Gait prediction method and system based on support-aware multi-scale temporal convolutional network

Abstract

This invention discloses a gait prediction method and system based on a support-aware multi-scale temporal convolutional network. The method includes the following steps: S1, collecting plantar pressure data through pressure insoles and ground reaction force data through a force plate to construct a gait dataset; S2, performing data preprocessing and feature enhancement on the dataset from step S1; S3, establishing a multi-scale support-aware temporal convolutional network (MSA-TCN) model; S4, training the network model from step S3 using supervised learning, inputting the dataset from step S2 into the trained network model for prediction, and outputting the predicted ground reaction force. This invention overcomes the shortcomings of existing technologies in multi-scale temporal feature extraction, gait phase-aware modeling, and synchronous high-precision prediction of multi-dimensional mechanical parameters.

Description

Technical Field

[0001] This invention belongs to the field of human gait prediction technology, specifically relating to a human gait prediction method and system based on a support-aware multi-scale temporal convolutional network. Background Technology

[0002] Ground reaction force (GRF), ground reaction torque (GRM), and center of pressure (COP) are key parameters for human gait analysis, rehabilitation assessment, and exercise training. GRF describes forces in three dimensions, GRM reflects rotational effects around the anatomical axis, and COP indicates the location of the resultant force application point. Together, they characterize lower limb load, joint dynamics, and balance control mechanisms.

[0003] While traditional force tables offer high precision and temporal resolution, spatial constraints limit their ability to monitor continuous gait and prevent them from recording natural motion variability across multiple gait cycles. Furthermore, multi-force table systems are costly and cannot be applied to complex motion scenarios such as uneven outdoor terrain or stair climbing. Wearable pressure sensing systems offer a portable and low-cost alternative, but predicting continuous multidimensional mechanical parameters from discrete pressure data presents challenges: the pressure-to-force mapping exhibits complex nonlinear and time-varying characteristics; the force generation mechanisms differ significantly across support phases; and inter-individual variations place higher demands on the model's generalization ability. Therefore, some new technologies have emerged to address the problems of traditional techniques.

[0004] 1) Traditional machine learning methods: These methods use stepwise linear regression, support vector regression, or random forests to establish a mapping from stress to GRF. These methods heavily rely on manual feature engineering and lack the ability to automatically learn deep features; they struggle to capture complex nonlinear mapping relationships and dynamic time-varying characteristics; while performing reasonably well in vertical force prediction, their accuracy is severely insufficient in horizontal shear force and three-dimensional moment prediction.

[0005] 2) Recurrent Neural Network Method: LSTM or GRU is used to process temporal stress signals. This type of method uses a sequential computation mechanism, which cannot be parallelized for training, resulting in low computational efficiency; it is susceptible to gradient vanishing when processing long gait sequences; and it processes all time scales uniformly, failing to distinguish between transient impacts upon foot strike (50-100ms) and the complete gait cycle dependence (1000-1200ms), leading to phase delays and amplitude attenuation in the prediction of rapidly changing signals.

[0006] 3) Convolutional Neural Network (CNN) method: This method uses a one-dimensional CNN or standard TCN to extract local features of the pressure signal. This type of method uses fixed-size convolutional kernels, which limits its ability to capture multi-scale temporal features. It cannot dynamically adjust the receptive field and struggles to simultaneously extract local high-frequency impact features and global low-frequency periodic features. Furthermore, it lacks an explicit perception and modeling mechanism for the gait support phase and fails to utilize structured prior information about the gait phase, resulting in limited prediction accuracy during the phase transition phase.

[0007] The following publicly available Chinese patent documents were found through a search:

[0008] Publication number CN110232412A proposes a human gait prediction method based on multimodal deep learning, which has the following shortcomings: it uses a single fixed-size convolutional kernel and lacks a multi-scale parallel branch design, making it difficult to capture short-term transient impacts and long-term periodic trends at the same time; it does not explicitly inject the gait support phase as a structured prior into the network and lacks an attention mechanism to adaptively adjust feature weights according to biomechanical state, which limits the prediction accuracy of the phase transition period.

[0009] Publication number CN110313918A proposes a gait phase recognition method and system based on plantar pressure, which has the following shortcomings: it focuses on discrete phase classification rather than continuous GRF / GRM numerical prediction; it relies on preset statistical assumptions and traditional probability models, and its feature capture and fitting ability is significantly weaker than that of deep neural networks when dealing with complex nonlinear mapping relationships.

