj )·W topo +b topo )

[0028] Among them, A topo is the topological attention matrix, representing the attention weight distribution for different plantar pressure points; G adj is the topological adjacency matrix predefined based on the plantar anatomical structure; φ(·) is the graph neural network operation, learning the biomechanical associations between plantar pressure points; softmax(·) is the normalized exponential function; W topo and b topo are respectively the learnable weight matrix and bias vector of the topological attention mechanism. The final pressure feature is expressed as:

[0029]

[0030] Among them, W p and b p are the learnable weight matrix and bias vector respectively.

[0031] Furthermore, the Phase-Sensitive Attention Network (PSA-Net) is used to process the inertial data. Since wavelet transform has been used for signal decomposition in the preprocessing, PSA-Net directly utilizes these decomposition results and enhances the key gait cycle features through the phase-sensitive attention mechanism:

[0032] M t = σ(W m ·[X] + b m )

[0033] Among them, X represents the set of inertial signals after empirical mode decomposition, T is the sequence length, S is the number of sensors, F is the feature dimension, including all components The residual term r n (t) and the approximate entropy ApEn value of each component, c i is the i-th intrinsic mode function, r n (t) is the residual term, representing the remaining part after decomposition, and n is the total number of mode functions; PSA-Net innovatively introduces a phase tracking module:

[0034]

[0035] Among them, ψ t represents the phase angle at time t, representing the instantaneous phase of the signal; arctan2(·) is the two-parameter arctangent function, calculating the angle of a point in the complex plane; the instantaneous phase of the signal is tracked through Hilbert transform, which is used to identify specific events in the gait cycle;

[0036]

[0037] Among them, f(·) is the phase encoding function, mapping the phase angle to a high-dimensional feature space; W ψ and b ψ are the learnable weight matrix and bias vector respectively; the final inertial feature representation is:

[0038]

[0039] Among them, ⊙ represents element-wise multiplication, enhancing the feature representation related to gait events through the phase attention mechanism; BiLSTM(·) is the bidirectional long short-term memory network.

[0040] Furthermore, an adaptive quantum state fusion mechanism is adopted to achieve feature fusion. Inspired by quantum information theory, different modality features are regarded as quantum states for fusion:

[0041]

[0042] Among them, ρ P and ρ I respectively represent the density matrix representations of pressure data and inertial data; and are respectively the transpose matrices of the pressure feature representation F p and the inertial feature representation F I ; tr(·) is the trace operation of the matrix; the degree of mutual correlation between modalities is evaluated by the quantum entanglement measure E(ρ P ,ρ I ):

[0043]

[0044] Among them, λ i is the singular value after the singular value decomposition of the joint representation matrix of pressure data and inertial data; the fusion weights are adaptively adjusted using the entanglement measure:

[0045]

[0046] λ I = 1 - λ p

[0047] Among them, λ P and λ I respectively represent the fusion weights of pressure and inertial data; exp(·) represents the exponential function; W λp and W λI are respectively weight matrices, b λP and b λI are bias matrices; diag(ρ P ) and diag(ρ I ) are respectively the diagonal elements of the density matrices of pressure and inertial data; the final fusion feature F fused is obtained through quantum state combination:

[0048]

[0049] Among them, W f is the weight matrix, b f is the bias matrix.

[0050] Furthermore, based on the quantum state fusion features, key indicators such as step length, step frequency, gait asymmetry index, and gait stability index are output; the quantum state fusion feature F fusedAs the input, perform a non - linear mapping through a three - layer fully connected network:

[0051] y = W3·(W2·(W1·F fused +b1)+b2)+b3

[0052] where W1, W2, W3 and b1, b2, b3 are weight matrices and bias vectors, and y is the predicted value of m gait parameters;

[0053] The calculation of the stride utilizes the time correlation of quantum state characteristics:

[0054]

[0055] where v(t) is the walking speed, t1 and t2 are the start and end times of a gait cycle, and the trace matrix Tr(·) is the sum of the diagonal elements of the matrix, is the stride observation operator;

[0056] The step frequency is based on the periodic characteristics of the quantum state:

[0057]

[0058] where T is the gait cycle time, is the step frequency observation operator;

[0059] Utilize the entanglement property of the quantum state to predict the gait cycle times T L and T R , and then calculate the gait asymmetry index:

[0060]

[0061] where, is the asymmetry observation operator; for each gait parameter y i , calculate its coefficient of variation CV i in N gait cycles, using the statistical properties of the quantum state:

[0062]

[0063] where y i,j is the i - th gait parameter in the j - th gait cycle, represents the average value of the gait parameter, and are the observation operators of the parameter mean and standard deviation respectively; The gait stability index stability is calculated using the overall stability characteristics of the quantum state:

[0064]

[0065] Among them, k is the number of parameters used to calculate stability, is the stability observation operator.

[0066] Furthermore, the plantar pressure sensing module is installed in a customized insole, and 5 key measurement areas are set on the plantar surface: the inner side of the forefoot, the heel, the entire forefoot, the toe area, and the outer side of the forefoot;

[0067] The pressure sensing module is a multifunctional flexible piezoelectric sensor based on the self-assembly of the MOF / MXene interface, including:

[0068] A gradient distribution of ZIF-8 intercalation space structure is formed between MXene nanosheets;

[0069] The surface of ZIF-8 is modified with carboxyl and amino bifunctional groups to enhance the hydrogen bond network and coordination with the -OH, -F, -O functional groups on the MXene surface;

[0070] The matrix material is formed by blending polyvinyl alcohol PVA and polyacrylonitrile PAN.

