Gait evaluation method and system based on human body nonlinear system analysis technology

Abstract

The invention provides a gait evaluation method and system based on a human body nonlinear system analysis technology, and the method comprises the steps: collecting a gait cycle six-channel high-precision time sequence signal outputted by a wearable inertial measurement unit, eliminating noise and gait difference through wavelet threshold denoising and Z-score standardization, extracting a chaotic feature vector composed of a Lyapunov index spectrum and a Kolmogorov entropy value, and carrying out the recognition of a gait signal, a wavelet neural network is adopted to realize nonlinear mapping of chaotic features and phase-space reconstruction parameters, and a self-adaptive feedback mechanism is introduced to dynamically optimize modeling parameters, so that the accuracy and personalized matching capability of gait pattern recognition and stability evaluation are effectively improved; quantitative characterization of the gait chaos level and real-time online model optimization can be achieved, and high-robustness support is provided for rehabilitation training and exercise aided decision making.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of chaotic feature analysis and individualized dynamic modeling of gait signals, and in particular to a gait evaluation method and system of human nonlinear system analysis technology. BACKGROUND

[0002] Human gait analysis as an important research direction in the field of sports medicine, rehabilitation engineering and intelligent health has been widely concerned in recent years. With the development of wearable sensor technology (especially inertial measurement unit IMU), high-precision gait data acquisition has become the basis for realizing individualized gait evaluation. The mainstream gait analysis system generally uses nonlinear dynamics method to reveal the implicit time sequence structure and complex dynamic behavior of gait signals through phase space reconstruction technology. Specifically, the accurate setting of time delay embedding method and related phase space reconstruction parameters (delay time τ and embedding dimension d) plays a decisive role in gait pattern recognition and stability analysis; The current common gait dynamic analysis process includes: collecting high-precision three-dimensional acceleration and angular velocity time series, using signal preprocessing (such as wavelet denoising and standardization), and then combining nonlinear dynamics feature extraction. Chaotic characteristic analysis, especially the calculation of maximum Lyapunov exponent and Kolmogorov entropy, is widely used to evaluate the complexity and stability of human gait system. As the core modeling method, the parameters of phase space reconstruction are usually set based on fixed schemes, such as autocorrelation function method, Cao criterion, etc. However, the existing technology basically relies on static parameter selection, which is difficult to adapt to the inherent differences between subjects, environmental disturbances and dynamic changes in the movement process; In recent years, some researches have explored multi-modal data fusion, dynamic feature driven parameter optimization and other schemes in order to improve the individualization and intelligentization level of gait analysis. For example, visual, electromyographic or pressure sensor data are fused with inertial data for multi-modal gait feature extraction; or the modeling parameters are adaptively adjusted according to the time-varying characteristics. However, these schemes are complex in design, and there are certain limitations in system implementation and real-time performance. Moreover, when relying only on single modal input, the accuracy of individualized modeling is still limited.

[0003] At present, the gait phase space reconstruction technology has the following typical problems: (1) Fixed reconstruction parameters. Most systems use preset delay time and embedding dimension, which cannot be dynamically adjusted according to individual signal specificity, resulting in insufficient modeling results for group differences and abnormal gait response ability; (2) Lack of adaptive mechanism. The parameter setting process rarely has real-time adaptive adjustment capability for gait chaotic characteristics (such as Lyapunov exponent and Kolmogorov entropy), which is easily disturbed by non-structural noise, environmental changes and other factors, affecting the accuracy and stability of identification; (3) Multimodal fusion threshold is high, although fusion of multi-source data can improve nonlinear modeling performance, but in practical application, due to hardware, synchronization and algorithm complexity, it is difficult to be widely deployed, and personalized analysis is limited under low modal data conditions; (4) Dynamic modeling is limited, and existing parameter optimization methods are mostly based on static features or initial modeling values for adjustment, which is difficult to respond to the dynamic changes of gait signal chaos with time in the movement process, causing the evaluation accuracy to fluctuate with the change of the state of the subject. SUMMARY

[0004] The present application provides a gait evaluation method of human nonlinear system analysis technology to solve the above technical problems.

[0005] The technical scheme of the present application is as follows: a gait evaluation method of human nonlinear system analysis technology, comprising: S1: collecting single modal high-precision time series signals in the gait cycle of the subject, the signals being derived from the acceleration and angular velocity output of the inertial measurement unit in three-dimensional space, and recording the corresponding time stamp and sampling frequency parameters; S2: wavelet threshold denoising and Z-score standardization processing are performed on the collected time series signals to eliminate the influence of sensor noise and individual gait rhythm difference on subsequent chaos feature recognition; S3: based on the preprocessed signal, the Lyapunov exponent spectrum and Kolmogorov entropy value are calculated as quantitative indicators of individual gait chaos characteristics, and a chaos intensity feature vector is output; S4: a nonlinear mapping model between the chaos feature index and the phase space reconstruction parameter is constructed, wavelet neural network is used to dynamically decouple the delay time τ and the embedding dimension d for training, and an intelligent prediction mechanism of individual chaos level and optimal parameter combination is established; S5: an adaptive feedback updating mechanism is designed to monitor the drift trend of the chaos feature vector in real time, and if it is judged that the deviation of the current chaos intensity from the initial modeling value exceeds the preset threshold, the dynamic reconfiguration of the phase space parameter is triggered; S6: the optimized phase space reconstruction parameter is input into the nonlinear modeling module, individualized gait pattern recognition and stability evaluation are performed based on the reconstructed phase space trajectory, and the evaluation result is output for rehabilitation training or motion assistance decision; S7: a plurality of chaos feature-parameter mapping model libraries are established for the gait chaos feature distribution of different subjects, and the closest initial modeling parameter group is automatically matched according to the individual chaos intensity feature during the initial use; S8: a dynamic adaptive evaluation mechanism is introduced in the nonlinear modeling process, and if it is detected that the modeling error exceeds the set tolerance, the chaos feature re-recognition and parameter re-optimization process is triggered to realize online adaptive updating of the model.

[0006] The application further provides a gait evaluation system of a human nonlinear system analysis technology, which adopts the gait evaluation method of the human nonlinear system analysis technology to perform gait evaluation.

[0007] The gait evaluation method of the human nonlinear system analysis technology provided by the application has the following beneficial effects: (1) The individual difference and dynamic change of the gait signal are directly reflected in the selection process of the phase space reconstruction parameter by the multidimensional quantification and standardization of the chaos index (maximum Lyapunov index, Kolmogorov entropy, etc.), and on the basis of the nonlinear modeling of the wavelet neural network, the delay time and embedding dimension most suitable for the current state of each subject and each gait cycle can be automatically output, the dynamic adaptation of the model to the individual signal complexity and dynamic characteristics is realized, and the fine-grained individualization of gait modeling is greatly improved. (2) The parameter decoupling selection is realized by the nonlinear mapping deep learning of the chaos index and the parameter, the blindness and local optimal trap of the parameter configuration are effectively avoided, the nonlinear reconstruction error is significantly reduced in the training set and actual application, the matching degree of the embedding parameter selection and the dynamic divergence characteristics is greatly improved, and the system stability and robustness are enhanced. (3) A series of adaptive feedback mechanisms such as sliding window chaos feature monitoring, Hurst index drift calculation and sliding mean square error threshold triggering are introduced, so that the parameter adjustment is changed from passive offline setting to active online response. Once important indicators such as the current gait chaos intensity deviate, the system can automatically re-identify the chaos characteristics, call the best parameter mapping model and complete the parameter reconfiguration in time. This feature not only improves the real-time adaptability to gait behavior fluctuations and environmental disturbances, but also significantly prolongs the model life cycle and does not require frequent manual intervention, and the modeling accuracy fluctuation can be stabilized within a set threshold (such as ±3%). (4) The application emphasizes relying only on a single IMU mode, realizes intelligent matching of neural network parameters driven by chaos, and avoids the technical path of needing to fuse multiple source sensors or complex human posture estimation. This greatly reduces the system hardware cost and backend computing resource consumption, enables the technology to be lightweight integrated in wearable, mobile or remote rehabilitation application scenarios, greatly improves the applicability and promotion value, meets the actual clinical and large-scale evaluation deployment requirements. (5) The application can realize efficient adaptation to unseen gait individuals and abnormal states by establishing a chaos feature-parameter multi-model database of the historical sample library and combining the initial individual automatic precise matching scheme. Meanwhile, the sliding window and dynamic error tolerance strategy are introduced, so that the online model can adaptively track long-period dynamic changes, and still maintain high modeling accuracy and system stability in the face of long-term wearing and large sample population environments. Attached Figure Description