[0010] In summary, existing technologies have significant shortcomings in multi-scale temporal feature extraction, gait phase sensing modeling, and high-precision synchronous prediction of multi-dimensional mechanical parameters, and new technical solutions are urgently needed. Summary of the Invention

[0011] In summary, the technical problem to be solved by the gait prediction method and system based on support-aware multi-scale temporal convolutional networks of this invention is as follows:

[0012] 1) Existing deep learning methods struggle to capture multi-scale temporal features, limiting their ability to capture complex motion signals: Ground reaction force prediction requires simultaneously analyzing rapid transient impacts from foot strikes lasting only 50-100ms and long-term gait periodic dependencies lasting 1000-1200ms. Existing recurrent neural networks (RNNs) suffer from gradient vanishing and uniform processing across all time scales, while standard temporal convolutional networks (TCNs) are limited by fixed kernel sizes, making it impossible for existing models to simultaneously capture such a wide range of multi-scale temporal features.

[0013] 2) Existing technologies lack explicit perception and effective fusion of support phase information, neglecting the structured priors of gait: gait includes different phases such as single-leg support, double-leg support, and swing phase, with significant differences in pressure distribution and force generation mechanisms at each stage. Existing research has not combined multi-scale temporal modeling with phase-aware attention, and the structural information of gait phases is not fully utilized, making it difficult for models to adaptively adjust feature extraction strategies according to the current support state (e.g., left foot, right foot, or double-leg support).

[0014] 3) Existing methods have insufficient prediction accuracy for multidimensional biomechanical parameters (especially horizontal shear force and torque): The mapping from pressure to force has complex nonlinear and time-varying characteristics. Existing methods (such as conventional regression or shallow networks) mainly focus on the estimation of vertical ground reaction force (Vertical GRF), and there are few studies on the prediction of horizontal shear force (Fx, Fy) and three-dimensional torque (GRM) with low accuracy, which makes it difficult to meet the high accuracy requirements of clinical diagnosis and motion assessment for full-dimensional biomechanical parameters.

[0015] The technical solution adopted by the present invention to solve its technical problem is as follows:

[0016] A gait prediction method based on a support-aware multi-scale temporal convolutional network includes the following steps:

[0017] S1. Collect foot pressure data through pressure insoles and ground reaction force data through force plates to construct gait datasets;

[0018] S2. Perform data preprocessing and feature enhancement on the dataset from step S1;

[0019] S3. Establish the MSA-TCN model of a multi-scale support-aware temporal convolutional network;

[0020] S4. Train the network model from step S3 using supervised learning, input the dataset from step S2 into the trained network model for prediction, and output the predicted ground reaction force.

[0021] This invention also discloses a gait prediction system based on a support-aware multi-scale temporal convolutional network for performing the above method, which includes the following modules:

[0022] Gait dataset construction module: Collects plantar pressure data through pressure insoles and ground reaction force data through force plates to construct gait datasets;

[0023] Data processing and feature enhancement module: performs data preprocessing and feature enhancement on the dataset;

[0024] MSA-TCN Model Construction Module: Establishes the multi-scale support-aware temporal convolutional network MSA-TCN model;

[0025] Training and prediction module: The MSA-TCN model is trained using supervised learning. The feature-enhanced dataset is input into the trained network model for prediction, and the predicted ground reaction force is output.

[0026] Compared with the prior art, the present invention has the following significant technical advantages:

[0027] 1. Effectively solves the technical challenge of simultaneously capturing transient impact and long-term trends: This invention designs a parallel multi-scale temporal convolutional module (MS-TCN) and utilizes convolutional kernels of different sizes (7, 9, 11) and progressively increasing dilated convolutions (dilation rates from 1 to 32) to achieve parallel extraction of multi-level temporal features. This technique can simultaneously and accurately capture rapid transient impact features of foot strike lasting only 50-100ms and complete gait cycle-dependent features lasting 1000-1200ms in the same model, effectively overcoming the shortcomings of standard convolutional networks, such as fixed receptive fields and difficulty in adapting to the multi-scale characteristics of biomechanical signals.

[0028] 2. Overcoming the limitation of traditional deep learning models lacking prior biomechanical knowledge: This invention innovatively introduces a bidirectional gated recurrent unit (BiGRU) as a support phase encoder, converting discrete gait support states (such as left-foot support, two-foot support, etc.) into continuous high-dimensional semantic representations. This design explicitly injects prior knowledge of gait structure into the neural network, enabling the model to possess "phase awareness" capabilities. This significantly reduces prediction errors at critical phase transition stages (such as the switch from two-foot support to one-foot support), solving the prediction lag or smooth transition problems caused by the lack of phase guidance in existing methods.