[0071] Furthermore, the preparation of the multifunctional flexible piezoelectric sensor includes:

[0072] Synthesis of ZIF-8: Weigh 2-methylimidazole and dissolve it in methanol. Separately, take zinc nitrate hexahydrate Zn(NO3)2·6H2O and dissolve it in methanol. Slowly pour the zinc salt solution into the 2-methylimidazole solution, stir at room temperature and then let it stand for 24 hours. Collect the formed white precipitate by centrifugation, wash it successively with methanol, and dry it under vacuum to obtain white ZIF-8 nanocrystals;

[0073] Preparation of MXene: Place Ti3AlC2 MAX phase powder in a Teflon container containing 40% HF, stir magnetically at room temperature, collect the product by centrifugation, wash it repeatedly with deionized water until pH > 6, and freeze-dry it to obtain lamellar Ti3C2Tx MXene;

[0074] Preparation of the polymer matrix: Weigh polyacrylonitrile and polyvinyl alcohol according to a mass ratio of 3:1, dissolve them in N,N-dimethylformamide, and prepare a mixed solution with a mass fraction of 10%. After magnetic stirring, a bubble-free transparent solution is obtained.

[0075] The present invention also provides a gait walking analysis device for stroke patients, including: a plantar pressure sensing module, an inertial measurement module, a data preprocessing module, a deep learning module, a wearable display unit, and a cloud data management platform;

[0076] The plantar pressure sensing module is fixed on the patient's plantar surface and is used to collect the pressure data generated during the patient's gait walking;

[0077] The inertial measurement module is worn on the affected ankle joint and is used to collect inertial data when the patient walks;

[0078] The input end of the data preprocessing module is respectively connected to the output ends of the plantar pressure sensing module and the inertial measurement module, and is used for missing data processing, data synchronization, noise reduction filtering and time-domain feature extraction of the original data;

[0079] The input end of the deep learning module is connected to the output end of the data preprocessing module, and is used to construct a two-stream neural network, extract the temporal correlation features between the plantar pressure distribution and ankle joint movement, and realize gait evaluation;

[0080] The wearable display unit is connected to the output end of the deep learning module, and is used to display gait parameters, abnormal warnings and rehabilitation progress in real time;

[0081] The cloud data management platform is connected to other modules, and is used to store the patient's historical data, manage the rehabilitation plan and remotely monitor the patient's gait changes.

[0082] Compared with the prior art, the present invention has the following beneficial effects:

[0083] 1) Multi-source sensing fusion: Combining plantar pressure and inertial measurement data to comprehensively capture gait features and improve information integrity;

[0084] 2) Deep learning empowerment: Using a deep learning model designed specifically for time-series data to achieve end-to-end learning from raw data to gait evaluation;

[0085] 3) Strong noise suppression ability: A multi-level noise processing mechanism, including both traditional wavelet filtering and the anti-noise learning ability of neural networks;

[0086] 4) Robustness to missing data: Even when part of the gait data is missing, the system can still perform accurate gait evaluation, improving the clinical application value;

[0087] 5) Real-time evaluation and feedback: Through the wearable display unit and the cloud data management platform, real-time display of gait monitoring results and long-term data tracking are realized. Brief Description of the Drawings

[0088] Figure 1 is a schematic diagram of the overall system architecture of the present invention;

[0089] Figure 2 is a schematic diagram of the layout of the plantar pressure sensing module and the inertial measurement module of the present invention;

[0090] Figure 3 is a data preprocessing flowchart of the present invention;

[0091] Figure 4 It is the structural diagram of the deep learning analysis module of the present invention;

[0092] Figure 5 It is the flowchart of a gait walking analysis method for stroke patients in the embodiment of the present invention. Detailed implementation manners

[0093] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0094] Embodiment 1: Overall system architecture

[0095] As Figure 1 shown, a flexible and portable gait monitoring device for stroke patients of the present invention includes a plantar pressure sensing module 10, an inertial measurement module 20, a data preprocessing module 30, a deep learning module 40, a wearable display unit 50, and a cloud data management platform 60.

[0096] The plantar pressure sensing module 10 is installed in a customized insole and includes a flexible piezoresistive sensing array. As Figure 2 shown, 5 key measurement areas are arranged on the sole: the medial part of the forefoot 11, the heel 12, the entire forefoot 13, the toe area 14, and the lateral part of the forefoot 15. Each area uses a high-precision piezoresistive sensing element with a sensitivity of 0.1 N and a sampling frequency of 100 Hz. This module is connected to the data preprocessing module 30 through low-power Bluetooth.

[0097] The inertial measurement module 20 adopts a miniaturized design and is worn at the patient's ankle joint. It includes an MPU-9250 nine-axis sensor, integrating a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer. The acceleration range is set to ±16 g, the angular velocity range is ±2000° / s, the sampling frequency is 100 Hz, and the data is transmitted through low-power Bluetooth.

[0098] The data preprocessing module 30 is responsible for receiving and preprocessing the original sensing data. As Figure 3 shown, it mainly consists of three parts: a time-domain synchronization unit 31, a frequency-domain filtering unit 32, and a multi-scale feature decomposition unit 33. The data preprocessing module is deployed on an edge computing node and uses an NVIDIA Jetson AGX Orin processor, equipped with 16 GB of LPDDR5 memory and 256 GB of NVMe SSD storage.