[0008] Fig. 1 This is a flowchart of a gait evaluation method for a human nonlinear system analysis technique according to the present invention; Fig. 2 This is a sub-flowchart of a gait evaluation method for a human nonlinear system analysis technique according to the present invention; Fig. 3 This is another sub-flowchart of the gait evaluation method of the human nonlinear system analysis technology of the present invention. Detailed Implementation

[0009] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0010] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0011] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having,” etc., specify the presence of the stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0012] Please see Figs. 1-3 As shown, a gait assessment method based on human nonlinear system analysis technology includes: S1: Acquire high-precision time-series signals of a single mode within the subject's gait cycle. The signals originate from the acceleration and angular velocity outputs of the inertial measurement unit in three-dimensional space, and record the corresponding timestamps and sampling frequency parameters. S2: Perform wavelet threshold denoising and Z-score normalization on the acquired time series signals to eliminate the influence of sensor noise and individual gait rhythm differences on subsequent chaotic feature recognition. S3: Based on the preprocessed signal, calculate its Lyapunov exponent spectrum and Kolmogorov entropy value as quantitative indicators characterizing the chaotic characteristics of individual gait, and output the chaotic intensity feature vector. S4: Constructing the nonlinear mapping model between the chaotic feature index and the phase space reconstruction parameter, using the wavelet neural network to dynamically decouple and train the delay time τ and the embedding dimension d, and establishing an intelligent prediction mechanism of the individual chaotic level and the optimal parameter combination; S5: Designing an adaptive feedback updating mechanism to monitor the drift trend of the chaotic feature vector in real time, and if the deviation between the current chaotic intensity and the initial modeling value exceeds the preset threshold, triggering the dynamic reconfiguration of the phase space parameter; S6: Inputting the optimized phase space reconstruction parameter into the nonlinear modeling module, and based on the reconstructed phase space trajectory, performing individualized gait pattern recognition and stability evaluation, and outputting the evaluation results for rehabilitation training or motion assistance decision; S7: Establishing a multi-group chaotic feature-parameter mapping model library for different subjects' gait chaotic feature distribution, and automatically matching the closest initial modeling parameter group according to the individual chaotic intensity characteristics during the first use; S8: Introducing a dynamic adaptability evaluation mechanism in the nonlinear modeling process, and if the modeling error exceeds the set tolerance, triggering the chaotic feature re-identification and parameter re-optimization process to realize online adaptive updating of the model.

[0013] The step S1: collecting single modal high-precision time series signals in the gait cycle of the subject, the signal is derived from the acceleration and angular velocity output of the inertial measurement unit in three-dimensional space, and the corresponding time stamp and sampling frequency parameters are recorded. Specifically, it includes: S1.1: In the data acquisition stage of the gait evaluation system, the acceleration and angular velocity signals of the subject's single lower limb in the gait cycle are obtained based on the wearable inertial measurement unit (IMU), the acceleration signal includes X, Y, Z three orthogonal direction linear acceleration components, and the angular velocity signal includes angular rate components around X, Y, Z axes, to form six-channel original signal input; The data input source in the acquisition stage is the original sensing signal output by the inertial measurement unit IMU worn on the subject's single lower limb, and the execution period covers the complete gait cycle; Three-axis acceleration sensors are used to collect X, Y, Z three orthogonal direction linear acceleration components, and the parameter setting is based on the definition of the inertial reference frame direction, realizing the full range capture of the lower limb linear motion state; Further, the three-axis gyroscope sensor is used to collect the angular velocity components around the X, Y, Z axes, and the parameter setting is based on the right-hand rule coordinate system, realizing the complete measurement of the corresponding lower limb rotation motion state, and obtaining a six-channel original signal matrix ; Further, by analyzing the Performing an ordered arrangement operation of the channel dimension, the acceleration components are placed in channel 1 to channel 3, and the angular velocity components are placed in channel 4 to channel 6, to generate a multi-channel time series data structure conforming to the input specification of the subsequent algorithm ; Further, by synchronously triggering with the gait cycle start and end point detection module, the collection window is strictly aligned to the gait starting time, ensuring the cross-cycle coherence in the subsequent chaos feature analysis; Through the above channel collection and structured reconstruction processing, the six-dimensional motion information of the subject's lower limbs in the complete gait cycle is effectively converted into multi-channel time series input data, realizing high-fidelity preservation of the original information and standardized output of the modeling basic data; For example, in a certain clinical rehabilitation training, the IMU is fixed to the front side of the right shank, the sampling range of the three-axis acceleration sensor is set to , the resolution is , the sampling range of the three-axis gyroscope is set to , the resolution is , and the sampling frequency is . Each sampling point obtained by collection is a vector containing channels, which is output as a matrix after data structure arrangement, with a matrix size of , covering a gait cycle signal of . After time stamp synchronization and noise suppression processing, the signal matrix retains the complete frequency spectrum component of the original motion feature, with a signal-to-noise ratio improved by about compared with the untreated, providing high-quality input for chaos feature calculation; S1.2: Time stamp synchronization markers are added to the collected six-channel raw signals, and high-precision real-time clock (RTC) modules are used to attach absolute time stamp information to each sampling point to ensure time consistency in the subsequent signal processing and modeling process; The multi-channel time series matrix output in the collection stage is synchronized and marked in the absolute time domain using the time reference signal (parameters: synchronization accuracy , sampling frequency ) of the high-precision real-time clock (RTC) module; Further, through the clock interrupt service program (parameters: hardware trigger frequency consistent with the IMU sampling frequency), the current value of the RTC counter register is read and attached to the corresponding sampling vector , realizing point-by-point correspondence between the sampling data and the global time reference; Further, the time stamp alignment algorithm (parameters: maximum allowed time deviation Cross-channel time series registration is performed on the six-channel data to correct the sampling phase offset caused by hardware delay or buffer differences in different channels, ensuring strict consistency of each channel in the time domain. Furthermore, monotonicity and equal interval checks are performed on the timestamp sequence after synchronization, based on differential calculation. Confirm that the deviation between the sampling interval and the preset sampling frequency does not exceed the allowable range; Furthermore, by correcting the interpolation method (parameter: linear interpolation or spline interpolation strategy is optional), the detected micro-scale gaps or overlaps are compensated to generate a six-channel synchronization signal matrix on a unified absolute time axis; By using the above timestamp synchronization marking and alignment processing methods, the original multi-channel gait signals are transformed into standardized acquisition data with a unified time reference and consistent phase across channels, thereby ensuring time consistency and phase accuracy of feature extraction in the subsequent modeling process. For example, in a rehabilitation monitoring application, the IMU output matrix The sampling frequency is set to The RTC module provides The absolute time base for resolution. When acquiring a 1-second signal, for each sampling point... Additional timestamp range covers from arrive Time difference between adjacent sampling points The mean is , compared with theoretical value The maximum deviation in seconds is less than Cross-channel time offsets were all corrected by the alignment algorithm. Within this range, the output matrix is ​​directly used for recording the sampling frequency parameters of S1.3, effectively eliminating the interference of phase difference on subsequent phase space reconstruction and chaotic feature analysis; S1.3: Record the sampling frequency parameter set by the inertial measurement unit during data acquisition. The sampling frequency shall not be lower than 200Hz to meet the Nyquist sampling theorem's requirement for complete capture of the frequency domain features of the gait signal, and use this frequency parameter as the time domain analysis basis for subsequent signal reconstruction and feature extraction. S1.4: Perform coordinate system normalization transformation on the acquired acceleration and angular velocity signals, converting the original local coordinate system data output by the sensor into a vector representation under the global reference coordinate system, so as to eliminate the influence of orientation uncertainty caused by the difference in sensor wearing angle; S1.5: The normalized acceleration and angular velocity signals are encapsulated into structured data frame formats, which include timestamps, sampling frequencies, channel identifiers and signal amplitude arrays, and output to the next processing module for wavelet thresholding and standardization operations.