[0029] 3. Achieved adaptive and precise fusion of pressure features and phase information: This invention utilizes a multi-head attention mechanism, guided by support phase features, to dynamically calculate the attention weights of pressure temporal features. This enables the model to adaptively adjust its focus based on the current motion state (e.g., focusing on arch pressure in a single support phase and heel pressure in a double support phase), thereby achieving optimal weighted fusion of multi-source information in complex dynamic motions and further improving the model's robustness.

[0030] 4. This invention achieves high-precision joint prediction of ground reaction force, ground reaction moment, and pressure center, and solves the technical challenge of simultaneous estimation of multidimensional mechanical parameters in existing technologies: Addressing the limitation of existing technologies to predicting only the vertical ground reaction force (Fz), this invention achieves full-dimensional biomechanical parameter estimation based on insole sensors through deep feature learning. Specifically, this invention can not only predict the vertical ground reaction force, but also accurately predict the horizontal shear force (Fx, Fy) and three-dimensional moments (Mx, My, Mz), which are difficult for existing technologies to handle. In particular, it can accurately capture the dynamic changes in inversion / outversion oscillations and torsional moments, thus providing complete multidimensional mechanical parameter data support for biomechanical analysis of complex motion scenarios (such as directional movements and balance assessments). Attached Figure Description

[0031] Figure 1 This is a flowchart of a preferred embodiment of the present invention, which describes a gait prediction method based on a support-aware multi-scale temporal convolutional network.

[0032] Figure 2 This is a diagram of a plantar pressure measurement system according to a preferred embodiment of the present invention.

[0033] Figure 3 This is a diagram of a force measuring plate system according to a preferred embodiment of the present invention.

[0034] Figure 4 This is a structural diagram of the MSA-TCN model according to a preferred embodiment of the present invention.

[0035] Figure 5 This is a block diagram of a gait prediction system based on a support-aware multi-scale temporal convolutional network, according to a preferred embodiment of the present invention. Detailed Implementation

[0036] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0037] This embodiment proposes a gait prediction method based on a support-aware multi-scale temporal convolutional network, such as... Figure 1 As shown, the main process includes four stages: constructing a gait dataset using pressure insoles and force plates, data preprocessing and feature enhancement, establishing a multi-scale support-aware temporal convolutional network (MSA-TCN), and model training and prediction output.

[0038] Phase 1: Data Acquisition and Preprocessing

[0039] 1. Data Acquisition Hardware Configuration: The data acquisition tool involved in this embodiment is an integrated data acquisition device, including a force plate system and a plantar pressure measurement system. For example... Figure 3As shown, the force measuring plate adopts a 60×60cm aluminum alloy upper and lower plate structure (10mm thick). Four three-dimensional force sensors (SBT301A type, Guangzhou Spartau Electronic Technology Co., Ltd.) are symmetrically placed between the upper and lower plates. The center-to-center distance between adjacent sensors is 550mm, and the distance from the edge of the force measuring plate is 25mm. Each sensor can simultaneously measure the force components in the X, Y, and Z directions (range 100kg / axis). Through an 8-channel signal transmitter (SBT908D type, Guangzhou Spartau Electronic Technology Co., Ltd.), the signal is conditioned and amplified to measure the forces (Fx1, Fx2, Fx3, Fy1, Fy2, Fy3, Fz1, Fz2, Fz3) in the three directions of each three-dimensional force sensor. The three-dimensional ground reaction force (Fx, Fy, Fz) is obtained by force composition. The force composition formula is as follows: Where i represents the number of the three-dimensional force sensor. See also Figure 2 Foot pressure data is collected via an embedded pressure insole. The insole sensors (in this embodiment, a sensor array using flexible printed circuit (FPC) technology, optionally manufactured by Luoyang Guantuo Electronic Technology Co., Ltd.) cover four weight-bearing areas: heel, arch, forefoot, and toes, with a total of 18 pressure sensing units. Ground truth is collected via a force plate containing four three-dimensional force sensors for synchronously measuring the three-dimensional ground reaction forces (Fx, Fy, Fz). Subjects wear the data acquisition device and perform designated actions to collect data. The sampling frequency of the device is set to 50Hz, and time-stamp alignment technology ensures a synchronization error of less than 1ms. Ground reaction torques (Mx, My, Mz) and (COP_x, COP_y) are calculated based on the ground reaction forces, using the following formula: , , The center of pressure is then calculated based on the ground reaction force / torque, using the following formula: , .