[0099] The time-domain synchronization unit 31 uses the cubic spline interpolation method to simultaneously solve the problems of inconsistent sampling times and missing data of different sensors. This method constructs a piecewise cubic polynomial function:

[0100] S i (x) = ai (x - x i ) 3 +b i (x - x i ) 2 +c i (x - x i ) + d i , x ∈ [x i , x i+1 )

[0101] Satisfy the following boundary conditions:

[0102] S i (x i ) = y i , S i (x i+1 ) = y i+1

[0103] S' i (x i+1 ) = S' i+1 (x i+1 ), S'' i (x i+1 ) = S'' i+1 (x i+1 )

[0104] By solving the linear equations, determine the coefficients a i , b i , c i , d i , and then realize the resampling and interpolation of multi - sensor data on a unified time basis. This method not only ensures that the interpolation results have continuous first - order and second - order derivatives at the nodes, ensuring data smoothness, but also can effectively handle problems such as inconsistent sampling frequencies, timestamp offsets, and data missing. For long - term data missing, the system also combines the slope information of adjacent data points, through natural boundary conditions (S''(x0) = S''(x n ) = 0) or complete boundary conditions, to optimize the interpolation effect and ensure that the interpolation results are consistent with the actual data trend.

[0105] The frequency - domain filtering unit 32 is responsible for eliminating various noise interferences and adopts a multi - level noise suppression strategy. The amplitude - frequency characteristic of the Butterworth low - pass filter is:

[0106]

[0107] where ω c is the cut - off frequency and n = 4 is the filter order. For wavelet denoising, the soft - threshold function is adopted:

[0108]

[0109] where ω is the wavelet coefficient, λ is the threshold, usually taking σ is the standard deviation of the noise, and N is the length of the signal. In addition, the system also integrates an adaptive Kalman filter, and its state update equation is:

[0110]

[0111] where, K k is the Kalman gain, z k is the observed value, achieving precise filtering of different types of noise.

[0112] The multi-scale feature decomposition 33 relies on a multi-dimensional analysis framework to decompose the key features of the sensing data. This unit realizes the multi-scale decomposition of the signal based on the discrete wavelet transform:

[0113]

[0114] where x(t) is the original signal, which is a function of time t; W ψ (j,k) is the wavelet transform coefficient, representing the signal component intensity at scale j and translation k; ψ(·) is the wavelet basis function, ψ * (·) is the complex conjugate of the wavelet basis function; combining the empirical mode decomposition technique, the signal x(t) is decomposed into:

[0115]

[0116] where, c i is the i-th intrinsic mode function, r n (t) is the residual term, representing the remaining part after decomposition, and n is the total number of mode functions. In addition, a comprehensive feature set is constructed by calculating a series of statistical features, frequency domain features, and non-linear features such as approximate entropy, providing a reliable data basis for subsequent pattern recognition and condition monitoring. The expression of approximate entropy is:

[0117] ApEn(m,r,N) = φ m (r) - φ m+1 (r)

[0118] where, ApEn(m,r,N) is the approximate entropy, m is the embedding dimension; r is the similarity threshold, used to judge whether two patterns are similar; N is the length of the time series; φ m (r) and φ m+1 respectively represent the logarithmic average values of the similarity measures in the m and m + 1 dimensions.

[0119] The deep learning module 40 is the core of the system, such as Figure 4As shown, it includes a pressure data processing stream 41 and an inertial data processing stream 42, as well as a feature fusion unit 43 and a gait evaluation unit 44. The deep learning module is also deployed on the edge computing node.

[0120] The pressure data processing stream 41 employs a Spatio-Temporal Fractal Convolutional Network (ST-FractalNet). This network constructs a recursive convolutional structure based on the principles of fractal geometry for the set of pressure signals after empirical mode decomposition where T is the sequence length, S is the number of sensors, F is the feature dimension, including all components the residual term r n (t) and the approximate entropy ApEn value of each component. The basic computational unit of ST-FractalNet can be expressed as:

[0121]

[0122] where represents the feature map output of the l-th layer of the fractal convolutional network; and represent the convolutional weight matrix and bias vector of the l-th layer respectively; * represents the convolution operation; α l ∈[0,1] is a path selection parameter for adaptive learning, allowing the network to dynamically switch between feature extraction paths of different complexities.

[0123] ST-FractalNet also introduces a plantar pressure point topology awareness mechanism:

[0124] A topo = softmax(φ(G adj )·W topo + b topo )

[0125] where A topo is the topological attention matrix, representing the attention weight distribution for different plantar pressure points; G adj is a pre-defined topological adjacency matrix based on the plantar anatomical structure, φ(·) is a graph neural network operation for learning the biomechanical associations between plantar pressure points; softmax(·) is the normalization exponential function; W topo and b topo are the learnable weight matrix and bias vector of the topological attention mechanism. The final pressure feature can be expressed as:

[0126]

[0127] where W p and bp They are the learnable weight matrix and the bias vector respectively.

[0128] This feature representation integrates plantar anatomical structure information, enhances the model's understanding of biomechanical associations, and improves the system's generalization ability.

[0129] The inertial data processing stream 42 adopts a Phase-sensitive Attention Network (PSA-Net). Considering that the feature extraction unit has used wavelet transform for signal decomposition, PSA-Net directly utilizes these decomposition results and enhances the key gait cycle features through a phase-sensitive attention mechanism:

[0130] M t = σ(W m · [X] + b m )

[0131] Where, represents the set of inertial signals after empirical mode decomposition, T is the sequence length, S is the number of sensors, F is the feature dimension, including all components The residual term r n (t) and the approximate entropy ApEn value of each component. PSA-Net introduces a phase tracking module:

[0132]

[0133] Where, ψ t represents the phase angle at time t, representing the instantaneous phase of the signal; arctan2(·) is a two-parameter arctangent function that calculates the angle of a point in the complex plane; this module tracks the instantaneous phase of the signal through the Hilbert transform and is used to identify specific events in the gait cycle (such as heel strike, toe off, etc.).

