Wearable human gait analysis system

Abstract

The application discloses a wearable human gait analysis system, comprising a synchronous data acquisition module, a plantar pressure topology graph module, a pre-training model module, a real scene model module and a pathological gait classification model module; the synchronous data acquisition module collects real force platform data, insole pressure time series data and inertial measurement unit data, and constructs a gait data set; the plantar pressure topology graph module topologically expresses the insole pressure time series data, and outputs a plantar pressure topology graph corresponding to a time sequence based on the insole pressure time series data; the pre-training model module constructs a pre-training model based on a gait data set and a space-time feature fusion network to output a gait dynamics prediction result; the real scene model module performs target domain semi-supervised self-adaptive fine-tuning on the pre-training model to generate a real scene model; and the pathological gait classification model module extracts multi-domain classification features based on the prediction results of the real scene model of each subject in a pathological state, and outputs a final classification result.

Description

Technical Field

[0001] This invention belongs to the field of biomedical engineering technology, specifically relating to a wearable human gait analysis system. Background Technology

[0002] Gait is an important external manifestation of the functional state of the human lower limbs, and parameters such as ground reaction force, ground reaction torque, and pressure centerline can further characterize key biomechanical features in the interaction between the foot and the ground. These parameters have significant application value in areas such as abnormal gait screening, rehabilitation efficacy assessment, sports injury early warning, auxiliary diagnosis of lower limb diseases, and human-computer interaction control. Therefore, establishing a technical means to continuously and quantitatively acquire gait dynamics information has clear engineering and clinical application significance.

[0003] Currently, the acquisition of gait dynamics parameters such as ground reaction force still mainly relies on laboratory equipment such as force tables and optical motion capture. Although such technologies can provide high-precision dynamic reference data, they usually require fixed sites, are costly, have complex testing procedures, and are significantly limited by space and equipment conditions. They are difficult to support long-term continuous monitoring of subjects in daily life environments, and also cannot meet the application requirements of portability, continuous operation, and low-barrier deployment in real-world scenarios.

[0004] With the development of wearable sensing technology, insole-based plantar pressure monitoring systems have gradually become an important technical approach for gait analysis. These systems can continuously record the contact information between the foot and the ground during natural walking, making them more suitable for data acquisition during free movement compared to fixed experimental equipment. However, most existing smart insoles or plantar pressure monitoring solutions still primarily focus on superficial tasks such as displaying plantar pressure distribution, estimating the center of pressure, or recognizing simple movements. Their output is usually limited to local force readings and struggles to generate higher-level gait dynamic parameters with greater biomechanical explanatory power, such as ground reaction force, ground reaction torque, and the center of pressure line.

[0005] Furthermore, existing plantar pressure sensor arrays are typically sensitive primarily to vertical contact loads, while human gait dynamics are simultaneously influenced by posture changes, load transfer, propulsion processes, and the coupling of forces in the anterior-posterior and lateral directions. Relying solely on simple integration of plantar pressure, empirical formula conversions, or ordinary time-series regression methods often fails to accurately describe the complex nonlinear mapping relationship between local plantar contact information and overall gait dynamics variables, resulting in deficiencies in the completeness, stability, and accuracy of the generated results. Especially when ground material, slope, movement speed, and environmental disturbances change, existing methods are more prone to prediction biases and struggle to maintain consistent output performance.

[0006] Furthermore, while some studies have begun to incorporate kinematic information, such as inertial measurement units (IMUs), to improve the perception of human motion states, existing technologies generally lack a unified processing framework for wearable multimodal signals. This framework cannot simultaneously consider the spatial distribution characteristics of plantar pressure signals, the temporal dynamics of inertial signals, and the data distribution variations in complex real-world scenarios, thus hindering the stable conversion from raw wearable signals to high-level gait dynamics representations. In other words, the main bottleneck of existing technologies lies not only in the deployment and acquisition of sensors themselves, but also in the lack of a gait dynamics analysis method capable of spatiotemporally coupled modeling of multimodal signals and maintaining good generalization ability in real-world scenarios.

[0007] Therefore, there is an urgent need to provide a wearable human gait analysis method and system to solve the following problems in the existing technology: laboratory equipment is difficult to use for continuous monitoring in real environments; existing insole systems are difficult to output high-level gait dynamic parameters; and existing data processing methods are difficult to reconstruct key dynamic information such as ground reaction force, ground reaction torque and pressure centerline from wearable signals such as plantar pressure and inertia stably and accurately. Summary of the Invention

[0008] This invention discloses a wearable human gait analysis system, which includes a synchronous data acquisition module, a plantar pressure topology map module, a pre-trained model module, a real-scene model module, and a pathological gait classification model module.

[0009] The synchronous data acquisition module includes a high-precision force measuring platform, a prototype insole, and a host computer. Subjects wear the prototype insole and perform multiple walking motion tests on the high-precision force measuring platform. The host computer collects real force measuring platform data, insole pressure time-series data, and inertial measurement unit data to construct a gait dataset.

[0010] The plantar pressure topology graph module performs topological representation of insole pressure time series data and outputs the corresponding plantar pressure topology graph based on the insole pressure time series data;

[0011] The pre-trained model module constructs a pre-trained model based on the gait dataset and spatiotemporal feature fusion network outputting gait dynamics prediction results;

[0012] The real-scene model module performs semi-supervised adaptive fine-tuning of the pre-trained model in the target domain to generate a real-scene model.

[0013] The pathological gait classification model module extracts multi-domain classification features based on the prediction results of the real-world scenario model for each pathological subject and outputs the final classification result.

[0014] Furthermore, the insole prototype includes 16 pressure-sensitive units distributed across the main load-bearing areas of the foot, including the heel, arch, first to fifth metatarsals, and big toe, to cover key pressure points on the sole of the foot.

[0015] Furthermore, the plantar pressure topology graph module abstracts each pressure-sensitive unit in the insole prototype as a graph node, thus forming a plantar pressure graph structure composed of 16 graph nodes; and constructs an adjacency matrix based on the physical adjacency relationships between the sensing nodes to obtain the plantar pressure topology graph. ,in, Represents a set of nodes. This indicates the connection relationships between nodes. This represents the adjacency matrix generated based on physical adjacency relationships.

[0016] Furthermore, the pre-trained model module includes an input layer, a graph structure input construction layer, a graph convolutional layer, a pooling layer, a feature concatenation layer, a bidirectional long short-term memory layer, and a fully connected regression layer, with each layer connected sequentially according to the data flow direction;

[0017] The input layer combines 16 channels of plantar pressure signals from the insole pressure time-series data at each sampling time and 6 channels of kinematic signals from the inertial measurement unit data into a 22-dimensional vector;

[0018] The graph structure input layer constructs a topology map of plantar pressure from 16 pressure sensor nodes. As input for node features;

[0019] Graph convolutional layers are used to extract local spatial correlation features of plantar pressure signals from graph-structured inputs; for any sampling time... Based on the plantar pressure topology at that moment Forming the node feature matrix at the sampling time This generates the corresponding graph structure features. ;

[0020] Pooling layers are used to compress node-level spatial features into time-level pressure space representations;

[0021] The feature splicing layer is used to fuse pressure space representation and synchronous inertial information to generate fused features;

[0022] A bidirectional long short-term memory layer is used to learn the temporal dependencies of fused feature sequences in the gait cycle;

[0023] The fully connected regression layer maps the temporal hidden state corresponding to each sampling moment to the dynamic prediction result at that moment. The dynamic prediction result includes three-dimensional ground reaction force, three-dimensional ground reaction moment and two-dimensional pressure centerline coordinates.