[0040] The second stage: Data processing and feature enhancement. The data processing workflow includes sensor selection, signal filtering, and normalization. To enhance the model's ability to perceive temporal changes, this embodiment constructs three types of derived features and divides the samples into fixed-length time windows. This stage performs data processing and feature enhancement based on the raw data collected in the first stage (including pressure values ​​from 18 pressure sensors, three-dimensional ground reaction forces Fx, Fy, and Fz synchronously collected by the force plate, and calculated three-dimensional ground reaction moments Mx, My, and Mz, and pressure center coordinates COP_x and COP_y) to construct input features and label data suitable for deep learning model training. The data processing workflow includes sensor selection, signal filtering, normalization, feature extraction, and sliding window segmentation. Specifically:

[0041] 2.1 Sensor Selection and Signal Filtering: To reduce data redundancy, an activity-based sensor selection strategy was adopted, selecting the top 10 sensor data points for each foot's plantar activity. To address baseline drift and high-frequency noise in the raw pressure signals, a 4th-order Chebyshev Type I bandpass filter was used, with the passband frequency set to 0.5–20 Hz. The actual GRF values ​​output by the force plate were smoothed using a Savitzky-Golay filter to reduce measurement noise. For the raw data from the 18 pressure sensors collected in the first stage, an activity-based sensor selection strategy was adopted to reduce data redundancy and improve model efficiency. The top 10 sensor data points for each foot's plantar activity were selected, resulting in 20-dimensional pressure features (10 sensors for each foot). To address baseline drift and high-frequency noise in the selected 20-dimensional pressure signal, a 4th-order Chebyshev Type I bandpass filter was used, with the passband frequency set to 0.5–20 Hz to preserve the effective frequency band of gait motion and filter out interference signals. For the actual GRF, GRM, and COP value tag data output from the first-stage force plate, a Savitzky-Golay filter (window length 7, polynomial order 3) was used for smoothing to reduce measurement noise and ensure the smoothness and reliability of the tag data.

[0042] 2.2 Data Normalization: To eliminate differences in feature dimensions between different sensors and accelerate network convergence, the Z-score normalization method is used to process all features. The normalization formula is as follows: .in, These are the normalized eigenvalues. denoted as the mean and standard deviation of the i-th feature on the training set, respectively.

[0043] 2.3 Feature Extraction

[0044] 1) First-order difference characteristic: Calculates the rate of pressure change, reflecting the dynamic trend; the formula is... ,in, Let be the pressure value of the i-th sensor at time t.

[0045] 2) Moving average feature: Smoothing and suppressing local fluctuations through a time window, the formula is as follows: .

[0046] 3) Supporting State Features: Constructing Supporting State Labels These correspond to four states: left foot support, right foot support, both feet support, and both feet off the ground.

[0047] First-order difference features and moving average features are feature enhancement processes. First-order difference features calculate the time derivative from the raw pressure data, adding information about time variations; moving average features smooth the raw pressure data, reducing noise. Support state features are additional sensor data.

[0048] 2.4 Input Vector Construction: The 20-dimensional original pressure features, 20-dimensional difference features, 20-dimensional moving average features, and 4-dimensional support phase labels are concatenated to form a 64-dimensional input feature vector.

[0049] 2.5 Sliding Window Segmentation: The sliding window method is used to segment the continuous time series data. The window length is set to 187 sample points (approximately 3.74 seconds), and the sliding step size is 10 sample points to cover the complete gait cycle and maintain the temporal continuity.

[0050] Phase 3: Network Model Construction

[0051] See Figure 3 This embodiment constructs a multi-scale support-aware temporal convolutional network (MSA-TCN), which includes three core modules:

[0052] 3.1 Parallel Multi-Scale Temporal Convolution Module (MS-TCN): This module is used to extract temporal features from stress data. It contains three parallel branches with kernel sizes of 7, 9, and 11, respectively, to accommodate feature extraction needs at different scales. Each branch employs a dilated convolution structure, with dilation factors set sequentially to 1, 2, 4, 8, 16, and 32, thereby exponentially expanding the receptive field without increasing computational cost. The outputs of each branch are concatenated along the channel dimension to form a fused stress temporal feature vector. .

[0053] 3.2 Support Phase Encoder Module: This module uses a Bidirectional Gated Recurrent Unit (BiGRU) as its core component. The input is a discrete sequence of support state features. The BiGRU integrates historical and look-ahead contextual information through a bidirectional processing mechanism, transforming the discrete support phase labels into a continuous high-dimensional semantic representation. This explicitly introduces biomechanical structural priors.