[0134]

[0135] Where, f(·) is a phase encoding function that maps the phase angle to a high-dimensional feature space; W ψ and b ψ are the learnable weight matrix and the bias vector respectively; the final inertial feature representation is:

[0136]

[0137] Where, ⊙ represents element-wise multiplication, and enhances the feature representation related to gait events through the phase attention mechanism; BiLSTM(·) is a bidirectional long short-term memory network.

[0138] The feature fusion unit 43 adopts the Adaptive Quantum State Fusion (AQSF) mechanism. Inspired by quantum information theory, it fuses different modality features by treating them as quantum states:

[0139]

[0140] Among them, ρ P and ρ I respectively represent the density matrix representations of pressure data and inertial data; and are respectively the transpose matrices of the pressure feature representation F p and the inertial feature representation F I . tr(·) is the trace operation of the matrix. This representation method can retain the quantum coherence between features and capture richer internal structure information of modalities than traditional feature connection methods.

[0141] The degree of mutual correlation between modalities is evaluated by the quantum entanglement metric E(ρ P ,ρ I ):

[0142]

[0143] Among them, λ i is the singular value after the singular value decomposition of the joint representation matrix of pressure data and inertial data; the system adaptively adjusts the fusion weights using this metric:

[0144]

[0145] λ I = 1 - λ p

[0146] Among them, λ P and λ I respectively represent the fusion weights of pressure and inertial data; exp(·) represents the exponential function; W λp and W λI are respectively weight matrices, b λP and b λI are bias matrices; diag(ρ P ) and diag(ρ I ) are respectively the diagonal elements of the density matrices of pressure and inertial data; the final fused feature is obtained through quantum state combination:

[0147]

[0148] Among them, W f is the weight matrix, b f is the bias matrix.

[0149] This quantum-inspired fusion method can dynamically adjust the weights of different modalities, automatically enhance the influence of reliable modalities in the case of gait abnormalities, and improve the robustness of the system.

[0150] Based on the quantum state fusion features, the gait assessment unit 44 outputs key indicators such as step length, cadence, gait asymmetry index, and gait stability index. This unit receives the fused quantum state feature vector F fused as input and performs a non-linear mapping through a three-layer fully connected network:

[0151] y = W3 · (W2 · (W1 · F fused + b1) + b2) + b3

[0152] where W1, W2, W3 and b1, b2, b3 are weight matrices and bias vectors, F fused is the fused density matrix obtained through quantum state combination, and y is the predicted value of m gait parameters.

[0153] The calculation of the step length stride utilizes the time correlation of quantum state features:

[0154]

[0155] where v(t) is the walking speed, t1 and t2 are the start and end times of a gait cycle, the trace matrix Tr(·) is the sum of the diagonal elements of the matrix, is the step length observation operator. The gait assessment unit predicts the step length by learning the relationship between quantum state features and the step length.

[0156] The cadence is based on the periodic characteristics of the quantum state:

[0157]

[0158] where T is the gait cycle time (in seconds), is the cadence observation operator.

[0159] The gait assessment unit predicts the gait cycle times T L and T R of the left and right feet using the quantum state entanglement property, and then calculates the gait asymmetry index:

[0160]

[0161] where, is the asymmetry observation operator. For each gait parameter y i (step length, cadence, etc.), calculate its coefficient of variation CV i in N gait cycles, and utilize the statistical properties of the quantum state:

[0162]

[0163] where y i,j is the i-th gait parameter in the j-th gait cycle, represents the average value of gait parameters, and are the observation operators of the parameter mean and standard deviation respectively. The gait stability index stability is calculated using the overall stability characteristics of the quantum state:

[0164]

[0165] where k is the number of parameters used to calculate stability, is the stability observation operator.

[0166] Through the analysis of the coherence and entanglement degree of the quantum state, the system can capture abnormal gait patterns more sensitively, especially the subtle changes that are difficult to detect by traditional methods. For each index, this module sets a healthy reference interval and a rehabilitation target threshold. For example, the normal range of the step length index is 0.9 - 1.1, and the rehabilitation target threshold is 0.95 - 1.05. It also calculates the percentage difference between the current value and the target value in real time and intuitively displays the patient's rehabilitation progress. The high-dimensional information of the quantum state fusion characteristics enables the system to more accurately predict the patient's rehabilitation trajectory and provide personalized rehabilitation suggestions.

[0167] The wearable display unit 50 adopts a flexible OLED touch screen design and can be worn on the patient's wrist or attached to the surface of the clothing. It has a size of 1.5 inches and a resolution of 240×240 pixels. The display interface includes the following functional areas:

[0168] 1) Gait parameter display area: It presents key parameters such as step length, step frequency, and gait symmetry and their healthy reference ranges in real time;

[0169] 2) Rehabilitation progress area: It shows the achievement degree of the current gait parameters and the rehabilitation target in the form of a circular progress bar and a percentage;

[0170] 3) Abnormal warning area: When an abnormal gait pattern is detected, it reminds the patient to pay attention through color coding and vibration;

[0171] 4) Training guidance area: It shows the content and key points of the rehabilitation training for the day and supports voice broadcast.

[0172] The display unit is wirelessly connected to the gait assessment unit using Bluetooth 5.2 technology with low power consumption, and the transmission delay is less than 50 ms. It is built-in with a 450 mAh lithium battery and supports 72 hours of continuous operation. It has an IP68 waterproof and dustproof rating and is suitable for daily living environments. The touch interface supports a variety of gesture operations, including tapping, swiping, and long pressing, which is convenient for elderly patients to operate. At the same time, it supports voice interaction function, and the display content can be switched or help information can be called up through simple voice commands.