[0024] Furthermore, the training process for the pre-trained model module includes:

[0025] First, the 16-channel plantar pressure data collected from the prototype insole are synchronized with the 6-channel inertial measurement unit data, and then matched with the three-dimensional ground reaction force, three-dimensional ground reaction moment and two-dimensional pressure centerline coordinates in the actual force table data.

[0026] Secondly, the distribution parameters of each input dimension and each output dimension in the training data are statistically analyzed, and the 22-dimensional temporal features of the input and the 8-dimensional dynamic labels of the output are normalized accordingly to reduce the impact of different dimensions and amplitude differences on the stability of model training.

[0027] Next, the standardized continuous time series data is divided into preset time lengths. The training samples were divided into multiple training samples. For each training sample, 16-channel plantar pressure sequence and 6-channel inertial measurement unit sequence within the corresponding time window were taken as model input, and the corresponding 8-dimensional real force table data sequence within the same time window was taken as supervision label.

[0028] Subsequently, each training sample is input into the spatiotemporal fusion model to perform forward computation;

[0029] Finally, a supervised training objective is constructed based on the error between the dynamic prediction sequence and the corresponding ground truth label sequence, and the parameters of the spatiotemporal fusion model are updated.

[0030] Furthermore, the semi-supervised adaptive fine-tuning of the real-world scene model module includes the following process:

[0031] Read gait time series samples based on smart foot pads in real-world scenarios within the target domain; read the target domain samples one file at a time, maintaining the original order of the time series within each file to preserve complete continuous dynamic gait information;

[0032] Monte Carlo Dropout inference is used to generate pseudo-labels for the target domain; for the first... Target domain samples Input it into the pre-trained model for processing The second random forward propagation, the... The prediction result of the random forward propagation is denoted as:

[0033]

[0034] in, Indicates the first Pre-trained model under sub-random inactivation conditions This represents the corresponding dynamic prediction time series; pseudo-labels for the target domain samples are generated based on the mean of each prediction result.

[0035]

[0036] in, For the first Pseudo-label time series of samples from the target domain;

[0037] The uncertainty of the target domain samples is calculated; let the length of the output sequence for a single sample be... The output dimension at each time step is , for the The target domain sample, at the _th ... The time step, the first The prediction standard deviation over each output dimension is denoted as:

[0038]

[0039] in, Indicates the first The target domain samples in the th... In the second random forward propagation, at the... The time step, the first Predicted values ​​on each output dimension; Indicates the same sample in the first... The predicted value at the corresponding position under the next random forward propagation, where , is used to represent the average of all random forward propagation results;

[0040] Then, the mean of the prediction standard deviations over all time steps and all output dimensions is calculated to obtain the average uncertainty index for the target domain sample:

[0041]

[0042] in, Used to characterize the The stability and reliability of the pseudo-label prediction results for each target domain sample;

[0043] After obtaining the average uncertainty index, a high-confidence screening is performed on the target domain samples, and an uncertainty threshold is preset. If and only if At that time, the first one is reserved. Target domain samples and their corresponding pseudo-labels As a pseudo-label sample for a high-confidence target domain; when When this happens, the samples in that target domain are discarded and will not participate in subsequent fine-tuning training;

[0044] The labeled source domain samples from the laboratory stage are merged with the pseudo-labeled target domain samples with high confidence to construct a joint training dataset.

[0045] When entering the target domain inference stage, the fine-tuned model and its corresponding normalizer are loaded, and the real scene data is predicted one by one according to the file. First, the input samples are subjected to the same standardization process as in the training stage. Then, the 22-dimensional input time series features are input into the fine-tuned model to obtain the dynamic prediction results in the normalized space. Finally, the prediction results are subjected to de-standardization to restore the ground reaction force, ground reaction torque and pressure centerline time series in the physical quantity space, and the results are saved in chronological order.

[0046] Furthermore, the joint fine-tuning training adopts a two-stage joint fine-tuning approach;

[0047] In the first stage, labeled samples from the laboratory and high-confidence pseudo-labeled samples from the target domain are used to perform initial adaptation training on the pre-trained model, so that the model can gradually learn the changes in the input distribution in the real scene.

[0048] In the second stage, the model parameters obtained in the first stage are further fine-tuned to enhance the model's adaptability to changes in gait patterns and ground conditions in the target domain.

[0049] Furthermore, the pathological gait classification model module includes the following processes:

[0050] All participants wore the same prototype insoles as described above and completed a walking test in a pre-defined walking scenario; the participants included normal participants, participants with clinically diagnosed knee osteoarthritis, and participants with clinically diagnosed flat feet.

[0051] The raw insole data of each subject were preprocessed to make the data format and feature scale consistent with the real scene model;

[0052] The 22-dimensional temporal input features of each subject are directly input into the real scene model to generate gait dynamics reconstruction results for the corresponding samples; dynamic prediction results are obtained for the normal subject group, the knee osteoarthritis subject group, and the flat-foot subject group, respectively.

[0053] Multi-domain classification features are extracted from the time series of dynamic prediction results corresponding to the subject population;

[0054] Establish a pathological gait classification model to distinguish between normal gait, knee osteoarthritis gait, and flatfoot gait;

[0055] The classification model employs a two-layer stacking ensemble model, which includes a base learner layer and a meta-classifier layer. The base learner layer and the meta-classifier layer are connected sequentially in the direction of data flow. The base learner layer receives multi-domain classification features as input and outputs the predicted probability of a sample belonging to each category. The meta-classifier layer receives the predicted probabilities output by each base learner as input and outputs the final classification result.

[0056] The multi-domain classification features of the sample to be tested are input into the trained pathological gait classification model, and the final classification result of whether the sample belongs to normal gait, knee osteoarthritis gait, or flatfoot gait is output.

[0057] The beneficial effects achieved by this invention are:

[0058] This invention highly integrates sensors into the insole, enabling convenient "wear-and-use" deployment. This breaks the reliance on professional laboratories for GRF, GRM, and COP measurements, allowing for widespread application in real-world scenarios such as homes, gyms, and outdoor sports fields, greatly promoting the popularization of the technology. Furthermore, based on a GCN-LSTM spatiotemporal feature fusion deep learning model, it can learn complex GRF, GRM, and COP mapping relationships from the spatial topology and temporal dynamics of foot movement, ensuring intelligent and accurate predictions across various movement modes. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of a foot sensor in a wearable human gait analysis system provided by the present invention;

[0060] Figure 2 This invention provides a schematic diagram of the sensor array data acquisition and processing flow in a wearable human gait analysis system.