[0054] 3.3 Multi-head Attention Fusion Module (Support-Aware Attention): This module aims to achieve adaptive fusion of stress features and phase information. The specific calculation process is as follows:

[0055] 1) To support phase features As the query vector (Query, Q), with stress time series features As key vectors (Key, K) and value vectors (Value, V) 17 .

[0056] The attention weights are calculated using the Softmax function and then summed using a weighted average. The formula is as follows:

[0057]

[0058] in, The dimension of the key (192 in this embodiment).

[0059] Adopting a multi-head mechanism ( This allows the model to focus on different feature interaction patterns in parallel (such as heel-forefoot transfer or medial-lateral balance), as shown in the formula: ; represents the attention heads in a multi-head attention mechanism, where head1 is the output of the first attention head and head2 represents the output of the attention head;

[0060] Among them, w o It is the output projection matrix, whose function is to concatenate the outputs of all attention heads and then map them back to the original dimension through linear transformation.

[0061] Finally, the output is processed through residual connections and layer normalization to obtain the final fused feature vector. ;

[0062]

[0063] Phase 4: Model Training and Predictive Output

[0064] This stage is based on the Multi-Scale Support-Aware Temporal Convolutional Network (MSA-TCN) model built in the third stage. It uses the training sample set (input feature matrix X and label matrix Y) obtained in the second stage to train the model and achieve high-precision prediction output of ground reaction force, ground reaction torque and pressure center.

[0065] 4.1 Model Input and Forward Propagation

[0066] During training, the 64-dimensional input feature vector constructed in the second stage is separated into two independent input streams:

[0067] 1) Pressure Feature Flow: Extract the first 60 dimensions of features (including 20 dimensions of original pressure, 20 dimensions of difference features, and 20 dimensions of moving average features), with dimensions of [Batch, 187, 60], and input them into the parallel multi-scale temporal convolution module (MS-TCN) constructed in the third stage for temporal feature extraction.

[0068] 2) Support phase flow: Extract the 4-dimensional support state features with dimensions [Batch, 187, 4] and input them into the support phase encoding module (Support Encoder based on BiGRU) constructed in the third stage for phase semantic encoding.

[0069] The MS-TCN module extracts multi-scale temporal features of stress through three parallel branches (with kernel sizes of 7, 9, and 11, and dilation factors increasing from 1 to 32), with an output dimension of [Batch, 187, C], where C is the number of channels after concatenation. The Support Encoder module converts the 4-dimensional discrete support states into a continuous high-dimensional semantic representation, with an output dimension of [Batch, 187, C].

[0070] 4.2 Feature Fusion and Enhancement

[0071] The third-stage multi-head attention fusion module (Support-Aware Attention) receives the stress temporal features output by MS-TCN and the phase semantic features output by Support Encoder. It uses the phase features as the query vector and the stress features as the key and value vectors, respectively, and calculates the attention weights through a multi-head attention mechanism.

[0072]

[0073] Where, d k The dimension is [Batch, 187, C]. A multi-head mechanism (4 heads in this embodiment) enables the model to focus on different feature interaction patterns in parallel. The attention-fused features are processed through residual connections and layer normalization to obtain a fused feature vector F_fused, with dimensions [Batch, 187, C]. The fused feature vector F_fused then undergoes nonlinear transformation and feature enhancement through a fully connected layer fc_mid. fc_mid consists of a linear layer, a batch normalization layer (BatchNorm1d), and a ReLU activation function, processing the features at each time step independently, while maintaining the output dimension [Batch, 187, C]. This fully connected layer design ensures sufficient spatial representation of the fused features, providing high-quality feature representations for subsequent prediction heads.

[0074] 4.3 Loss Function Design: The model training uses a weighted smooth L1 loss function (SmoothL1 Loss), and the total loss function is defined as: ,in, All are 1.0. Used to constrain the prediction error of three-dimensional forces and moments. The prediction error for constraining the pressure center is used. Model training process: A weighted smoothed L1 loss function is used to separately calculate the force prediction loss and the pressure center prediction loss. The optimizer uses AdamW (learning rate 1e-4, weight decay 1e-4) combined with the ReduceLROnPlateau learning rate scheduler (patience=12, decay factor 0.5) and gradient clipping (clip_value=1.0), employing mixed-precision training to accelerate convergence.

[0075] 4.4 Prediction Output: The fused feature vector F is processed by a fully connected layer and then split into two prediction heads.

[0076] 1) Force / Torque Prediction Head: Outputs a 6-dimensional vector, corresponding to Fx' (forward and backward shear force), Fy' (vertical reaction force), Fz' (inward and outward shear force), Mx (torque about the X-axis), My (torque about the Y-axis), and Mz (torque about the Z-axis).