[0173] The cloud data management platform 60 is a distributed system based on a microservices architecture. Its main functional modules include:

[0174] 1) Data storage module: Adopting a hierarchical storage strategy, the original data is stored in a time series database, and the processed features and evaluation results are stored in a relational database. It realizes compressed data storage and supports querying of historical data up to 5 years.

[0175] 2) Data analysis module: Integrating statistical analysis and machine learning algorithms to realize gait pattern mining, rehabilitation trend prediction, and generation of personalized training plans. It supports two modes: batch processing and stream processing to meet the needs of offline analysis and real-time monitoring.

[0176] 3) Visualization display module: Providing a multi-dimensional data visualization interface, including time series charts, etc. It supports parameter comparison, historical backtracking, and anomaly marking.

[0177] 4) Remote monitoring module: Providing a patient management interface for medical staff, supporting real-time monitoring, historical query, and remote intervention. When a continuous abnormal gait is detected, an early warning message is automatically pushed to the medical staff's terminal.

[0178] 5) Doctor-patient interaction module: Establishing a doctor-patient communication channel, supporting the formulation of rehabilitation plans, training guidance, and remote consultation. Providing a feedback mechanism to collect patients' subjective feelings and training compliance data.

[0179] The platform adopts the HTTPS encryption transmission protocol and the OAuth2.0 identity authentication mechanism to ensure the security of data transmission and storage. It is deployed in an elastic cloud service architecture, supporting high-concurrency access and dynamic expansion. It provides standardized API interfaces, is compatible with medical information systems such as hospital information systems (HIS) and electronic medical record systems (EMR), and supports multi-terminal access, including web browsers, mobile applications, and professional medical devices.

[0180] Through daily gait monitoring and evaluation, the system can track the patient's rehabilitation progress, dynamically adjust the training plan, and provide objective data support for medical staff.

[0181] Embodiment 2: Gait walking analysis method

[0182] An embodiment of the present invention also provides a gait walking analysis method for stroke patients, including the following steps:

[0183] Collect pressure data and inertial data generated during the patient's gait walking based on a plantar pressure sensing module and an inertial measurement module;

[0184] Perform data preprocessing on the pressure data and inertial data;

[0185] Construct a two-stream neural network to extract pressure features and inertial features from the pressure data and inertial data respectively, and perform feature fusion. Calculate key indicators such as step length, step frequency, gait asymmetry index, and gait stability index based on the fused features, so as to realize gait evaluation;

[0186] Transmit the evaluation results to a flexible OLED touch screen, and display gait parameters, rehabilitation progress, abnormal warnings, and training guidance information in real time through the flexible OLED touch screen.

[0187] Data preprocessing includes time-domain synchronization, frequency-domain filtering, and multi-scale feature decomposition;

[0188] In time-domain synchronization, a cubic spline interpolation method is used to solve the problems of inconsistent sampling times and data missing of different sensors;

[0189] Frequency-domain filtering is realized through a Butterworth low-pass filter and an adaptive Kalman filter;

[0190] Multi-scale feature decomposition uses wavelet transform technology, combines empirical mode decomposition technology to decompose the signal into intrinsic mode functions and residual terms, calculates non-linear features, and constructs a multi-dimensional feature set.

[0191] The specific implementations of the plantar pressure sensing module, inertial measurement module, data preprocessing, two-stream neural network, etc. are the same as those in a flexible portable stroke patient gait monitoring device, and the present invention will not describe them.

[0192] Embodiment 3: Preparation method of plantar pressure sensing module

[0193] In this embodiment, the flexible piezoresistive sensing array adopts a multifunctional flexible piezoelectric sensor based on MOF / MXene interface self-assembly, and its preparation method includes the following steps:

[0194] Synthesis of ZIF-8: Weigh 2-methylimidazole (3.24 g, 40 mmol) and dissolve it in 70 mL of methanol. Separately, weigh zinc nitrate hexahydrate (Zn(NO3)2·6H2O, 0.59 g, 2 mmol) and dissolve it in 30 mL of methanol. Slowly pour the zinc salt solution into the 2-methylimidazole solution. After stirring at room temperature for 1 hour, let it stand for 24 hours. Collect the formed white precipitate by centrifugation (8000 rpm, 10 minutes), wash it with methanol three times successively, and dry it in vacuum at 60 °C for 12 hours to obtain white ZIF-8 nanocrystals.

[0195] Preparation of MXene: Place Ti3AlC2 MAX phase powder (2 g) in a Teflon container containing 40% HF, stir magnetically at room temperature (500 rpm) for 24 hours, collect the product by centrifugation (3500 rpm, 5 minutes), wash it repeatedly with deionized water until pH > 6, and freeze-dry it for 48 hours to obtain lamellar Ti3C2Tx MXene.

[0196] Preparation of polymer matrix: Weigh polyacrylonitrile (PAN, Mw = 150,000) and polyvinyl alcohol (PVA, degree of hydrolysis > 99%, Mw = 89,000 - 98,000) according to a mass ratio of 3:1, dissolve them in N,N-dimethylformamide (DMF) to prepare a 10% mass fraction mixed solution, and stir magnetically at 50 °C (300 rpm) for 12 hours to obtain a bubble-free transparent solution.

[0197] Disperse 1 g of ZIF-8 nanocrystals in 200 mL of 0.3 mol / L acetic acid solution, and ultrasonically treat it in an ice bath (power 300 W, intermittent mode) for 30 minutes to partially expose the zinc ion active sites.