[0014] The step S2: wavelet threshold denoising and Z-score standardization processing is performed on the collected time series signal to eliminate the influence of sensor noise and individual gait rhythm difference on subsequent chaotic feature recognition. Specifically, it includes: S2.1: Select wavelet basis function and set decomposition level for the three-dimensional acceleration and angular velocity time series signals collected in S1, perform discrete wavelet transform based on Mallat algorithm to obtain approximate coefficients and detail coefficients at multiple scales; For the three-dimensional acceleration and angular velocity time series signals collected in S1 and completed time stamp synchronization and structured processing, wavelet basis function selection algorithm (parameter setting according to: signal frequency spectrum distribution characteristics and target noise frequency range) is used to determine the optimal wavelet basis type, considering the resolution requirements of time domain and frequency domain; Further, through the decomposition level setting method (parameter: the number of layers L satisfies The frequency band width is not less than the target minimum information frequency), the upper limit of wavelet multi-scale decomposition and the lower limit of resolution are matched to ensure that the low frequency trend and high frequency details in the gait signal are effectively described; Further, through Mallat decomposition algorithm (parameter: orthogonal wavelet filter bank length and transition band steepness match the actual sampling frequency ), the multi-channel signal is iteratively decomposed to obtain the approximate coefficient and detail coefficient component set at each scale; Further, through the cross-channel synchronous decomposition strategy (parameter: the same decomposition level L and wavelet basis type are used for each channel), the phase and amplitude correspondence of different motion components at multiple scales is realized, avoiding feature mismatch caused by asynchronous decomposition; Through the above wavelet basis selection and decomposition level configuration method, the original gait signal is converted into a coefficient matrix distributed in the low-frequency approximate sub-band and multi-level high-frequency detail sub-band, laying the multi-scale data foundation for subsequent threshold denoising and standardization processing, and realizing effective separation of noise and signal components in the scale domain; For example, in a rehabilitation evaluation scene, for three-dimensional acceleration and angular velocity signals with a sampling frequency of Hz, based on frequency spectrum analysis, it is determined that the signal main frequency energy is concentrated between 0.5 Hz and 40 Hz, and the noise is mainly concentrated in the frequency band higher than 50 Hz, and Daubechies-8 wavelet basis (db8) is used for decomposition. According to the frequency band division formula With L=5 layers, the lowest approximation sub-band covers the 0-4Hz range, fully encompassing the fundamental frequency component of the gait cycle. The Mallat algorithm is applied to decompose the six-channel signal layer by layer, obtaining one set of approximation coefficients and five sets of detail coefficients for each channel, while maintaining consistency in decomposition parameters across channels. In the resulting scale-domain coefficient matrix, the approximation coefficients completely preserve the overall trend of the gait, while the detail coefficients significantly enhance the instantaneous impact information and foot strike event characteristics in sub-bands above the third layer, meeting the input requirements of the subsequent soft-threshold denoising strategy in S2.2. Actual measurements show that its high-frequency noise energy is reduced by approximately 60% compared to the original signal, effectively suppressing high-frequency noise from the sensor while preserving signal structure information. S2.2: Based on the wavelet detail coefficients obtained in S2.1, the wavelet threshold of each layer of signal is calculated using the universal threshold method, and soft thresholding is performed on the detail coefficients to suppress non-stationary noise components, thereby obtaining the set of denoised wavelet coefficients. For the multi-scale detail coefficient matrix obtained by S2.1 decomposition, the wavelet threshold of each decomposition layer is calculated by the general threshold method (parameter: the threshold is calculated based on the variance of the detail coefficients at each scale) to achieve adaptive estimation of noise energy at each scale. Furthermore, the first is calculated using the following formula. Common threshold for layer detail factor : in, For the first Standard deviation of layer detail factor This is the noise scaling factor. This represents the total number of signal sampling points for this channel. Furthermore, the median absolute deviation (MAD) method is used to estimate the noise standard deviation. The calculation formula is: in For the first Layer detail coefficient vector; Furthermore, in obtaining the threshold values ​​for each layer... Then, a soft thresholding function is applied to the detail coefficients to smoothly reduce high-amplitude noise spikes and suppress low-amplitude random fluctuations. Furthermore, soft thresholding is performed independently on the detail coefficients of all decomposition layers according to the above formula, while keeping the approximation coefficients unchanged, so as to ensure that the low-frequency skeleton structure information of the signal is completely preserved. Through the general threshold calculation and soft thresholding described above, the high-frequency components mainly caused by sensor noise in the multi-scale detail coefficients are effectively weakened, forming a set of denoised wavelet coefficients. , lay the foundation for the inverse wavelet transform of S2.3, and achieve the technical effect of suppressing non-stationary noise components and preserving the transient characteristics of gait signals; For example, in a rehabilitation training monitoring, the sampling frequency of the three-dimensional acceleration and angular velocity signal is Hz, and the detail coefficient matrix is obtained at the first to fifth scales after decomposition by S2.1. The MAD method is used to estimate at the third scale, and the general threshold value is calculated by substituting the formula. The soft threshold value processing is performed on the scale detail coefficients, the coefficients below the threshold value are set to zero, and the coefficients above the threshold value are attenuated by the offset, and the reduction ratio is about . After processing, the multi-scale high-frequency noise energy is reduced by about , and the step frequency main energy retention rate is higher than , verifying the optimal balance effect between high-frequency noise suppression and fidelity; S2.3: Perform inverse wavelet transform on the processed wavelet coefficients in S2.2 to reconstruct the three-dimensional acceleration and angular velocity time series signal after noise reduction to obtain a clean signal that removes high-frequency noise and drift interference; S2.4: Based on the reconstructed clean signal in S2.3, the mean and standard deviation of each channel signal are calculated, and the signal is standardized in Z-score to eliminate non-target interference caused by individual gait rhythm and amplitude difference; S2.5: Perform time-frequency consistency verification on the standardized time series signal in S2.4 to detect the distribution stability of the signal in the time and frequency domain, and output the preprocessed signal meeting the requirements of chaotic characteristic analysis for subsequent steps.