[0077] 2) Center of Pressure Prediction Head (COP Head): Outputs a 2D vector, corresponding to the COP_x and COP_y coordinates respectively.

[0078] The supervised learning methods involved in step S4 are briefly described as follows: 1. Loss function design: Loss function (L force and L cop The core mechanism of supervised learning is to quantify the prediction error by comparing the model's predicted values ​​with the ground truth collected by the force plate. 2. Prediction output: The parameters of both prediction heads (force / torque prediction head and COP prediction head) are trained using supervised learning. During training, the output of the prediction head is compared with the ground truth, and the parameters are updated using the backpropagation algorithm to make the predicted values ​​approximate the true values.

[0079] The following experiment will be conducted to verify the technical advantages of this invention.

[0080] Table 1 Performance of MSA-TCN across various prediction dimensions

[0081] Center of Force / Pressure MAE RMSE R² Fx' 0.0605 0.0566 0.9931 Fy' 0.0647 0.0584 0.9922 Fz' 0.0564 0.0512 0.9939 Mx 0.0510 0.0704 0.9950 My 0.0465 0.0640 0.9959 Mz 0.0648 0.0890 0.9921 COP_x 0.0438 0.0553 0.9964 COP_y 0.0491 0.0507 0.9954

[0082] As shown in Table 1, the MSA-TCN model of this invention achieved excellent performance across all eight prediction dimensions. In the prediction of three-dimensional ground reaction forces, the mean absolute errors (MAE) of Fx', Fy', and Fz' were 0.0605, 0.0647, and 0.0564, respectively, and the root mean square errors (RMSE) were 0.0566, 0.0584, and 0.0512, respectively. The coefficients of determination (R²) were all above 0.99, with Fz exhibiting the highest prediction accuracy (R² = 0.9939). In the prediction of three-dimensional ground reaction moments, the MAEs of Mx, My, and Mz ranged from 0.0465 to 0.0648, the RMSEs ranged from 0.0640 to 0.0890, and the R² values ​​reached 0.9921 to 0.9959. Among these, My performed best (MAE = 0.0465, R² = 0.9959). In the prediction of two-dimensional pressure centers, the MAEs for COP_x and COP_y are 0.0438 and 0.0491, respectively, the RMSEs are 0.0553 and 0.0507, and the R² values ​​reach 0.9964 and 0.9954, respectively. This invention achieves high-precision mapping of plantar pressure signals to multi-dimensional biomechanical parameters by fusing a multi-scale temporal convolutional network with a support phase-aware attention mechanism. Experimental data show that the coefficients of determination (R²) for all prediction dimensions exceed 0.99, the MAE is controlled within 0.065, and the RMSE is below 0.09, achieving an accuracy level comparable to force table measurements.

[0083] Compared with existing technologies, the present invention has the following technical effects: (1) It realizes the synchronous high-precision prediction of 8 biomechanical parameters, solving the problems of low single-dimensional prediction accuracy and incomplete information in traditional methods; (2) Through multi-scale feature extraction, it effectively captures the pressure change law under different time scales, so that the prediction R² of complex torque components is still above 0.99; (3) The prediction accuracy of pressure center (R²>0.996) provides reliable technical support for balance function assessment and gait abnormality detection; (4) It uses portable foot pressure insoles to replace force measuring platform equipment, which significantly reduces the measurement cost while maintaining high accuracy, and provides technical feasibility for the daily application of gait analysis.

[0084] like Figure 5 As shown, this embodiment discloses a gait prediction system based on a support-aware multi-scale temporal convolutional network, used to perform the above method, including the following modules:

[0085] Gait dataset construction module: Collects plantar pressure data through pressure insoles and ground reaction force data through force plates to construct gait datasets;

[0086] Data processing and feature enhancement module: performs data preprocessing and feature enhancement on the dataset;

[0087] MSA-TCN Model Construction Module: Establishes the multi-scale support-aware temporal convolutional network MSA-TCN model;

[0088] Training and prediction module: The MSA-TCN model is trained using supervised learning. The feature-enhanced dataset is input into the trained network model for prediction, and the predicted ground reaction force is output.

[0089] Other aspects of this embodiment can be found in the above method embodiments.