[0198] Add 0.6 g of N-(3-dimethylaminopropyl)-N'-ethylcarbodiimide hydrochloride (EDC·HCl) and 0.4 g of N-hydroxysuccinimide (NHS) to the above solution in sequence, stir at room temperature for 30 minutes, then add a mixture of 0.5 g of glycine and 0.5 g of 4-aminobenzoic acid, and raise the temperature to 60 °C for reaction for 3 hours.

[0199] After the reaction is completed, collect the product by centrifugation (12000 rpm, 15 minutes), wash it with methanol, acetone, and deionized water three times successively, and dry it in vacuum at 50 °C for 24 hours to obtain ZIF-8 modified with carboxyl and amino groups on the surface (denoted as f-ZIF-8).

[0200] Disperse 200 mg of MXene nanosheets in 100 mL of deionized water (2 mg / mL), and obtain a deep black stable dispersion by probe-type ultrasonic treatment (power 400 W, pulse mode, 2 s on / 1 s off) for 30 minutes.

[0201] Disperse 100 mg of f-ZIF-8 in 200 mL of methanol (0.5 mg / mL), and ultrasonicate for 10 minutes to form a homogeneous dispersion. Slowly add it dropwise to the MXene dispersion at a rate of 1 mL / min using a syringe pump while maintaining gentle stirring at 200 rpm.

[0202] After the dropwise addition is complete, slowly add 20 mL of ammonia water with a pH of 8.0 to the mixed system to stabilize the pH of the system at 7.5 ± 0.2, promoting the formation of coordination and hydrogen bond interactions between the surface functional groups of f-ZIF-8 and MXene.

[0203] Stir and react at room temperature for 8 hours, centrifuge (8000 rpm, 10 minutes) to collect the product, and wash it three times with a water-methanol mixed solution (1:1) and pure methanol in sequence to obtain a composite structure with a gradient distribution of f-ZIF-8 between the MXene layers (denoted as MX / ZIF).

[0204] Weigh 15 g of PAN / PVA polymer solution, add 0.75 g of MX / ZIF composite material (i.e., 5 wt%) to it in batches, ultrasonically disperse it for 30 minutes under ice bath conditions, and then magnetically stir (400 rpm) for 12 hours to obtain a gray spinning precursor solution with appropriate viscosity.

[0205] Load the mixed solution into a 10 mL syringe and use a customized electrospinning device to prepare fibers. The electrospinning process parameters are set as follows: supply voltage 15 kV, working distance 15 cm, liquid delivery rate 0.8 mL / h, ambient temperature 25 ± 2 °C, relative humidity 40 ± 5%, and collection roller rotation speed 1200 rpm.

[0206] To achieve the oriented arrangement of nanofibers, the collection roller is coated with conductive aluminum foil, and auxiliary electrodes are set on both sides of the main electric field to form a bias electric field of 13 kV to control the fiber deposition orientation.

[0207] Continue electrospinning for 24 hours to obtain a composite nanofiber membrane with a uniform thickness (about 100 ± 10 μm).

[0208] After cutting the obtained composite nanofiber membrane into the required shape, place it in a -80 °C refrigerator for pre-cooling for 2 hours, and then place it in a freeze dryer for freeze drying at a cold trap temperature of -50 °C for 48 hours to maintain the porous structure of the fiber network.

[0209] Heat the dried sample to 65 °C at a heating rate of 1 °C / min under nitrogen protection (flow rate 50 mL / min) and hold for 4 hours for thermal annealing treatment to enhance the interfacial bonding force while maintaining the structural integrity of ZIF-8.

[0210] The heat-treated composite film is cut into shapes corresponding to the 5 key plantar measurement areas shown by using precision laser cutting technology, and conductive silver paste (RS-TKA10 silver glue) is coated on the surface as the upper and lower electrodes, and dried at 120 °C for 30 minutes. Figure 2 The sensing element with electrodes is placed in the PDMS prepolymer after demoulding treatment (silicone: curing agent = 10:1) and cured at 60 °C for 2 hours to form a protective encapsulation layer with a thickness of 0.2 mm.

[0211] According to the characteristics of plantar biomechanics, the sensing units in the 5 areas are arranged according to the layout shown, and each unit is connected by a flexible printed circuit (FPC) to form an array structure.

[0212] The prepared sensing module is subjected to a standard loading test (in the range of 0 - 160 N) using a material mechanics test system (Instron 5567). The measured sensitivity is 5.9 V / N, the response time is 98 ms, and the recovery time is 55 ms, meeting the high-precision requirements of gait analysis. Figure 2 The prepared sensing module is subjected to a standard loading test (in the range of 0 - 160 N) using a material mechanics test system (Instron 5567). The measured sensitivity is 5.9 V / N, the response time is 98 ms, and the recovery time is 55 ms, meeting the high-precision requirements of gait analysis.

[0213] A 10,000-cycle pressure test (loading force 40 N, frequency 1 Hz) is carried out, and the signal intensity attenuation is less than 3.5%, proving that the sensor has excellent durability.

[0214] Force-voltage calibration curves for 20 different pressure points are established for the sensor units in the five areas respectively, and the calibration parameters are stored in the data preprocessing module to ensure that the measurement accuracy reaches 0.1 N as required by the design.

[0215] This piezoelectric sensor array is based on the ZIF-8 / MXene composite system, showing good sensitivity (5.9 V / N) at lower pressures (<20 N), and maintaining a stable linear response in the high-pressure region (20 - 160 N), with a response time of less than 100 ms, suitable for real-time monitoring applications of plantar pressure.

[0216] This piezoelectric sensor array is based on the ZIF-8 / MXene composite system, showing good sensitivity (5.9 V / N) at lower pressures (<20 N), and maintaining a stable linear response in the high-pressure region (20 - 160 N), with a response time of less than 100 ms, suitable for real-time monitoring applications of plantar pressure.