[0015] The step S3: Based on the preprocessed signal, calculate its Lyapunov exponent spectrum and Kolmogorov entropy value as a quantitative index representing the chaotic characteristics of individual gait, and output the chaotic intensity feature vector. As shown in Fig. 2 , specifically includes: S3.1: Calculate the autocorrelation function of the time series using the mutual information method for the three-dimensional acceleration and angular velocity time series signal after wavelet threshold denoising and Z-score standardization processing, to determine the initial value independence and time correlation of the signal, and obtain the minimum delay time estimate value of the individual gait signal; S3.2: Based on the estimated minimum delay time, use Cao criterion to preliminarily estimate the embedding dimension of the time series, and output the candidate embedding dimension combination by analyzing the convergence of the phase space reconstruction error, providing an initial parameter space for subsequent chaotic feature recognition; Based on the minimum delay time estimate obtained from S3.1, the Cao criterion is used to make a preliminary estimate of the embedding dimension of the preprocessed three-dimensional acceleration and angular velocity time series signals in order to determine the minimum dimension range required for phase space reconstruction. The Cao criterion algorithm is adopted (parameter: minimum delay time). Maximum embedding dimension limit Threshold criteria This allows for trajectory distribution consistency analysis across reconstructed phase spaces for each candidate embedding dimension, and the calculation of the mean-nearest neighbor ratio function. ; Furthermore, the embedding dimension is calculated using the following formula. and Rate of change of mean proximity ratio between ; Furthermore, through dimension-by-dimensional recursive calculation and Sequence, monitoring when tending to stabilize and The inflection point where increasing the dimension no longer significantly decreases is used to determine the minimum effective embedding dimension. ; Furthermore, by conducting convergence analysis on the mean square error (MSE) of phase space reconstruction under different embedding dimension combinations, the following formula is used: in, For real signal points, Embedding dimension The predicted points below, The total number of samples; Furthermore, through the convergence criterion, when and The difference is less than the preset convergence threshold When this dimension is used, it is added to the candidate embedding dimension combination; By combining the above-mentioned Cao criterion with error convergence analysis, the minimum delay time obtained in the previous step is mapped to a set of optimal candidate embedding dimension combinations, thereby providing an initial parameter space for subsequent chaotic feature recognition. For example, in a gait assessment experiment, S3.1 estimates the minimum delay time. Sampling points, with a maximum embedding dimension. Threshold criterion .right Calculate sequentially from 2 to 10. and The results showed that when hour tending to stabilize and The rate of change is less than 0.02; then... =4 to 6, perform phase space MSE convergence test, when the MSE difference between is less than , take =5 as the optimal initial value, =6 as the alternative value, form the candidate embedding dimension array combination {5, 6}, which effectively guarantees the accuracy and stability of the subsequent S3.3 and S3.4 in the calculation process of chaotic features; S3.3: Based on the reconstructed phase space trajectory, the maximum Lyapunov exponent of the time series is estimated using the small data method. By linear fitting the divergence rate of adjacent orbits, one of the chaotic strength indicators of individual gait signals, the maximum Lyapunov exponent value, is obtained; Based on the minimum delay time and candidate embedding dimension array output by S3.2, the normalized preprocessed three-dimensional acceleration and angular velocity signals are reconstructed into phase space to generate high-dimensional phase space trajectory point sequences as the spatial representation basis for chaotic feature calculation; The small data method (parameters: phase space reconstruction dimension , candidate dimension, delay time determined by S3.1, neighborhood radius set to 5% of the trajectory mean square distance) is used to realize the nearest neighbor search of each reference point in the trajectory point set, and the points within the Theiler window that are too close in time are removed to avoid autocorrelation interference; Further, by the divergence rate calculation method, the Euclidean distance change between each reference point and its nearest neighbor point after step-by-step iteration is calculated, where k is the iteration step number, and the formula is: where is the neighbor distance of the ith point after k iterations, and M is the total number of reference points; Further, by taking the logarithm of the distance change, the logarithmic divergence curve is obtained, and the least squares linear fitting method (the fitting window length is determined by the Lyapunov exponent convergence section) is used to fit the slope, which is the maximum Lyapunov exponent ; Further, by statistically analyzing the values obtained by fitting different reference points, the mean and standard deviation are calculated as the maximum Lyapunov exponent feature of the time series. The mean value is used to represent the concentration trend of chaotic strength, and the standard deviation is used to represent the fluctuation amplitude of the dynamic state; By using a small data volume method and linear fitting, the phase space trajectory structure information is transformed into a numerical maximum Lyapunov exponent, thereby achieving a reliable quantitative characterization of the chaotic intensity of gait signals. For example, in a rehabilitation gait assessment scenario, the sampling frequency of the IMU worn by the subject... Hz, S3.1 estimated S3.2 gives the candidate embedding dimension. In this implementation, Reconstruct the phase space and set the neighborhood radius. The Theiler window length is 5% of the mean square distance of the trajectory. Point. Select M=300 reference points, calculate the neighbor distances of each reference point after iterative iterations, and average them to obtain... The curve varies with k. Within the iteration range of k=2 to k=20, the curve... Curve fitting slope to obtain a single reference point =0.136. Statistical analysis of 300 reference points showed a mean of 0.132 and a standard deviation of 0.011, indicating that the dynamic system of the subject's gait signal has significant chaotic characteristics, providing a highly reliable indicator of chaotic intensity for subsequent Kolmogorov entropy calculation and chaotic feature vector construction; S3.4: Based on the topological complexity of phase space trajectories, the Kolmogorov entropy value of the time series is calculated using the approximate entropy algorithm. By weighted summation of the non-uniformity of the state transition probability distribution, the chaotic disorder index of the individual gait system is obtained as a supplementary parameter for the chaotic intensity feature. Based on the maximum Lyapunov exponent result and phase space trajectory point set obtained in S3.3, the approximate entropy algorithm is adopted (parameter: embedding dimension). Take the optimal value of the output of S3.2, and the delay time. Take the estimated value from S3.1, similarity tolerance. Set to 15% of the root mean square amplitude of the trajectory, to achieve local pattern complexity quantification of the gait phase space trajectory; Furthermore, by using the sequence block generation method, the trajectory point set is divided into... 3D delayed coordinates form a vector sequence ,in Using the starting point index, global coverage of the embedding mode is achieved; Furthermore, the Chebyshev distance algorithm (parameter: distance metric norm) is used... ) Calculate any two The maximum component difference between dimensional embedding vectors is statistically less than The conditional probability is obtained from the total number of matching pairs. and take the average of the natural logarithm, get The average matching probability of the dimension , the formula is: Wherein is the number of trajectory points; Further, the embedding dimension is expanded to , repeat the above process to calculate ; Further, by the approximate entropy formula: The relative uncertainty of the state pattern sequence is quantified, wherein The higher the value, the more irregular the sequence pattern; Further, by the approximate relationship of Kolmogorov entropy, the is normalized and converted into the Kolmogorov entropy value (unit: bits / step), the formula is: And get the chaotic disorder degree index of individual gait system, which is used to supplement the dynamic divergence represented by the maximum Lyapunov exponent; Through the above approximate entropy and Kolmogorov entropy calculation processing mode, the topological complexity of the phase space trajectory is converted into comparable chaotic disorder degree value, which realizes the supplement of gait complexity characteristics in the chaotic intensity vector; Exemplarily, in a gait evaluation scene of a subject, the known collected signal is processed by S3.1 and S3.2 to obtain , , the sample point number , the root mean square amplitude of the trajectory is 0.85 m / s², and the tolerance . Under m-dimensional embedding, the average matching probability is , under m+1-dimensional embedding, the probability average is , the approximate entropy is calculated . The conversion Kolmogorov calculation entropy value , represents the disorder level of the gait phase space of the subject. The results show that the Kolmogorov entropy value combined with the maximum Lyapunov exponent can effectively distinguish the gait chaotic characteristics of different rehabilitation stages, and provide high stability input for the fusion of the chaotic feature vector of S3.5; S3.5: The obtained maximum Lyapunov index value is normalized and fused with the Kolmogorov entropy value to construct a chaotic intensity feature vector, and the Euclidean distance or cosine similarity is used to quantitatively represent the individual chaotic level, and the chaotic feature input vector for subsequent nonlinear mapping model training is output.