[0061] Figure 3 This invention provides a schematic diagram of the structure of the foot pad hardware system in a wearable human gait analysis system;

[0062] Figure 4 This invention provides a schematic diagram of plantar pressure distribution in a wearable human gait analysis system;

[0063] Figure 5 This invention provides a schematic diagram of representative plantar pressure distribution under genu valgum, normal and genu varum conditions in a wearable human gait analysis system;

[0064] Figure 6 This is a schematic diagram of the plantar pressure mapping results under different movement states in a wearable human gait analysis system provided by the present invention;

[0065] Figure 7This invention provides a schematic diagram of signal processing for an inertial measurement unit in a wearable human gait analysis system;

[0066] Figure 8 This invention provides a schematic diagram of plantar pressure topology modeling in a wearable human gait analysis system.

[0067] Figure 9 This invention provides a schematic diagram of a neural network in a wearable human gait analysis system;

[0068] Figure 10 This is a schematic diagram comparing the performance of a wearable human gait analysis system with other systems provided by the present invention. Detailed Implementation

[0069] The present invention will be further described below with reference to specific embodiments, and the advantages and features of the present invention will become clearer as a result. However, these embodiments are merely exemplary and do not constitute any limitation on the scope of the present invention. Those skilled in the art should understand that modifications or substitutions can be made to the details and form of the technical solutions of the present invention without departing from the spirit and scope of the present invention, but all such modifications and substitutions fall within the protection scope of the present invention.

[0070] like Figure 1 As shown, the present invention provides a wearable human gait analysis system, which includes a synchronous data acquisition module, a plantar pressure topology map module, a pre-trained model module, a real-scene model module, and a pathological gait classification model module.

[0071] The synchronous data acquisition module is designed for laboratory environments and consists of a high-precision force measuring platform, a prototype insole, and a host computer.

[0072] Subjects wore the prototype insole and completed a walking test on a high-precision force measuring platform. During the test, the high-precision force measuring platform collected real-time data on the three-dimensional ground reaction force, three-dimensional ground reaction torque, and pressure centerline during the subjects' movement, which served as standard label data for subsequent gait dynamics prediction model training and testing. Preferably, the insole includes 16 pressure-sensitive units, distributed in the main load-bearing areas of the foot, such as the heel, arch, first to fifth metatarsals, and big toe, to cover key pressure points on the sole.

[0073] The prototype insole uses a flexible printed circuit board (FPCB) as its carrier, and deploys multiple pressure-sensitive units in key load-bearing areas of the sole, such as the heel, arch, first to fifth metatarsals, and toes, to collect information on plantar pressure distribution. Simultaneously, an inertial measurement unit (IMU) is integrated into the prototype to synchronously collect kinematic information during gait. Preferably, the IMU collects triaxial acceleration and triaxial angular velocity signals, forming six channels of inertial measurement data to characterize the subject's posture changes and motion state during the gait cycle, thus creating a complementary multimodal input with the plantar pressure data.

[0074] To address the signal crosstalk and measurement errors caused by parallel leakage current that are prone to occur during data acquisition of array-type resistance sensors, the insole prototype employs a two-scan acquisition circuit. This circuit performs forward and reverse potential scans on the same sensing node, and calculates the resistance value of the corresponding sensing node based on the output voltage ratio Vout_1 / Vout_2 obtained from the two scans. This suppresses parallel path interference and enables independent and accurate measurement of the pressure signal from each sensing node. The main control module in the insole prototype further completes pressure array scanning, analog-to-digital conversion, resistance calculation, inertial measurement unit data reading, and sampling timing control.

[0075] Subjects were invited to perform multiple normal walking and other movement tests on a force measuring platform. During this process, the following data were collected simultaneously: real force measuring platform data, insole pressure time-series data, and inertial measurement unit data. Among them, the real force measuring platform data included three-dimensional ground reaction force, three-dimensional ground reaction torque, and two-dimensional pressure centerline coordinates.

[0076] The collected plantar pressure data and inertial measurement unit data are wirelessly transmitted to the host computer via a low-power Bluetooth module, and are synchronized and paired with the standard tag data collected by the high-precision force measuring platform to form a multimodal raw sample dataset.

[0077] The host computer preprocesses the multimodal raw sample dataset. The preprocessing includes at least the following: Butterworth low-pass filtering of the inertial measurement unit data; unified time axis alignment of the plantar pressure data, inertial measurement unit data, and force table label data; establishing a one-to-one correspondence between pressure signals, inertial signals, and label signals according to the sampling time; and standardizing the input features and output labels respectively.

[0078] Thus, a supervised learning dataset for model training and testing is constructed, in which each input sample includes synchronized plantar pressure time-series data and inertial measurement unit time-series data, and each label sample includes its corresponding ground reaction force, ground reaction torque and pressure centerline time-series data.

[0079] In this embodiment, the high-precision force stage is the FP5060-06-PT type high-precision force stage manufactured by Bertec Corporation of the United States, with a size of 49.5cm×60cm; the piezoresistive thin-film pressure sensor is the FSR400 type sensor.

[0080] The plantar pressure topology module represents the plantar pressure time-series data acquired from the insole prototype in a topological manner. Specifically, with the lower left corner of the insole as the origin of the two-dimensional coordinate system, a first coordinate axis is defined along the length of the insole, and a second coordinate axis is defined along the width of the insole. Based on the actual placement of each pressure sensor node in the insole, the spatial coordinates of the center of the effective acquisition area of ​​each sensor node are determined, and a standard foot sensor topology is constructed accordingly.

[0081] Each pressure-sensitive unit is abstracted as a graph node, thus forming a plantar pressure graph structure composed of 16 graph nodes, where each graph node corresponds to the spatial location of a pressure sensing node and its pressure response information.

[0082] After determining the spatial coordinates of each node in the graph, an adjacency matrix is ​​constructed based on the physical adjacency relationships between the sensor nodes. Physical adjacency refers to the following: when two pressure sensor nodes are spatially adjacent in the insole, or located in adjacent foot bearing areas and have a direct force transmission relationship, an edge is determined to exist between the two nodes; when the spatial distance between two pressure sensor nodes exceeds a preset threshold, or they do not belong to adjacent bearing areas, an edge is determined to exist between the two nodes. This yields the plantar pressure topology graph. ,in, Represents a set of nodes. This indicates the connection relationships between nodes. This represents the adjacency matrix generated based on physical adjacency relationships. The adjacency matrix is ​​used to characterize the spatial distribution characteristics and local coupling relationships of the plantar pressure sensor array, thereby avoiding treating each pressure channel as an independent ordinary vector.

[0083] Preferably, the standard foot sensor topology map is generated in a semi-automatic manner. First, the spatial coordinates of the center of the effective acquisition area of ​​each pressure sensing node are determined based on the insole layout design file, sensor arrangement template, or calibration results. Then, the connection relationships and adjacency matrix are automatically generated according to preset physical adjacency rules. For insoles with different shoe sizes, different numbers of sensing nodes, or different layout methods, only the corresponding node spatial coordinates and adjacency rules need to be updated to obtain the corresponding foot sensor topology map.