[0217] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. The patent protection scope of the present invention is subject to the claims. Those skilled in the art can make various changes, modifications, substitutions and variations within the essence of the present invention, and these changes, modifications, substitutions and variations are all within the protection scope of the present invention.

Claims

1. A gait walking analysis method for stroke patients, characterized in that, Including: Collecting pressure data and inertial data generated during the gait walking of patients based on a plantar pressure sensing module and an inertial measurement module; Performing data preprocessing on the pressure data and inertial data; Constructing a two-stream neural network to extract pressure features and inertial features of the pressure data and inertial data respectively, and performing feature fusion, and calculating key indicators such as stride length, stride frequency, gait asymmetry index, and gait stability index based on the fused features, so as to realize gait evaluation; Transmitting the evaluation results to a display device to display gait parameters, rehabilitation progress, abnormal warnings, and training guidance information in real time.

2. The gait walking analysis method for stroke patients according to claim 1, characterized in that: The data preprocessing includes time domain synchronization, frequency domain filtering, and multi-scale feature decomposition; In time domain synchronization, a cubic spline interpolation method is used to solve the problems of inconsistent sampling times and data missing of different sensors; Frequency domain filtering is achieved through a Butterworth low-pass filter and an adaptive Kalman filter; Multi-scale feature decomposition uses wavelet transform technology, combines empirical mode decomposition technology to decompose the signal into intrinsic mode functions and residual terms, calculates non-linear features, and constructs a multi-dimensional feature set.

3. A method for analyzing the gait walking of stroke patients according to claim 2, characterized in that: First, multi-scale decomposition of the signal is realized based on discrete wavelet transform: where x(t) is the original signal, which is a function of time t; W ψ (j, k) is the wavelet transform coefficient, representing the signal component intensity at scale j and translation k; ψ(·) is the wavelet basis function, and ψ * (·) is the complex conjugate of the wavelet basis function; combined with the empirical mode decomposition technique, the signal x(t) is decomposed into: Among them, c i is the i-th intrinsic mode function, and r n (t) is the residual term, representing the remaining part after decomposition, and n is the total number of mode functions; in addition, an approximate entropy is calculated to construct a comprehensive feature set for providing a data basis, and the expression of the approximate entropy is as follows: ApEn(m,r,N) = φ m (r) - φ m+1 (r) Among them, ApEn(m,r,N) is the approximate entropy, m is the embedding dimension; r is the similarity threshold, used to determine whether two patterns are similar; N is the length of the time series; φ m (r) and φ m+1 respectively represent the logarithmic average values of the similarity measures in the m and m + 1 dimensions.

4. The gait walking analysis method for stroke patients according to claim 1, characterized in that: The spatio-temporal fractal convolution network ST-FractalNet is used to learn the preprocessed pressure data for the input pressure sensor data where P represents the set of pressure signals after empirical mode decomposition, T is the sequence length, S is the number of sensors, F is the feature dimension, including all components Residual term r n (t) and the approximate entropy ApEn value of each component, c i is the i-th intrinsic mode function, r n (t) is the residual term, representing the remaining part after decomposition, and n is the total number of mode functions; The basic computational unit of ST-FractalNet can be expressed as: Among them, represents the feature map output of the l-th layer of the fractal convolutional network; and respectively represent the convolutional weight matrix and the bias vector of the l-th layer; * represents the convolution operation; α l ∈[0,1] is the path selection parameter for adaptive learning, allowing the network to dynamically switch between feature extraction paths with different complexities; ST-FractalNet also introduces a plantar pressure point topological perception mechanism: A topo = softmax(φ(G adj )·W topo + b topo ) Among them, A topo is a topological attention matrix, representing the attention weight distribution for different pressure points on the sole; G adj is a predefined topological adjacency matrix based on the sole anatomical structure; φ(·) is a graph neural network operation for learning the biomechanical associations between sole pressure points; softmax(·) is a normalized exponential function; W topo and b topo are the learnable weight matrix and bias vector of the topological attention mechanism respectively, and the final pressure feature representation is: Among them, W p and b p are the learnable weight matrix and bias vector, respectively.

5. The gait walking analysis method for stroke patients according to claim 1, characterized in that: Using a phase-sensitive attention network PSA-Net to process inertial data. Since wavelet transform has been used for signal decomposition in the preprocessing, PSA-Net directly uses these decomposition results and enhances key gait cycle features through a phase-sensitive attention mechanism: M t = σ(W m ·[X] + b m ) Among them, X represents the set of inertial signals after empirical mode decomposition, T is the sequence length, S is the number of sensors, F is the feature dimension, including all components Residual term r n (t) and the approximate entropy ApEn value of each component, c i is the i-th intrinsic mode function, r n (t) is the residual term, representing the remaining part after decomposition, and n is the total number of mode functions; PSA-Net introduces a phase tracking module: where ψ t represents the phase angle at time t and represents the instantaneous phase of the signal; arctan2(·) is a two-parameter arctangent function that calculates the angle of a point in the complex plane; the instantaneous phase of the signal is tracked through Hilbert transform and is used to identify specific events in the gait cycle; where f(·) is a phase encoding function that maps phase angles to a high-dimensional feature space; W ψ and b ψ are learnable weight matrix and bias vector respectively; the final inertial feature representation is: Among them, ⊙ represents element-wise multiplication, and the feature representation related to gait events is enhanced through the phase attention mechanism; BiLSTM(·) is a bidirectional long short-term memory network.