[0016] The step S4: constructing a nonlinear mapping model between the chaotic feature index and the phase space reconstruction parameter, dynamically decoupling the delay time τ and the embedding dimension d by using the wavelet neural network, and establishing an intelligent prediction mechanism of the individual chaotic level and the optimal parameter combination. As shown in the figure, it specifically includes: Fig. 3 S4.1: The chaotic intensity feature vector of the individual gait signal is normalized to eliminate the dimensional differences of the chaotic feature distribution between different subjects, and a standardized chaotic feature input vector is obtained as the training input of the wavelet neural network; The chaotic intensity feature vector output by S3.5 is The feature value range detection is performed, and the extreme value normalization method (parameters: minimum value , maximum value ) is used to realize the standardization of the chaotic feature numerical domain between different individuals; The normalization formula is The input chaotic feature vector is mapped to the closed interval [0, 1], where is the current feature value, is the normalized result; Further, the Z-score method (parameters: mean value , standard deviation ) is used to perform standardization preprocessing, and the formula is to eliminate the dimensional and amplitude differences between different feature components and enhance the comparability of the feature distribution; Further, the Pearson correlation coefficient is analyzed to eliminate the feature components with strong redundancy, and the correlation coefficient threshold is used to filter out the components highly correlated with the reserved features, and reduce the dimension redundancy of the input vector; Further, the principal component analysis (PCA) method (parameters: cumulative contribution rate threshold ) is used to solve the characteristic value and characteristic vector of the feature matrix after reducing the redundancy, and the first principal components with a cumulative contribution rate greater than are selected to form the final input vector; ​Through the above processing mode of normalization combined with standardization, the original chaotic intensity feature vector is converted into a standardized chaotic feature input vector, realizing the consistency of the feature space distribution between different subjects, and providing scale-unified and redundancy-optimized training data for the wavelet neural network input of S4.2. Exemplarily, in a gait modeling scene of once rehabilitation, the chaotic intensity feature vector output by S3.5 is , = , = . Through normalization formula calculation, the first component = 0.277, and the second component = 0.5. In the standardization stage, the mean value is calculated as , and the standard deviation is calculated as , and the standardized feature is obtained as = . Through correlation analysis, no redundant features are deleted, and PCA is projected into a two-dimensional principal component space with a cumulative contribution rate of 99.1%. The standardized chaotic feature input vector is directly used as the input node data of the wavelet neural network, effectively improving the convergence speed and prediction stability of S4.2 network training; S4.2: Based on the standardized chaotic feature input vector, a three-layer feedforward wavelet neural network structure with a multi-scale wavelet basis function is constructed, wherein the number of input layer nodes is consistent with the dimension of the chaotic feature, the Morlet wavelet basis function is used for the hidden layer nodes, and the output layer nodes correspond to the delay time τ and the embedding dimension d two parameters; Based on the standardized chaotic feature input vector, a three-layer feedforward wavelet neural network construction method (parameters: network layer number = 3, input layer node number = n , output layer node number = 2) is used to realize the establishment of a nonlinear mapping structure of chaotic features and phase space reconstruction parameters; Further, through the hidden layer node wavelet mother function selection algorithm (parameters: candidate wavelet basis set = {Morlet, Mexican Hat, Daubechies}, shape parameter value range ), the Morlet wavelet with the smallest Fourier frequency domain distribution fitting error is selected as the hidden layer activation function to ensure the high-fidelity approximation ability to the multi-scale and non-stationary mode of the chaotic feature space; Further, the hidden layer scale and translation factor initialization method (parameters: scale initialization range , translation The initialization range is based on the mean value of chaotic characteristics ± 3σ range, and a multi-scale perception window is established for the Morlet wave hidden layer to capture the mapping detail characteristics at different chaotic levels. Further, based on the forward transfer formula: wherein is the output of the jth hidden layer node, is the Morlet wave function, is the i-th component of the input vector, and are the scale and translation variables, respectively, to realize the nonlinear transformation of chaotic characteristics and multi-scale response of the hidden layer. Further, according to the output layer mapping relationship: wherein is the output weight matrix, is the hidden layer output vector, is the output threshold vector, to realize the parallel prediction output of the delay time and embedding dimension parameters. Through the above multi-scale wavelet basis neural network structure, the standardized chaotic characteristic input vector is mapped to the core parameters required for phase space reconstruction, realizing the adaptive nonlinear modeling ability for different chaotic levels. For example, in a rehabilitation gait analysis experiment, the input chaotic characteristic vector dimension , the input layer node configuration is 2, the number of hidden layer nodes is empirically set to 8, the Morlet wave shape parameter , the scale initialization range , the translation range is determined by the mean value and the standard deviation of the training sample characteristics as and . The hidden layer output is calculated using the forward transfer formula, and the weight matrix mapping output delay time prediction value is obtained. The embedding dimension prediction value is obtained. The mean square error of the real optimal configuration is , which verifies the high precision and stability of the network structure in the chaotic characteristic-parameter mapping task. S4.3: Levenberg-Marquardt optimization algorithm is used to initialize and train the wavelet neural network, and historical subject sample data set is used for supervised learning, taking the actual phase space reconstruction parameters as the expected output, and minimizing the mean square error between the network output and the real parameters. Based on the standardized chaotic feature input vector and the constructed wavelet neural network structure, the Levenberg-Marquardt (LM) optimization algorithm (parameter: damping factor) is adopted. initial value Gain coefficient minimum gradient The network weights and thresholds are initialized and trained step by step to combine the advantages of gradient descent and Newton's method, so as to achieve fast convergence and high-precision approximation. Furthermore, the network prediction output is defined through the error function construction stage. Compared with actual annotation parameters Mean squared error loss function between This is used to quantify the degree of deviation between the network output and the actual parameters; Furthermore, during the LM iterative optimization process, a weight update formula is constructed, and the damping factor is adjusted. The scaling up and down achieves a smooth transition from gradient descent to quasi-Newton update; Furthermore, in the early stages of training, maintain a relatively large Values ​​are used to ensure search stability; when iterations make When descending, shrink proportionally. To accelerate the convergence speed; when As it rises, it increases. This allows the search to revert to a stable search range, thus preventing it from getting trapped in local minima. Furthermore, using historical subject sample datasets (size: Standardized chaotic eigenvectors and their corresponding optimal Supervised learning is performed using parameters. Samples are input into the network one by one, forward propagation is performed to obtain the predicted output, the error is calculated by comparing it with the labeled parameters, and reverse weight correction is performed according to the LM weight update formula until the convergence criterion is met. Or reach the maximum number of iterations ; By using the LM optimization training strategy, the weights and thresholds of the wavelet neural network are adjusted to the optimal state, maximizing the accuracy of the nonlinear mapping between chaotic features and phase space reconstruction parameters, and achieving high-precision prediction capabilities for delay time and embedding dimension. For example, in a personalized modeling task for rehabilitation gait, the training sample size... Each sample contains a two-dimensional standardized chaotic feature vector and the corresponding true parameter combination. Equal to Different tags for the same class. LM initialization parameters are... Gain coefficient Maximum iteration Second. During the training process, initially Maintain stability to ensure search directionality, when the first iteration, the error drops significantly, and the is automatically reduced to accelerate convergence. In the second iteration, the validation set mean square error reaches and remains stable, and the training is terminated in advance. The final model's delay time prediction mean square error on the test set is only sampling points, and the embedding dimension prediction mean square error is , which verifies the convergence efficiency and prediction accuracy of the training strategy; S4.4: In the training process, a dynamic decoupling mechanism is introduced to adjust the output nodes of delay time τ and embedding dimension d independently by error back propagation, ensuring that the two achieve nonlinear decoupling in the parameter space, avoiding coupling interference between parameters, and improving the independence and accuracy of parameter prediction; S4.5: After training, the normalized chaotic features of the current subject are input into the optimized wavelet neural network, and the corresponding optimal delay time τ and embedding dimension d parameter combination is output, and the parameter combination is packaged as a phase space reconstruction parameter configuration file for the nonlinear modeling module to call and execute.