[0084] The pre-trained model module is a spatiotemporal fusion model used to learn the mapping relationship between wearable multimodal temporal signals and gait dynamics parameters. The spatiotemporal fusion model uses plantar pressure signals to characterize the plantar contact state and load transfer process, uses inertial measurement unit signals to supplement the posture and kinematic changes in the gait cycle, and outputs ground reaction force, ground reaction torque, and pressure centerline.

[0085] The spatiotemporal fusion model includes an input layer, a graph structure input construction layer, a graph convolutional layer, a pooling layer, a feature concatenation layer, a bidirectional long short-term memory layer, and a fully connected regression layer. Each layer is connected sequentially according to the data flow direction.

[0086] The input layer receives multimodal timing data after time synchronization processing. For any length of... Time series samples, in the first At each sampling time, the input data consists of two parts: the first part is 16 channels of plantar pressure signals, denoted as... The second part consists of six kinematic signals acquired by the inertial measurement unit, denoted as... The six kinematic signals preferably include triaxial acceleration signals and triaxial angular velocity signals. Therefore, the input features at a single sampling moment can be represented as a 22-dimensional vector with a length of... The input sample can be represented as by A 22-dimensional multimodal temporal feature sequence composed of sampling times.

[0087] The output of the spatiotemporal fusion model is a sequence of dynamic variables that corresponds one-to-one with the input time series. For the th At each sampling time, the output variable is denoted as... Preferably, the dynamic variables are 8-dimensional, including three-dimensional ground reaction force, three-dimensional ground reaction moment, and two-dimensional pressure centerline coordinates. Therefore, the length is... The output label sample can be represented as... An 8-dimensional dynamic time sequence composed of sampling times.

[0088] The graph structure input building layer is used to construct the plantar pressure topology graph. The input is converted into a graph structure suitable for spatial structure modeling. Specifically, based on the foot sensor topology graph, the 16 pressure sensing nodes are treated as graph nodes, and an adjacency matrix pre-constructed based on the physical adjacency relationships of the sensors is used. Represents the spatial connectivity between nodes. In the first... At each sampling time point, the 16 plantar pressure values ​​at that time are assigned to the corresponding graph nodes as node feature inputs, thus forming the graph structure input for that time point. The adjacency matrix remains fixed within the same insole structure to constrain the spatial propagation and local correlation modeling of the plantar pressure signals.

[0089] Graph convolutional layers are used to extract local spatial correlation features of plantar pressure signals from graph-structured inputs.

[0090] For any sampling time The 16 plantar pressure signals at that moment are assigned to the corresponding graph nodes to form the node feature matrix at that sampling moment. This generates the corresponding graph structure features. As the sampling time progresses continuously, multiple graph structure features arranged in chronological order are obtained sequentially, forming a graph sequence data. The graph sequence data simultaneously contains the spatial topological relationship and temporal evolution relationship of the plantar pressure signal, which can serve as input data for subsequent graph convolutional networks to extract spatial correlation features and temporal networks to extract dynamic dependency features.

[0091] For the At each sampling time, let the original node features be... The graph convolutional layer performs neighborhood aggregation and linear transformation on node features to obtain updated node representations. The calculation relationship is as follows:

[0092]

[0093] in, To add the adjacency matrix after self-connection, This is the degree matrix corresponding to the adjacency matrix. For the first Learnable weight parameters of layer graph convolutional layers For the first The bias parameters of the layer diagram convolutional layer. It is a nonlinear activation function. Through graph convolution, each pressure node retains its own pressure information while fusing the force information of its neighboring nodes, thereby extracting the local spatial correlation between different contact areas of the sole.

[0094] Pooling layers are connected after graph convolutional layers to compress node-level spatial features into time-level compressed spatial representations. Preferably, average pooling is performed on all node features output by the graph convolutional layer to obtain the first... Pressure space representation vector corresponding to each sampling time The pressure space representation vector is used to characterize the overall force distribution pattern of the sole of the foot and the local load transfer state at that moment.

[0095] The feature splicing layer, connected after the pooling layer, is used to fuse the spatial representation of plantar pressure with synchronous inertial information. In the first... Each sampling time point will represent the pressure space vector. Features of the inertial measurement unit acquired synchronously The features are concatenated along the feature dimension to form a fused feature vector. ,Right now:

[0096]

[0097] Arrange the fused feature vectors at each sampling time in chronological order to obtain a length of... fusion feature sequence .

[0098] A bidirectional long short-term memory (LSTM) layer is connected after the feature concatenation layer to learn the temporal dependencies of the fused feature sequence during the gait cycle. The LSM layer simultaneously performs forward and backward temporal modeling on the fused feature sequence to capture the dynamic relationships between time steps during gait, obtaining the temporal hidden state corresponding to each sampling moment. By using a bidirectional long short-term memory layer, the model can simultaneously utilize contextual information before and after the current moment to improve its ability to represent continuous dynamic changes such as load acceptance, support propulsion, and liftoff transitions during the gait cycle.

[0099] A fully connected regression layer, connected after a bidirectional long short-term memory layer, maps the temporal hidden state corresponding to each sampling time step to the dynamic prediction result at that time step. Specifically, it maps the temporal hidden state... Input a fully connected regression layer and output the dynamic prediction value at the corresponding sampling time. Arrange the dynamic prediction values ​​at all sampling times in chronological order to obtain a continuous dynamic prediction sequence of the same length as the input sequence.

[0100] The training process for the spatiotemporal fusion model is as follows:

[0101] First, supervised training samples are constructed based on synchronously acquired data. Specifically, 16 channels of plantar pressure data acquired from the insole prototype are synchronized with 6 channels of inertial measurement unit data, and then matched with the true values ​​of ground reaction force, ground reaction torque, and pressure centerline acquired by the high-precision force measuring platform.

[0102] Secondly, the input features and output labels are standardized. Specifically, the distribution parameters of each input dimension and each output dimension in the training data are statistically analyzed, and the 22-dimensional temporal features of the input and the 8-dimensional dynamic labels of the output are normalized accordingly to reduce the impact of differences in scale and amplitude on the stability of model training. The normalized parameters are saved after training and are consistently used in subsequent validation, testing, and inference processes.

[0103] Next, the standardized continuous time series data is divided into preset time lengths. The dataset is divided into multiple training samples. For each training sample, 16 plantar pressure sequences and 6 inertial measurement unit sequences within the corresponding time window are used as model inputs, and the corresponding 8-dimensional dynamic variable sequences within the same time window are used as supervision labels. This forms an input sample set and a label sample set suitable for batch training.