[0090] The key technical points of this invention are summarized as follows:

[0091] 1. Design of Parallel Multi-Scale Temporal Convolutional Module (MS-TCN): This invention designs a convolutional architecture with three parallel branches to process plantar pressure data sequences. To simultaneously accommodate the extraction of transient impact and long-term gait features, the three branches employ different convolutional kernels of sizes 7, 9, and 11, respectively. Furthermore, within each branch, a progressively increasing dilated convolution strategy is used, with dilation factors set to 1, 2, 4, 8, 16, and 32, respectively. This structural design allows the model to simultaneously capture both rapid transient impact features lasting only 50-100ms and complete gait cycle trend features lasting 1000-1200ms within the same layer, effectively solving the problem that the fixed receptive field of standard convolutional networks cannot match multi-scale biomechanical signals.

[0092] 2. Bidirectional Gated Recurrent Unit (BiGRU) Support Phase Encoder: This invention introduces a BiGRU as a support phase encoder, specifically designed to process discrete support state labels (including four states: left foot support, right foot support, both feet support, and both feet off the ground). Through the bidirectional processing mechanism of the BiGRU, the model can integrate historical and prospective contextual information at each time point, transforming discrete support phase labels into continuous high-dimensional semantic representations. This key point explicitly injects structured prior information about gait into the neural network, enabling the model to possess "phase awareness" capabilities and adjust prediction strategies based on the current biomechanical state, thus overcoming the deficiency of existing deep learning methods in explicitly modeling gait phases.

[0093] 3. Supporting a Multi-Head Attention Fusion Mechanism: This invention develops a multi-head attention mechanism to achieve adaptive fusion of pressure features and phase information. Specifically, this mechanism uses the phase features output by the BiGRU as the query vector and the temporal pressure features extracted by MS-TCN as the key and value vectors. By calculating attention weights, the model can dynamically learn phase-dependent feature weighting patterns. For example, in a single support phase, it automatically focuses on pressure changes in the arch or forefoot, while in a dual support phase, it shifts to pressure features in the heel or toes, thus significantly improving the feature fusion effect during complex gait transitions.

[0094] 4. Synchronous High-Precision Prediction Architecture for Multi-Dimensional Biomechanical Parameters: This invention constructs an end-to-end deep prediction model. The output includes two independent prediction heads, used to generate 6-dimensional forces / torques (Fx, Fy, Fz, Mx, My, Mz) and 2-dimensional pressure centers (COP_x, COP_y), respectively. Trained using a weighted smoothing L1 loss function, this model achieves, for the first time, the overall prediction of 3D ground reaction force, 3D torque, and pressure center based on wearable insole sensors. Experimental results show that this key technology enables the model to achieve a coefficient of determination (R²) exceeding 0.99 across all eight prediction dimensions, significantly reducing prediction errors for horizontal shear force and torsional torque compared to existing technologies.

[0095] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A gait prediction method based on a support-sensory multi-scale temporal convolutional network, characterized by: Includes the following steps: S1. Collect foot pressure data through pressure insoles and ground reaction force data through force plates to construct gait datasets; S2. Perform data preprocessing and feature enhancement on the dataset from step S1; S3. Establish the MSA-TCN model of a multi-scale support-aware temporal convolutional network; S4. Train the network model from step S3 using supervised learning, input the dataset from step S2 into the trained network model for prediction, and output the predicted ground reaction force.

2. The gait prediction method based on a support-aware multi-scale temporal convolutional network as described in claim 1, characterized in that, In step S1, foot pressure data is collected through an embedded pressure insole sensor. The insole sensor layout covers four weight-bearing areas: heel area, arch area, forefoot area, and toe area.

3. The gait prediction method based on a support-aware multi-scale temporal convolutional network as described in claim 1 or 2, characterized in that, step... In S1, the actual value of the ground reaction force is acquired through a force plate. The force plate has a built-in three-dimensional force sensor for synchronously measuring the three-dimensional ground reaction force (Fx, Fy, Fz). The sampling frequency is set to 50Hz and aligned using timestamps. The ground reaction torque (Mx, My, Mz) and pressure center (COP_x, COP_y) are calculated based on the ground reaction force, using the following formula: Where Fx is the ground reaction force in the x direction, Fy is the ground reaction force in the y direction, Fz is the ground reaction force in the z direction, and 1, 2, 3, and 4 represent the numbers of the three-dimensional force sensors. The pressure center is calculated based on the ground reaction force or moment, using the following formula: 、 In the above formula, 'a' represents the horizontal distance from the center of the three-dimensional force sensor to the center of the force measuring plate; the force measuring plate is placed horizontally, and the coordinate system is defined as follows: the Z-axis is perpendicular to the force measuring plate and pointing upwards, the X-axis is parallel to the force measuring plate and pointing forwards, and the Y-axis is pointing to the right. The sensors are numbered 1-4 clockwise.