[0017] The step S5: design an adaptive feedback update mechanism to monitor the drift trend of the chaotic feature vector in real time. If it is judged that the deviation of the current chaotic intensity from the initial modeling value exceeds the preset threshold, the dynamic reconfiguration of the phase space parameters is triggered. Specifically, it includes: S5.1: Perform sliding window sampling on the chaotic feature vector to obtain the dynamic change sequence of chaotic intensity within the continuous gait cycle, which is used for subsequent drift trend analysis; S5.2: Based on the chaotic intensity sequence within the sliding window, the Hurst index calculation method is used to evaluate the time correlation of the chaotic feature, and the long-term memory index of the chaotic feature is obtained; S5.3: Calculate the Euclidean distance between the long-term memory index of the chaotic feature and the chaotic feature value at the initial modeling time to quantify the drift degree of the chaotic intensity; Based on the long-term memory index of the chaotic feature output by step S5.2 and the initial modeling chaotic feature value vector, the Euclidean distance calculation method (parameters: feature vector dimension , normalized state is the normalized feature distribution) is used to quantify the drift of the chaotic feature in the feature space; Further, by constructing the current window feature vector , which is composed of the value and the current chaotic intensity feature component, the following Euclidean distance formula is used to calculate the drift degree: ​​ wherein, is the initial modeling phase's first chaotic feature component, is the current monitoring phase's corresponding component; Further, for the different component weights of the chaotic feature, a weighted Euclidean distance algorithm (parameters: weight vector determined by feature importance ranking results) is adopted, and the optimization formula is: Through this weighted processing, the importance of the feature component that has a greater impact on drift degree determination is highlighted; Further, to reduce the impact of instantaneous fluctuations, a first-order exponential smoothing processing (parameters: smoothing coefficient ) is performed on the distance sequence , and the smoothed drift index is calculated according to the real-time and anti-interference requirements set in : This formula effectively smooths the high-frequency jitter while preserving the sensitive response of the drift trend; Further, the smoothed drift index is compared with the system-set normal fluctuation tolerance interval to form the drift degree quantization result, which provides input basis for the threshold comparison and triggering mechanism of S5.4; Through the Euclidean distance and weighted smoothing processing, the global change of the chaotic intensity feature relative to the initial state within the time window is converted into a single quantifiable value, realizing reliable drift degree determination index output; For example, in a real-time tracking test of a rehabilitation gait, the initial modeling chaotic intensity feature vector is , the current window feature vector is , and the feature dimension is adopted. The unweighted Euclidean distance calculation is: The calculation result is . In the weighted distance algorithm using weight vector , the weighted distance is reduced to , reflecting that the third component has a smaller impact on drift determination. Combined with exponential smoothing, the result is , which is within the system-set deviation tolerance . This index is further used in S5.4 to generate a determination signal for whether to perform phase space reconstruction parameter reconfiguration, realizing early response to gait modeling accuracy degradation; S5.4: Perform threshold comparison operation on the degree of chaos intensity drift, if the drift exceeds the preset deviation threshold, generate a parameter reconfiguration trigger signal; S5.5: Based on the parameter reconfiguration trigger signal, call the update mechanism in the chaos feature-parameter mapping model library to dynamically re-optimize the configuration of the current delay time τ and the embedding dimension d.

[0018] The step S6: input the optimized phase space reconstruction parameters into the nonlinear modeling module, perform individualized gait pattern recognition and stability evaluation based on the reconstructed phase space trajectory, and output the evaluation results for rehabilitation training or motion assistance decision. Specifically, it includes: S6.1: input the chaos feature-driven optimized phase space reconstruction parameters τ and d into the nonlinear modeling module as the initial configuration parameters of phase space reconstruction, to ensure that the modeling process fully adapts to the chaos characteristics of individual gait signals; S6.2: Based on the input delay time τ and embedding dimension d, perform phase space reconstruction operation on the preprocessed three-dimensional acceleration and angular velocity time series signals of inertial measurement unit, generate trajectory point set in high-dimensional phase space, to represent the topological structure of individual gait dynamic behavior; Based on the optimized delay time and embedding dimension , perform phase space reconstruction on the preprocessed three-dimensional acceleration and angular velocity signals of inertial measurement unit; Adopt time delay embedding method (parameters: delay time derived from S6.1 output, embedding dimension is the predicted optimal value), select single channel or multi-channel combination to form state vector in multi-channel signal, shift step along the time axis on the original time series to construct delay components one by one, realize dimension expansion of state space; Further, by constructing embedding matrix , each state vector is defined as: Where is the preprocessed signal sample value, is the sampling time index, which ensures that each state vector covers the phase evolution characteristics of the signal; Further, perform normalization processing (parameters: adopt zero mean unit variance normalization standard) on the embedding matrix to prevent bias in space measurement of different dimension signals and improve geometric structure preservation ability; Further, use multi-dimensional data interpolation and missing value correction processing to correct the missing values caused by and To address the boundary truncation issue, we fill in the state vectors in the insufficient length regions at the beginning and end, ensuring the integrity and continuity of the trajectory set. Furthermore, Euclidean distance is used to map each state vector to a set of trajectory points in a high-dimensional coordinate system, while maintaining the time index mapping relationship, so that subsequent nonlinear analysis methods such as recursive graphs can track dynamic evolution. By using time delay embedding and normalization mapping, single or multi-channel time series can be transformed into a high-dimensional phase space trajectory point set, thereby enabling the reconstruction and representation of the topological structure of individual gait in the dynamic state space. For example, in a gait analysis during rehabilitation training, a delay time is set. Sampling points, embedding dimension For sampling frequency of A scalar time series is formed by axial synthesis of three-dimensional acceleration signals of Hz. The dimension of the embedding matrix generated by delayed embedding is × ( From After normalization to zero mean and unit variance, the variances of each dimension are approximately equal to... After interpolating and completing the boundary data, the size of the trajectory point set is... Closed trajectories and periodic perturbation features can be observed in high-dimensional space, and a spiral attractor structure can be seen after visualization projection onto a three-dimensional subspace. This reconstruction provides complete dynamic trajectory input for subsequent recursive graph analysis, contributing to the accurate evaluation of stability and pattern recognition. S6.3: The recursive graph analysis method is used to extract dynamic features from the reconstructed phase space trajectory point set, calculate the proximity and repeatability indices between trajectory points, and obtain nonlinear dynamic characteristic parameters such as recursion rate, determinism rate and average trajectory length to quantify the stability and repeatability of individual gait cycles. S6.4: Construct individual gait pattern feature vectors based on extracted nonlinear dynamic feature parameters, and identify gait patterns through a support vector machine classifier, outputting gait cycle classification results and stability scores to achieve qualitative and quantitative assessment of individual gait status; S6.5: The gait pattern recognition results and stability scores are visualized through the assessment output module, and personalized suggestions are generated by combining them with preset rehabilitation training or exercise assistance strategies. These suggestions are then output to the terminal device for clinical or training decision-making.