[0104] Subsequently, each training sample is input into the spatiotemporal fusion model for forward computation. Specifically, at each sampling time, 16 plantar pressure signals are combined with a pre-constructed adjacency matrix and input into a graph convolutional layer. After the graph convolutional layer extracts local spatial correlation features, a pooling layer is used to obtain a time-level pressure spatial representation. The pressure spatial representation is then concatenated with the features of 6 inertial measurement units at the same sampling time to form a fusion feature vector. The fusion feature vectors corresponding to each sampling time within the entire time window are combined into a fusion feature sequence and input into a bidirectional long short-term memory layer for temporal modeling. Finally, a fully connected regression layer outputs a dynamic prediction sequence that corresponds one-to-one with each sampling time within the time window.

[0105] Furthermore, a supervised training objective is constructed based on the error between the predicted gait sequence and the corresponding ground truth label sequence, and the parameters of the spatiotemporal fusion model are updated. The error is preferably the overall statistic of the regression error at each time step and each output dimension. During training, the output error is propagated layer by layer to the fully connected regression layer, the bidirectional long short-term memory layer, and the graph convolutional layer through the backpropagation algorithm, jointly updating the learnable weight parameters and bias parameters in each layer, thereby gradually optimizing the model's ability to fit the mapping relationship between multimodal inputs and gait dynamics outputs.

[0106] During the training iteration, all training samples are input into the model in batches to repeatedly perform forward computation, error calculation, and parameter updates until the model converges or reaches the preset training termination condition. Preferably, validation data is further set during the training process to monitor the changes in the model's prediction error on data that have not participated in parameter updates, and to determine the timing of training termination and save the model parameters with the best prediction performance.

[0107] After training, the network parameters of the spatiotemporal fusion model and the corresponding input-output normalized parameters are saved. During testing or actual inference, the 16 real-time acquired plantar pressure signals and 6 inertial measurement unit signals are first normalized in the same way as during training. Then, they are input into the trained spatiotemporal fusion model to obtain the dynamic prediction results in the normalized space. Finally, the prediction results are denormalized to restore the continuous time series of ground reaction force, ground reaction moment, and pressure centerline in the physical quantity space.

[0108] Through the above structure and training process, the spatiotemporal fusion model realizes the mapping from the spatial distribution of plantar pressure and inertial kinematic information to continuous gait dynamic variables. Among them, the graph convolutional layer is used to extract the spatial correlation features of the plantar pressure signal, the bidirectional long short-term memory layer is used to extract the temporal dependence features of the fusion features, and the fully connected regression layer is used to output continuous dynamic results, thereby completing the reconstruction of ground reaction force, ground reaction moment and pressure centerline.

[0109] The real-scene model module performs semi-supervised adaptive fine-tuning of the pre-trained model based on real-scene scenarios with high-confidence pseudo-labels.

[0110] To improve the model's generalization ability in real-world scenarios, a semi-supervised adaptive fine-tuning of the target domain was performed on the pre-trained laboratory model. Gait patterns in real-world environments differ from those under controlled laboratory conditions. Changes in ground material and slope alter the foot-ground contact process, leading to posture adjustments and gait temporal variations, thus reducing the predictive performance of the model trained solely on labeled laboratory data in target scenarios. Therefore, a semi-supervised adaptive strategy based on high-confidence pseudo-labels was adopted. Pre-training was first performed on labeled laboratory data, and then pseudo-labels were generated using unlabeled samples from real-world scenarios. These pseudo-labels were then jointly fine-tuned with source domain samples to enhance the model's adaptability in real-world scenarios.

[0111] The target domain input is consistent with the laboratory stage, including 6 inertial measurement unit (IMU) signals and 16 plantar pressure signals, totaling 22 temporal features. The model output remains a sequence of dynamic variables corresponding to each time step of the input time series, preferably including ground reaction force, ground reaction moment, and pressure centerline. During the semi-supervised adaptive fine-tuning process, the model network structure remains consistent with the spatiotemporal fusion model in the pre-trained model module. That is, it still uses graph convolutional layers to extract the spatial correlation features of the plantar pressure signals, concatenates them with the synchronously acquired IMU features, and inputs them into a bidirectional long short-term memory network for temporal dependency modeling. Finally, the dynamic prediction results for each time step are output through a fully connected regression layer.

[0112] The semi-supervised adaptive fine-tuning of the real-world model module includes six steps: target domain sample reading and preprocessing, pseudo-label generation, uncertainty calculation, high-confidence sample screening, joint fine-tuning training, and target domain inference.

[0113] Read gait time-series samples based on smart footpads in real-world scenarios within the target domain. Target domain samples are preferably sourced from real-world ground conditions different from laboratory environments, including but not limited to outdoor hard surfaces, grass, and uphill and downhill scenarios. The target domain samples are read into files one by one, maintaining the original order of the time series within each file to preserve complete and continuous gait dynamic information.

[0114] Secondly, the input samples in the target domain are subjected to the same input standardization process as in the laboratory stage. Specifically, the input normalizer saved after the laboratory stage training is called, and the same normalization transformation is performed on the 16 plantar pressure signals and 6 inertial measurement unit signals in each target domain sample. This ensures that the target domain input and the laboratory source domain input are consistent in feature scale, avoiding the impact of distribution scale differences on subsequent pseudo-label generation and model fine-tuning.

[0115] Monte Carlo Dropout inference is used to generate pseudo-labels for the target domain. Specifically, a pre-trained lab model is loaded, and Dropout units in the model are kept active during the inference phase. For the... Target domain samples Input it into the pre-trained model for processing The next random forward propagation, preferably... . No. The prediction result of the random forward propagation is denoted as:

[0116]

[0117] in, Indicates the first Pre-trained model under sub-random inactivation conditions This represents the corresponding dynamic prediction time series. Pseudo-labels for the target domain samples are generated based on the mean of each prediction result.

[0118]

[0119] in, For the first Pseudo-label time series of samples from the target domain.

[0120] The uncertainty of the target domain samples is calculated. Let the length of the output sequence for a single sample be... The output dimension at each time step is Preferably, For the first The target domain sample, at the _th ... The time step, the first The prediction standard deviation over each output dimension is denoted as:

[0121]

[0122] Then, the mean of the prediction standard deviations over all time steps and all output dimensions is calculated to obtain the average uncertainty index for the target domain sample:

[0123]

[0124] in, Used to characterize the The stability and reliability of pseudo-label prediction results for target domain samples.

[0125] After obtaining the average uncertainty index, a high-confidence screening is performed on the target domain samples. Specifically, an uncertainty threshold is preset. If and only if:

[0126]

[0127] At that time, the first one is reserved. Target domain samples and their corresponding pseudo-labels As a pseudo-label sample for a high-confidence target domain; when If a sample from that target domain is discarded, it will not be included in subsequent fine-tuning training. This screening process reduces the risk of error propagation caused by low-confidence pseudo-labels in model parameter updates, allowing adaptive learning to utilize only target domain samples with high prediction stability.