4. The gait prediction method based on a support-aware multi-scale temporal convolutional network as described in claim 3, characterized in that, In step S2, data preprocessing includes sensor selection and signal filtering, and normalization, as detailed below: Sensor selection and signal filtering: An activity-based sensor selection strategy was adopted to select the top 10 sensor data of plantar activity for each foot; a 4th-order Chebyshev Type I bandpass filter was used for processing, with the passband frequency set to 0.5-20Hz; and a Savitzky-Golay filter was used to smooth the true value of the ground reaction force (GRF) output by the force plate. Data normalization: All features are processed using the Z-score normalization method; The normalization formula is as follows: ;in, μ is the normalized eigenvalue. i and σ i Let be the mean and standard deviation of the i-th feature on the training set, respectively; Feature extraction: 1) First-order difference characteristic: Calculates the rate of pressure change, reflecting the dynamic trend. The formula is: Where, p i (t) represents the pressure value of the i-th sensor at time t; 2) Moving average feature: Smoothing and suppressing local fluctuations through a time window, the formula is: 3) Supporting State Features: Constructing Supporting State Labels These correspond to four states: left foot support, right foot support, both feet support, and both feet off the ground. Input vector construction: The 20-dimensional original pressure feature, 20-dimensional difference feature, 20-dimensional moving average feature and 4-dimensional support phase label are concatenated to form a 64-dimensional input feature vector; Sliding window segmentation: The sliding window method is used to segment continuous time series data. The window length is set to 187 sample points and the sliding step size is 10 sample points.

5. The gait prediction method based on a support-aware multi-scale temporal convolutional network as described in claim 4, characterized in that, In step S3, the multi-scale support-aware temporal convolutional network MSA-TCN includes the following modules: Parallel multi-scale temporal convolution module: Used to extract temporal features from stress data, it contains three parallel branches with kernel sizes of 7, 9, and 11 respectively. Each branch employs a dilated convolution structure with dilation factors set to 1, 2, 4, 8, 16, and 32 according to the hierarchy. The outputs of each branch are concatenated along the channel dimension to form the fused stress temporal feature vector H. p ; Support Phase Encoding Module: Employs a Bidirectional Gated Recurrent Unit (BiGRU) as the core component; the input is a discrete sequence of support state features. The BiGRU integrates historical and look-ahead contextual information through a bidirectional processing mechanism, transforming the discrete support phase labels into a continuous high-dimensional semantic representation H. s ; Multi-head self-attention fusion module: designed to achieve adaptive fusion of stress features and phase information; the specific process is as follows: To support phase feature H s As the query vector (Query, Q), with pressure time series features H p As key vectors (Key, K) and value vectors (Value, V) 17 ; The attention weights are calculated using the Softmax function and then summed using a weighted average. The formula is as follows: Where, d k The dimension of the key; A multi-head mechanism is adopted, allowing the model to focus on different feature interaction patterns in parallel. The formula is as follows: Where h=4, W o It is the output projection matrix, where head1 represents the output of the first attention head and head2 represents the output of the attention head; Finally, the output is processed by residual connections and layer normalization to obtain the final fused feature vector F. 。 6. The gait prediction method based on a support-aware multi-scale temporal convolutional network as described in claim 5, characterized in that, In step S4, the supervised learning method is as follows: Loss function design: The total loss function is defined as follows: Where λ1 and λ2 are both 1.0; L force L is used to constrain the prediction error of three-dimensional forces and moments. cop Prediction error used to constrain the center of pressure; Prediction output: The fused feature vector F, after being processed by a fully connected layer, is split into two prediction heads: 1) Force / Torque Prediction Head: Outputs a 6-dimensional vector, corresponding to forward and backward shear force Fx', vertical reaction force Fy', inward and outward shear force Fz', torque around the X-axis Mx, torque around the Y-axis My, and torque around the Z-axis Mz, respectively; 2) Pressure center prediction head: Outputs a 2D vector, corresponding to the COP_x and COP_y coordinates respectively.

7. A gait prediction system based on a support-sensory multi-scale temporal convolutional network, used to perform the method as described in any one of claims 1-6, characterized in that, Includes the following modules: Gait dataset construction module: Collects plantar pressure data through pressure insoles and ground reaction force data through force plates to construct gait datasets; Data processing and feature enhancement module: performs data preprocessing and feature enhancement on the dataset; MSA-TCN Model Construction Module: Establishes the multi-scale support-aware temporal convolutional network MSA-TCN model; Training and prediction module: The MSA-TCN model is trained using supervised learning. The feature-enhanced dataset is input into the trained network model for prediction, and the predicted ground reaction force is output.

Images