[0019] Step S7: For the chaotic gait feature distribution of different subjects, establish multiple sets of chaotic feature-parameter mapping model libraries, and automatically match the closest initial modeling parameter set based on individual chaotic intensity characteristics upon first use. Specifically, this includes: S7.1: Cluster analysis is performed on the gait chaos feature data of the historical subject population, and the K-means algorithm is used to divide the feature vector space composed of Lyapunov exponent spectrum and Kolmogorov entropy value, to generate representative chaos feature prototype categories; S7.2: Based on each chaos feature prototype category, a corresponding wavelet neural network model is trained, and a nonlinear mapping relationship between the chaos feature index and the phase space reconstruction parameter (τ, d) in this category is established, to form an independent model unit in the chaos feature-parameter mapping model library; S7.3: Each independent model unit and its corresponding chaos feature prototype label are stored in the model library, and an indexing mechanism is established to realize fast retrieval and calling of the model; The input conditions are the multiple independent wavelet neural network model units and their corresponding chaos feature prototype category labels output in step S7.2; The structured data packaging method (parameters: network weight matrix of the model unit, bias vector, activation function type, chaos feature prototype label vector) is used to realize standardized storage of each wavelet neural network model and its feature description; Further, through the unique identifier generation algorithm (parameter: hash code calculated based on the chaos feature prototype category label), one-to-one mapping between the model unit and the label is realized, and repeated index conflict is avoided; Further, the high-performance key-value database is used to establish an indexing mechanism (parameter: key is unique identifier, value is model unit storage address and metadata), to realize fast retrieval of the model unit with O(1) time complexity, and support multi-thread concurrent calling; Further, through the redundant index construction method (parameter: main index is hash key, auxiliary index is Euclidean distance range index), the approximate retrieval function based on feature similarity is realized in addition to exact matching, and a candidate model unit list is generated for the upper calling strategy selection; Further, the version control mechanism (parameters: timestamp , version number , update log) is used to record the training data time and parameter update history of each model unit in the model library, to ensure the traceability of subsequent maintenance and performance backtracking analysis; Through multi-level indexing and structured storage, the results of the previous step are converted into an efficient chaos feature-parameter mapping model library that can be efficiently retrieved and called, realizing the expected technical effect of the system in running time for different subjects to quickly load the corresponding mapping model; Exemplarily, in a rehabilitation training system deployment, the input is 5 independent wavelet neural network models corresponding to 5 chaotic feature prototype categories, with an input layer node number of 2 (normalized values of Lyapunov exponent and Kolmogorov entropy), 10 Morlet wavelet basis functions in the hidden layer, and an output layer of delay time and embedding dimension Two-dimensional parameters. A unique identifier is generated by using the SHA-256 algorithm on the label array [0.85, 0.67] (normalized maximum Lyapunov exponent and Kolmogorov entropy values, respectively), and the output is a 64-bit length hash string as the primary index key. The key value is stored in a key-value database supporting memory-mapped files, and the value contains the model weight tensor (a 3.2 KB floating-point matrix), the label vector, and the version information (version number 1.0, timestamp ). The auxiliary index is filtered in the range of the Euclidean distance ≤ , and the calculation formula is: wherein , are the two-dimensional components of the target chaotic feature vector, , are the two-dimensional components of the candidate feature prototype in the index library. Based on this index configuration, the system can complete model retrieval within 1 ms, and automatically switch to the approximate matching mode when the matching fails, with a retrieval delay of 5 ms, significantly improving the response speed and individualization ability of real-time gait assessment; S7.4: When the system is first used, perform a similarity matching algorithm on the chaotic intensity feature vector of the current subject, calculate the Euclidean distance or cosine similarity between it and each chaotic feature prototype, and identify the closest chaotic feature prototype category; S7.5: According to the similarity matching result, call the corresponding model unit from the chaotic feature-parameter mapping model library, output the initial phase space reconstruction parameter combination (τ0, d0), and input it as the initial configuration parameter of the nonlinear modeling module.

[0020] The step S8: Introduce a dynamic adaptability evaluation mechanism in the nonlinear modeling process. If the modeling error is detected to exceed the set tolerance, trigger the chaotic feature re-identification and parameter re-optimization process, and realize online adaptive update of the model. Specifically, it includes: S8.1: Based on the phase space trajectory reconstruction result output by the nonlinear modeling module, calculate the residual sequence between the current modeling trajectory and the actual gait signal as the quantitative input of the modeling error, to evaluate the goodness of fit of the model; S8.2: Perform wavelet packet transform processing on the residual sequence, extract the energy distribution characteristics of each frequency band, and generate an error feature vector to represent the distribution characteristics of the modeling error in different frequency bands; S8.3: Based on the error feature vector, use a sliding window mean square error (MSE) evaluation mechanism to calculate the average modeling error in the current window, and compare it with a preset dynamic error tolerance threshold to determine whether to trigger the parameter re-optimization process; S8.4: If the modeling error exceeds the set tolerance, start the chaos feature re-identification process, and perform Lyapunov exponent spectrum and Kolmogorov entropy calculation on the current gait time sequence signal to generate an updated chaos intensity feature vector as a new input for parameter optimization; S8.5: Input the updated chaos intensity feature vector into the pre-trained wavelet neural network model to perform dynamic decoupling prediction of the delay time τ and the embedding dimension d, output the optimized phase space reconstruction parameter group, and feedback to the nonlinear modeling module for model parameter update; S8.6: After parameter update, continue to monitor the modeling error change trend, and adjust the training weight of the wavelet neural network based on error feedback to realize continuous online learning and adaptive optimization of the model parameters, and improve the modeling robustness and individual matching ability of the system in long-term use.

[0021] The application also provides a gait evaluation system for human nonlinear system analysis technology, which uses the gait evaluation method of the human nonlinear system analysis technology for gait evaluation.

[0022] So far, the technical solutions of the application have been described in combination with the preferred embodiments shown in the drawings, but those skilled in the art can easily understand that the protection scope of the application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to related technical features without departing from the principles of the application, and the technical solutions after the changes or replacements will fall within the protection scope of the application.