[0128] After obtaining high-confidence target domain pseudo-label samples, joint fine-tuning training is performed. Specifically, the labeled source domain samples from the laboratory phase are merged with the high-confidence target domain pseudo-label samples to construct a joint training dataset. The source domain samples include 22-dimensional input temporal features and their corresponding real dynamic labels for each time step, while the target domain samples include 22-dimensional input temporal features and high-confidence pseudo-labels generated through Monte Carlo Dropout inference. Normalization is then performed on all inputs and outputs in the joint training dataset to maintain consistency between the feature space and label space during the laboratory and real-world training phases.

[0129] Joint fine-tuning training uses the parameters of the pre-trained laboratory model as initialization parameters. Specifically, the parameters of the graph convolutional layer, bidirectional long short-term memory layer, and fully connected regression layer saved after training in the pre-trained model module are loaded as the initial network weights for adaptive training in the target domain, instead of using random initialization. By preserving the fundamental mapping relationship between the multimodal inputs and dynamic outputs learned in the laboratory stage, the convergence efficiency and stability of fine-tuning in real-world scenarios can be improved.

[0130] The joint fine-tuning training employs a two-stage joint fine-tuning approach. In the first stage, labeled laboratory samples and high-confidence pseudo-labeled samples from the target domain are used to initially adapt the pre-trained model, enabling the model to gradually learn the input distribution changes in real-world scenarios. In the second stage, fine-tuning continues based on the model parameters obtained in the first stage to further enhance the model's adaptability to changes in gait patterns and ground conditions in the target domain. During the two-stage fine-tuning process, the model's data flow remains consistent with the pre-trained model module: First, based on the plantar pressure topology map and adjacency matrix constructed in the plantar pressure topology map module, the 16 plantar pressure signals at each moment are represented as graph node features and input into a graph convolutional layer to extract local spatial correlations. Then, average pooling is performed on the graph convolution output to obtain the pressure spatial representation at that moment. Subsequently, this pressure spatial representation is concatenated with the features from 6 synchronously acquired inertial measurement units to form a fused feature sequence. The fused feature sequence is then input into a bidirectional long short-term memory network for temporal modeling. Finally, a fully connected regression layer outputs dynamic prediction results corresponding to each sampling moment.

[0131] During joint fine-tuning training, the true dynamic labels of source domain samples are used as supervision signals; for target domain samples, the corresponding high-confidence pseudo-labels are used as supervision signals. For any input time series, the predicted dynamic sequence is obtained after forward computation by the model. The regression error is then calculated using the source domain true label or the target domain pseudo label, and the model parameters are updated through backpropagation. Preferably, the following joint regression loss function can be used:

[0132]

[0133] in, This indicates the supervised regression loss on labeled samples in the source domain. This represents the pseudo-supervised regression loss on high-confidence pseudo-labeled samples in the target domain. The loss coefficients are set to the target domain loss coefficients. By jointly optimizing the above loss functions, the model can gradually adapt to the signal distribution and gait changes of the real scene in the target domain while retaining the source domain dynamic mapping capability.

[0134] After completing the two-stage joint fine-tuning, save the model parameters and corresponding normalizers after the target domain adaptation.

[0135] When entering the target domain inference stage, the fine-tuned model and its corresponding normalizer are loaded, and the real scene data is predicted one by one according to the file. First, the input samples are subjected to the same standardization process as in the training stage. Then, the 22-dimensional input time series features are input into the fine-tuned model to obtain the dynamic prediction results in the normalized space. Finally, the prediction results are subjected to de-standardization to restore the ground reaction force, ground reaction torque and pressure centerline time series in the physical quantity space, and the results are saved in chronological order.

[0136] Through the above-mentioned semi-supervised adaptive fine-tuning process based on high-confidence pseudo-labels in real-world scenarios, model transfer can be completed using labeled laboratory samples and unlabeled samples in real-world scenarios, even in the absence of manual annotation of the target domain. This enables the spatiotemporal fusion model to stably reconstruct ground reaction force, ground reaction moment, and pressure centerline in real-world environments such as outdoor hard surfaces, grass, uphill and downhill slopes, while maintaining good consistency.

[0137] The pathological gait classification model module reconstructs gait dynamics from raw walking data collected from normal subjects, subjects with knee osteoarthritis, and subjects with flat feet under the condition of wearing insoles. Based on this, it constructs gait dynamics datasets for the three groups and further outputs the final classification results of whether the sample belongs to normal gait, knee osteoarthritis gait, or flat foot gait.

[0138] The pathological gait classification model module consists of five parts: pathological data acquisition experiment, pathological gait dynamics data generation, classification feature construction, classification model training, and model inference output.

[0139] Pathological data acquisition experiment

[0140] First, a walking experiment was conducted with three groups of participants, preferably including normal subjects, subjects with clinically diagnosed knee osteoarthritis, and subjects with clinically diagnosed flat feet. All participants wore the same prototype insole as described above and completed the walking test in a pre-defined walking scenario.

[0141] Preferably, the walking scenario includes walking on flat indoor ground. Subjects can walk continuously at a natural pace or a preset pace to obtain stable and continuous gait data. Each subject can complete multiple repeated walking trials to improve the richness and stability of subsequent modeling samples.

[0142] During the walking experiment, the insole prototype simultaneously collected 16 channels of plantar pressure time-series data and 6 channels of inertial measurement unit time-series data to form 22-dimensional original walking input features.

[0143] To facilitate subsequent supervised modeling and classification evaluation, a corresponding subject ID and category label were recorded for each subject sample. The category label was assigned based on the subject's group: normal subjects received a normal gait label, subjects with knee osteoarthritis received a knee osteoarthritis gait label, and subjects with flat feet received a flat foot gait label. This yielded a raw dataset of pathological gait with subject group information.

[0144] Preprocessing of raw pathological gait data

[0145] After completing the pathological data collection experiment, the raw insole data of each subject were preprocessed. The preprocessing process was consistent with that of the pre-trained model module and the real-world scenario model module to ensure consistency in the input data format and feature scale.

[0146] Generation of pathological gait dynamics data based on real-world scene models:

[0147] After preprocessing, the 22-dimensional temporal input features are directly input into the real scene model to generate gait dynamics reconstruction results for the corresponding samples.

[0148] Specifically, for any sample to be analyzed, its normalized 16-channel plantar pressure time series data and 6-channel inertial measurement unit time series data are read, and the model outputs the gait dynamics prediction results corresponding to each moment of the input time series.

[0149] Preferably, the gait dynamics prediction results include three-dimensional ground reaction force, three-dimensional ground reaction torque, and two-dimensional pressure centerline coordinates, thus forming an 8-dimensional gait dynamics time series. For each walking test of each subject, a corresponding set of continuous gait dynamics waveforms can be obtained.

[0150] Thus, gait dynamics data were obtained for the normal subject group, the knee osteoarthritis subject group, and the flat-foot subject group, respectively.

[0151] Construction of pathological gait classification features

[0152] After obtaining gait dynamics data from the three groups, gait dynamics time series were used as classification input instead of raw plantar pressure signals and inertial measurement unit signals, thereby improving the biomechanical interpretability and pathological characterization of classification features.