6. The gait walking analysis method for stroke patients according to claim 1, wherein: Using an adaptive quantum state fusion mechanism to realize feature fusion. Inspired by quantum information theory, different modal features are regarded as quantum states for fusion: where ρ P and ρ I represent the density matrix representations of the pressure data and the inertial data, respectively; and are the transposed matrices of the pressure feature representation F p and the inertial feature representation F I respectively; tr(·) is the trace operation of the matrix; the degree of mutual correlation between the modes is evaluated by the quantum entanglement metric E(ρ P , ρ I ): Among them, λ i is the singular value after the singular value decomposition of the joint representation matrix of pressure data and inertial data; the fusion weight is adaptively adjusted using entanglement measurement: λ I = 1 - λ p where, λ P and λ I represent the fusion weights of pressure and inertial data respectively; exp(·) represents the exponential function; W λp and W λI are weight matrices respectively, b λP and b λI are bias matrices; diag(ρ P ) and diag(ρ I ) are the diagonal elements of the pressure and inertial data density matrices respectively; the final fused feature F fused is obtained by quantum state combination: Among them, W f is the weight matrix, and b f is the bias matrix.

7. The gait walking analysis method for stroke patients according to claim 1, characterized in that: Output key indicators such as step length, step frequency, gait asymmetry index, and gait stability index based on the fused features of quantum states; use the fused features of quantum states F fused as the input and perform non-linear mapping through a three-layer fully connected network: y = W3 · (W2 · (W1 · F fused + b1) + b2) + b3 Among them, W1, W2, W3 and b1, b2, b3 are weight matrices and bias vectors, and y is the predicted value of m gait parameters; The calculation of the stride length stride utilizes the time correlation of quantum state features: where \(v(t)\) is the walking speed, \(t_1\) and \(t_2\) are the start and end times of a gait cycle, and the trace matrix \(\text{Tr}(\cdot)\) is the sum of the diagonal elements of the matrix, is the step length observation operator; The stride frequency is based on the periodic characteristics of the quantum state: where T is the gait cycle time, is the step frequency observation operator; Predicting the gait cycle time T of the left and right feet by utilizing the entanglement characteristics of quantum states L and T R , and then calculating the gait asymmetry index: Among them, is an asymmetry observation operator; for each gait parameter y i , calculate its coefficient of variation CV in N gait cycles i , and utilize the statistical characteristics of quantum states: where y i,j is the i-th gait parameter in the j-th gait cycle, represents the average value of gait parameters, and are the observation operators of the parameter mean and standard deviation respectively; the gait stability index stability is calculated using the overall stability characteristics of the quantum state: where k is the number of parameters used to calculate stability, is the stability observation operator.

8. The gait walking analysis method for stroke patients according to claim 1, characterized in that: The plantar pressure sensing module is installed in a customized insole, and 5 key measurement areas are arranged on the sole: the inner side of the forefoot, the heel, the entire forefoot, the toe area, and the outer side of the forefoot; The pressure sensing module is a multifunctional flexible piezoelectric sensor based on MOF / MXene interface self-assembly, including: Forming a gradient distribution of ZIF-8 intercalated space structure between MXene nanosheets; The surface of ZIF-8 is modified with carboxyl and amino bifunctional groups to enhance the hydrogen bond network and coordination with -OH, -F, -O functional groups on the MXene surface; The matrix material is formed by blending polyvinyl alcohol PVA and polyacrylonitrile PAN.

9. The gait walking analysis method for stroke patients according to claim 8, characterized in that: The preparation of the multifunctional flexible piezoelectric sensor includes: Synthesis of ZIF-8: Weigh 2-methylimidazole and dissolve it in methanol. Separately, take zinc nitrate hexahydrate Zn(NO3)2·6H2O and dissolve it in methanol. Slowly pour the zinc salt solution into the 2-methylimidazole solution, stir at room temperature and then let it stand for 24 hours. Collect the formed white precipitate by centrifugation, wash it successively with methanol, and dry it under vacuum to obtain white ZIF-8 nanocrystals; Preparation of MXene: Place Ti3AlC2 MAX phase powder in a Teflon container containing 40% HF, stir magnetically at room temperature, collect the product by centrifugation, wash it repeatedly with deionized water until pH > 6, and freeze-dry to obtain lamellar Ti3C2Tx MXene; Preparation of polymer matrix: Weigh polyacrylonitrile and polyvinyl alcohol according to a mass ratio of 3:1, dissolve them in N,N-dimethylformamide, and prepare a mixed solution with a mass fraction of 10%. After magnetic stirring, a bubble-free transparent solution is obtained.

10. A gait walking analysis device for stroke patients, characterized in that, Including: A plantar pressure sensing module, an inertial measurement module, a data preprocessing module, a deep learning module, a wearable display unit, and a cloud data management platform; The plantar pressure sensing module is fixed on the patient's sole and is used to collect the pressure data generated during the patient's gait walking; The inertial measurement module is worn on the patient's ankle joint and is used to collect the inertial data during the patient's walking; The input end of the data preprocessing module is respectively connected to the output ends of the plantar pressure sensing module and the inertial measurement module, and is used to perform missing data processing, data synchronization, noise reduction filtering, and time-domain feature extraction on the original data; The input end of the deep learning module is connected to the output end of the data preprocessing module, and is used to construct a two-stream neural network, extract the temporal correlation features of the plantar pressure distribution and ankle joint movement, and realize gait assessment; The wearable display unit is connected to the output end of the deep learning module and is used to display gait parameters, abnormal warnings, and rehabilitation progress in real time; The cloud data management platform is connected to other modules and is used to store the patient's historical data, manage the rehabilitation plan, and remotely monitor the patient's gait changes.

Images