[0023] The above description is only the preferred embodiments of the application and is not used to limit the application; for those skilled in the art, the application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A gait assessment method of human nonlinear system analysis technology, characterized in that, The method comprises the following steps: S1: collecting single modal high-precision time series signals of a subject in a gait cycle, and recording corresponding timestamp and sampling frequency parameters; S2: preprocessing the collected time series signals to generate preprocessed time series signals; S3: calculating the Lyapunov exponent spectrum and Kolmogorov entropy value of the preprocessed time series signals, and outputting a chaotic strength feature vector; S4: constructing a nonlinear mapping model between the chaotic feature index and the phase space reconstruction parameter, dynamically decoupling the delay time and the embedding dimension by using a wavelet neural network, and establishing an intelligent prediction mechanism of the individual chaotic level and the optimal parameter combination; S5: designing an adaptive feedback updating mechanism to monitor the drift trend of the chaotic feature vector in real time, and if it is judged that the deviation of the current chaotic strength from the initial modeling value exceeds a preset threshold, triggering the dynamic reconfiguration of the phase space parameter; S6: inputting the optimized phase space reconstruction parameter into a nonlinear modeling module, performing individualized gait pattern recognition and stability evaluation based on the reconstructed phase space trajectory, and outputting the evaluation result; S7: establishing a plurality of chaotic feature-parameter mapping model libraries for different gait chaotic feature distributions of subjects, and automatically matching the closest initial modeling parameter group according to the individual chaotic strength feature during the first use; S8: introducing a dynamic adaptability evaluation mechanism in the nonlinear modeling process, and if it is detected that the modeling error exceeds the set tolerance, triggering the chaotic feature re-identification and parameter re-optimization process.

2. The gait assessment method of human nonlinear system analysis technique according to claim 1, wherein, The step S1 specifically comprises: In the data acquisition stage of the gait evaluation system, the acceleration and angular velocity signals of the subject's single lower limb in the gait cycle are obtained based on the wearable inertial measurement unit to form six-channel original signal input; The collected six-channel original signals are respectively marked with timestamp synchronization, and absolute timestamp information is added to each sampling point by using a high-precision real-time clock module; The sampling frequency parameter set by the inertial measurement unit during data acquisition is recorded; The acceleration and angular velocity signals collected are subjected to coordinate system normalization conversion to convert the local coordinate system data originally output by the sensor into vector representation in the global reference coordinate system; The normalized acceleration and angular velocity signals are respectively packaged into a structured data frame format, and the data frame contains timestamp, sampling frequency, channel identifier and signal amplitude array.

3. The gait assessment method of human nonlinear system analysis technique according to claim 2, wherein, In the step S1, the inertial measurement unit worn on the subject's lower limb synchronously obtains the acceleration and angular velocity components in X, Y and Z directions, the sampling frequency is greater than or equal to 200Hz, the data acquisition window is strictly aligned with the start and end points of gait, all original signals are synchronized by a unified clock, and the maximum time deviation is less than 0.5ms.

4. The gait assessment method of human nonlinear system analysis technique according to claim 1, wherein, The step S2 specifically comprises: The three-dimensional acceleration and angular velocity time series signals collected are subjected to wavelet basis function selection and decomposition layer setting, discrete wavelet transform is performed, and approximate coefficients and detail coefficients at multiple scales are obtained; Based on the detail coefficients, the wavelet threshold of each layer signal is calculated, and soft threshold processing is performed on the detail coefficients to obtain a set of denoised wavelet coefficients; Perform inverse wavelet transform on the processed wavelet coefficients to reconstruct the denoised three-dimensional acceleration and angular velocity time series signals, and obtain clean signals free of high-frequency noise and drift interference; Based on the clean signals, the mean and standard deviation of each channel signal are calculated, and the signal is standardized; Perform time-frequency consistency verification on the standardized time series signal to detect the distribution stability of the signal in the time and frequency domain, and output the preprocessed three-dimensional acceleration and angular velocity time series signal.

5. The gait assessment method of human nonlinear system analysis technique according to claim 4, wherein, In step S2, the signal preprocessing adopts wavelet basis function selection and multi-scale Mallat algorithm to decompose the six-channel signal. The soft threshold value is calculated for each scale of the detail coefficient using the general threshold method, and the soft threshold value is independently calculated for all detail components. Then, the noise-free signal is reconstructed by inverse transform, and each channel is standardized. The time-frequency consistency verification pass flag is output.

6. The gait assessment method of human nonlinear system analysis technique according to claim 1, wherein, The step S3 specifically includes: Using the mutual information method, the autocorrelation function of the time series is calculated based on the preprocessed three-dimensional acceleration and angular velocity time series signal, and the minimum delay time estimate value of the individual gait signal is obtained; Based on the minimum delay time estimate value, the embedding dimension of the time series is preliminarily estimated, and the convergence of the reconstructed phase space error is analyzed to output the candidate embedding dimension combination; Based on the reconstructed phase space trajectory, the maximum Lyapunov exponent of the time series is estimated, and the maximum Lyapunov exponent value of the individual gait signal is obtained by linear fitting the divergence rate of the adjacent orbits; Based on the topological structure complexity of the phase space trajectory, the Kolmogorov entropy value of the time series is calculated, and the chaotic disorder degree index of the individual gait system is obtained by weighted summation of the non-uniformity of the state transition probability distribution; The maximum Lyapunov exponent value and the Kolmogorov entropy value are normalized and fused to construct a chaotic intensity feature vector. The Euclidean distance or cosine similarity is used to quantitatively represent the individual chaotic level, and the chaotic feature vector is output.

7. The gait assessment method of human nonlinear system analysis technique according to claim 6, wherein, In step S3, the minimum delay time of the signal is estimated by the mutual information method, and the candidate embedding dimension data set is output based on Cao criterion and mean square error convergence analysis. Then, the maximum Lyapunov exponent is estimated using the small data divergence method for the reconstructed trajectory set, and the chaotic feature combination vector is obtained by combining the Kolmogorov entropy normalization curve using the approximate entropy algorithm.

8. The gait assessment method of human nonlinear system analysis technique according to claim 1, wherein, The step S4 specifically includes: The chaotic intensity feature vector is normalized to obtain a standardized chaotic feature input vector; Based on the standardized chaotic feature input vector, a three-layer feedforward wavelet neural network structure with a multi-scale wavelet basis function is constructed, wherein the number of input layer nodes is consistent with the dimension of the chaotic feature, the Morlet wavelet basis function is used for the hidden layer nodes, and the output layer nodes correspond to the delay time and embedding dimension parameters; The wavelet neural network is initialized and trained, and the historical subject sample data set is used for supervised learning. The actual phase space reconstruction parameters are used as the expected output, and the mean square error between the network output and the true parameters is minimized. A dynamic decoupling mechanism is introduced in the training process to independently adjust the output nodes of the delay time and the embedding dimension through error back propagation; After the training is completed, the normalized chaotic feature of the current subject is input into the optimized wavelet neural network, and the corresponding optimal delay time and embedding dimension parameter combination is output, and the parameter combination is packaged as a phase space reconstruction parameter configuration file.

9. The gait assessment method of human nonlinear system analysis technique according to claim 1, wherein, The step S5 specifically includes: Performing sliding window sampling on the chaotic feature vector to obtain a chaotic intensity dynamic change sequence in a continuous gait cycle; Based on the chaotic intensity sequence in the sliding window, the time correlation of the chaotic feature is evaluated by the calculation method to obtain a chaotic feature long-term memory index; Performing Euclidean distance calculation on the chaotic feature long-term memory index and the chaotic feature value at the initial modeling to quantify the chaotic intensity drift degree; Performing a threshold comparison operation on the chaotic intensity drift degree, and if the drift amount exceeds a preset deviation threshold, a parameter reconfiguration trigger signal is generated; Based on the parameter reconfiguration trigger signal, an update mechanism in the chaotic feature-parameter mapping model library is called to dynamically re-optimize the configuration of the current delay time and embedding dimension.

10. A gait assessment system of human nonlinear system analysis technology, characterized by: The gait evaluation method of the human nonlinear system analysis technology according to any one of claims 1-9 is used for gait evaluation.

Images