[0153] Specifically, multi-domain classification features are extracted from the time series corresponding to the three-dimensional ground reaction force, the three-dimensional ground reaction moment, and the two-dimensional pressure centerline. These multi-domain classification features preferably include statistical features, temporal morphological features, key point features, frequency domain features, wavelet energy features, and nonlinear dynamic features.

[0154] Among them, statistical features are used to characterize the overall distribution information of the dynamic waveform, such as mean, peak value, standard deviation, coefficient of variation, integral area, and peak-to-valley difference; temporal morphological features are used to characterize the duration and waveform change trend of the load acceptance phase, support phase, and off-ground propulsion phase; key point features are used to characterize local peaks, valleys, and their relative positions in the gait cycle; frequency domain features are used to characterize the differences in gait rhythm and frequency component distribution under different pathological conditions; wavelet energy features are used to characterize the energy distribution law of the waveform at different time scales; and nonlinear dynamic features are used to characterize the complexity, stability, and dynamic change characteristics of the gait dynamic waveform.

[0155] Through the above feature extraction, the gait dynamics data of the three groups are further converted into a pathological gait classification feature set suitable for pattern recognition.

[0156] Training of pathological gait classification model

[0157] After obtaining the pathological gait classification feature set, a classification model was established to distinguish between normal gait, knee osteoarthritis gait, and flatfoot gait.

[0158] Preferably, the classification model employs a two-layer stacking ensemble classification model. This two-layer stacking ensemble classification model includes a base learner layer and a meta-classifier layer, which are sequentially connected in the data flow direction. The base learner layer receives multi-domain classification features as input and outputs the predicted probability of a sample belonging to each category; the meta-classifier layer receives the predicted probabilities output by each base learner as input and outputs the final classification result.

[0159] Preferably, the base learner layer includes one or more of the following: HistGradientBoosting classifier, XGBoost classifier, RandomForest classifier, and Radial Basis Kernel Support Vector Machine classifier; the meta-classifier layer preferably uses a Logistic Regression classifier. By combining multiple base learners, complementary pathological gait discrimination information can be extracted from different feature spaces and different discrimination mechanisms, thereby improving the overall classification performance and robustness.

[0160] To avoid data leakage caused by different samples from the same subject appearing simultaneously in both the training and validation sets, it is preferable to group the subjects according to their subject IDs and use GroupKFold cross-validation for model training and validation. During training, automatic hyperparameter optimization methods can be further incorporated to search for relevant parameters of each base learner and meta-classifier to obtain parameter combinations with superior overall performance.

[0161] After training, save the pathological gait classification model and its corresponding feature processing flow.

[0162] Pathological gait inference output

[0163] During the model inference phase, for any test sample, 16 channels of plantar pressure time-series data and 6 channels of inertial measurement unit time-series data are first collected under the condition of wearing insoles. Then, the data is preprocessed and standardized in the same way as in the training phase. The processed 22-dimensional time-series input features are then input into the real scene model to obtain the gait dynamics prediction result corresponding to the sample. Further, multi-domain classification features are extracted. Finally, the classification features are input into the trained pathological gait classification model to output the final classification result of whether the sample belongs to normal gait, knee osteoarthritis gait, or flatfoot gait.

[0164] Preferably, the final output includes not only the category label of the sample, but also the reconstructed gait dynamics curve and the characteristic parameters related to each pathological category, thereby providing a basis for abnormal gait screening, auxiliary judgment of lower limb diseases and rehabilitation assessment.

[0165] like Figure 10 As shown, as a comparative case, under the same training dataset, test dataset, input data format, output label format, and evaluation metrics, the spatiotemporal fusion model used in this invention is compared and evaluated with existing common temporal or deep learning models. The comparison models include CNN, ResNet, Transformer, LSTM, and GCN; the evaluation metric used is the correlation coefficient. and root mean square error RMSE, where, A higher RMSE indicates greater consistency between the generated ground reaction isodynamic curve and the true curve in terms of morphological changes and temporal response; a lower RMSE indicates smaller amplitude error between the generated curve and the true curve. Based on comparison and ablation results ( Figure 10 When the scheme of this invention is adopted, higher correlation and lower error are observed in each output component. When the spatial modeling module or the temporal modeling module is removed, the dynamic reconstruction performance decreases significantly. However, when the two are combined, the model achieves the best and more stable prediction results in each output component. This indicates that the synergy between spatial relationship modeling and temporal dependency modeling in the scheme of this invention is the key to improving the performance of high-level gait dynamic reconstruction.

[0166] The comparative examples above demonstrate that the method of this invention has the following data-level advantages over existing technologies in generating ground reaction force curves: First, the generated curves have a higher correlation with the standard true curves, indicating that the model can more intelligently learn the complex nonlinear mapping relationship between plantar pressure, inertial information, and gait dynamics; second, the generated curves have lower root mean square errors, indicating that the model has higher accuracy in curve amplitude reconstruction; third, the model output exhibits better stability across multiple dynamic components, demonstrating that this invention has good generalization consistency across different output channels. These results corroborate the beneficial effects of this invention, namely, its intelligence and accuracy.

[0167] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the scope of protection of the present invention; all technical solutions formed by equivalent transformations or equivalent substitutions fall within the scope of protection of the present invention; the parts of the present invention not described in detail are well-known technologies to those skilled in the art.

Claims

1. A wearable human gait analysis system, characterized in that, The wearable human gait analysis system includes a synchronous data acquisition module, a plantar pressure topology map module, a pre-trained model module, a real-scene model module, and a pathological gait classification model module. The synchronous data acquisition module includes a high-precision force measuring platform, a prototype insole, and a host computer. Subjects wear the prototype insole and perform multiple walking motion tests on the high-precision force measuring platform. The host computer collects real force measuring platform data, insole pressure time-series data, and inertial measurement unit data to construct a gait dataset. The plantar pressure topology graph module performs topological representation of insole pressure time series data and outputs the corresponding plantar pressure topology graph based on the insole pressure time series data; The pre-trained model module constructs a pre-trained model based on the gait dataset and spatiotemporal feature fusion network outputting gait dynamics prediction results; The real-scene model module performs semi-supervised adaptive fine-tuning of the pre-trained model in the target domain to generate a real-scene model. The pathological gait classification model module extracts multi-domain classification features based on the prediction results of the real-world scenario model for each pathological subject and outputs the final classification result.

2. The wearable human gait analysis system according to claim 1, characterized in that, The prototype insole includes 16 pressure-sensitive units distributed across the main load-bearing areas of the foot, including the heel, arch, first to fifth metatarsals, and big toe, to cover key pressure points on the sole of the foot.

3. The wearable human gait analysis system according to claim 1, characterized in that, The plantar pressure topology graph module abstracts each pressure-sensitive unit in the insole prototype as a graph node, thus forming a plantar pressure graph structure composed of 16 graph nodes; and constructs an adjacency matrix based on the physical adjacency relationships between the sensing nodes to obtain the plantar pressure topology graph. ,in, Represents a set of nodes. This indicates the connection relationships between nodes. This represents the adjacency matrix generated based on physical adjacency relationships.

4. The wearable human gait analysis system according to claim 1, characterized in that, The pre-trained model module includes an input layer, a graph structure input construction layer, a graph convolutional layer, a pooling layer, a feature concatenation layer, a bidirectional long short-term memory layer, and a fully connected regression layer. Each layer is connected sequentially according to the data flow direction. The input layer combines 16 channels of plantar pressure signals from the insole pressure time-series data at each sampling time and 6 channels of kinematic signals from the inertial measurement unit data into a 22-dimensional vector; The graph structure input layer constructs a topology map of plantar pressure from 16 pressure sensor nodes. As input for node features; Graph convolutional layers are used to extract local spatial correlation features of plantar pressure signals from graph-structured inputs; For any sampling time Based on the plantar pressure topology at that moment Forming the node feature matrix at the sampling time This generates the corresponding graph structure features. ; Pooling layers are used to compress node-level spatial features into time-level pressure space representations; The feature splicing layer is used to fuse pressure space representation and synchronous inertial information to generate fused features; A bidirectional long short-term memory layer is used to learn the temporal dependencies of fused feature sequences in the gait cycle; The fully connected regression layer maps the temporal hidden state corresponding to each sampling moment to the dynamic prediction result at that moment. The dynamic prediction result includes three-dimensional ground reaction force, three-dimensional ground reaction moment and two-dimensional pressure centerline coordinates.

5. The wearable human gait analysis system according to claim 4, characterized in that, The training process for the pre-trained model module includes: First, the 16-channel plantar pressure data collected from the prototype insole are synchronized with the 6-channel inertial measurement unit data, and then matched with the three-dimensional ground reaction force, three-dimensional ground reaction moment and two-dimensional pressure centerline coordinates in the actual force table data. Secondly, the distribution parameters of each input dimension and each output dimension in the training data are statistically analyzed, and the 22-dimensional temporal features of the input and the 8-dimensional dynamic labels of the output are normalized accordingly to reduce the impact of different dimensions and amplitude differences on the stability of model training. Next, the standardized continuous time series data is divided into preset time lengths. The training samples were divided into multiple training samples. For each training sample, 16-channel plantar pressure sequence and 6-channel inertial measurement unit sequence within the corresponding time window were taken as model input, and the corresponding 8-dimensional real force table data sequence within the same time window was taken as supervision label. Subsequently, each training sample is input into the spatiotemporal fusion model to perform forward computation; Finally, a supervised training objective is constructed based on the error between the dynamic prediction sequence and the corresponding ground truth label sequence, and the parameters of the spatiotemporal fusion model are updated.

6. The wearable human gait analysis system according to claim 1, characterized in that, The semi-supervised adaptive fine-tuning of the real-world scene model module includes the following process: Read gait time series samples based on smart foot pads in real-world scenarios within the target domain; read the target domain samples one file at a time, maintaining the original order of the time series within each file to preserve complete continuous dynamic gait information; Monte Carlo Dropout inference is used to generate pseudo-labels for the target domain; for the first... Target domain samples Input it into the pre-trained model for processing The second random forward propagation, the... The prediction result of the random forward propagation is denoted as: ； in, Indicates the first Pre-trained model under sub-random inactivation conditions This represents the corresponding dynamic prediction time series; pseudo-labels for the target domain samples are generated based on the mean of each prediction result. ； in, For the first Pseudo-label time series of samples from the target domain; The uncertainty of the target domain samples is calculated; let the length of the output sequence for a single sample be... The output dimension at each time step is , for the The target domain sample, at the _th ... The time step, the first The prediction standard deviation over each output dimension is denoted as: ； in, Indicates the first The target domain samples in the th... In the second random forward propagation, at the... The time step, the first Predicted values ​​on each output dimension; Indicates the same sample in the first... The predicted value at the corresponding position under the next random forward propagation, where , is used to represent the average of all random forward propagation results; Then, the mean of the prediction standard deviations over all time steps and all output dimensions is calculated to obtain the average uncertainty index for the target domain sample: ； in, Used to characterize the The stability and reliability of the pseudo-label prediction results for each target domain sample; After obtaining the average uncertainty index, a high-confidence screening is performed on the target domain samples, and an uncertainty threshold is preset. If and only if At that time, the first one is reserved. Target domain samples and their corresponding pseudo-labels As a pseudo-label sample for a high-confidence target domain; when When this happens, the samples in that target domain are discarded and will not participate in subsequent fine-tuning training; The labeled source domain samples from the laboratory stage are merged with the pseudo-labeled target domain samples with high confidence to construct a joint training dataset. When entering the target domain inference stage, the fine-tuned model and its corresponding normalizer are loaded, and the real scene data is predicted one by one according to the file. First, the input samples are subjected to the same standardization process as in the training stage. Then, the 22-dimensional input time series features are input into the fine-tuned model to obtain the dynamic prediction results in the normalized space. Finally, the prediction results are subjected to de-standardization to restore the ground reaction force, ground reaction torque and pressure centerline time series in the physical quantity space, and the results are saved in chronological order.

7. The wearable human gait analysis system according to claim 6, characterized in that, The joint fine-tuning training adopts a two-stage joint fine-tuning approach; In the first stage, labeled samples from the laboratory and high-confidence pseudo-labeled samples from the target domain are used to perform initial adaptation training on the pre-trained model, so that the model can gradually learn the changes in the input distribution in the real scene. In the second stage, the model parameters obtained in the first stage are further fine-tuned to enhance the model's adaptability to changes in gait patterns and ground conditions in the target domain.

8. The wearable human gait analysis system according to claim 1, characterized in that, The pathological gait classification model module includes the following workflow: All participants wore the same prototype insoles as described above and completed a walking test in a pre-defined walking scenario; the participants included normal participants, participants with clinically diagnosed knee osteoarthritis, and participants with clinically diagnosed flat feet. The raw insole data of each subject were preprocessed to make the data format and feature scale consistent with the real scene model; The 22-dimensional temporal input features of each subject are directly input into the real scene model to generate the gait dynamics reconstruction results of the corresponding samples; Kinetic prediction results were obtained for normal subjects, subjects with knee osteoarthritis, and subjects with flat feet, respectively. Multi-domain classification features are extracted from the time series of dynamic prediction results corresponding to the subject population; Establish a pathological gait classification model to distinguish between normal gait, knee osteoarthritis gait, and flatfoot gait; The classification model employs a two-layer stacking ensemble model, which includes a base learner layer and a meta-classifier layer. The base learner layer and the meta-classifier layer are connected sequentially in the direction of data flow. The base learner layer receives multi-domain classification features as input and outputs the predicted probability of a sample belonging to each category. The meta-classifier layer receives the predicted probabilities output by each base learner as input and outputs the final classification result. The multi-domain classification features of the sample to be tested are input into the trained pathological gait classification model, and the final classification result of whether the sample belongs to normal gait, knee osteoarthritis gait, or flatfoot gait is output.

Images