A gait plantar pressure intelligent prediction system based on Transformer structure

Abstract

The present application relates to the technical field of gait analysis, and discloses a gait plantar pressure intelligent prediction system based on a Transformer structure, which comprises a gait collection platform, a plantar pressure collection device, a gait kinematics collection device, a central processing server and a display terminal. The system obtains initial plantar pressure distribution through a lower limb multi-rigid-body chain biomechanical model and foot-ground contact mechanics solution; and constructs a gated prefix memory sliding window Transformer in the central processing server, performs time series modeling and spatial correction on the initial pressure matrix and gait characteristics, and generates plantar pressure prediction results conforming to the biomechanical law. The system effectively solves the problems of unstable prediction, weak generalization ability and lack of interpretability of traditional methods, realizes higher-precision plantar pressure prediction, and can be used for gait evaluation, rehabilitation training and orthosis customization.

Description

Technical Field

[0001] This invention relates to the field of gait analysis technology, and in particular to an intelligent prediction system for plantar pressure based on a Transformer structure. Background Technology

[0002] Current plantar pressure analysis methods largely rely on direct measurement using pressure plates or indirect prediction based on traditional statistical learning models. However, purely data-driven methods (such as CNNs and RNNs) lack the ability to model dynamic changes in gait phases and have significant limitations in cross-individual generalization, identification of abnormal pressure peaks, and stability under complex gait conditions. Meanwhile, while traditional biomechanical inversion methods have a physical basis, they are computationally complex, parameter-sensitive, and poorly adaptable to rapid changes within the gait cycle, failing to meet the dual requirements of real-time performance and accuracy in clinical gait laboratories. With breakthroughs in temporal modeling using attention mechanisms and specific computational models such as Transformers, researchers have begun to explore their application to plantar pressure prediction. However, existing methods generally lack the ability to jointly model gait structural features, temporal dependence patterns, and spatial distribution of plantar pressure, making it difficult to obtain temporally consistent, spatially reliable, and clinically interpretable prediction results. Therefore, there is an urgent need for an intelligent plantar pressure prediction system based on the Transformer architecture to improve the accuracy, stability, and adaptability to gait changes in plantar pressure prediction, providing more reliable technical support for gait assessment and rehabilitation applications. Summary of the Invention

[0003] This invention proposes an intelligent gait plantar pressure prediction system based on the Transformer structure, aiming to address the problems of insufficient gait phase modeling, unstable prediction results, and lack of interpretability in existing plantar pressure prediction methods. The system deeply integrates a lower limb multi-rigid-body chain biomechanical model with specific computational models (neural networks, attention mechanisms, and Transformers with prefix memory structures) to achieve structured, temporal, and interpretable prediction of plantar pressure distribution. The system first acquires plantar pressure and gait kinematic data using a gait acquisition platform, constructing an initial biomechanical pressure distribution that includes joint reaction forces, segmental postures, and foot-ground contact characteristics. Based on this, a gated prefix memory sliding window Transformer is introduced to perform temporal dependency modeling, local spatial correlation analysis, and nonlinear correction on the initial pressure matrix and gait features, generating pressure prediction results that are more consistent with biomechanical principles and possess adaptive capabilities. Compared with existing technologies, this invention significantly enhances the system's adaptability to dynamic characteristics of gait phases, nonlinear pressure patterns, and cross-individual differences by utilizing a specific computational model. It achieves high-precision, high-stability, and interpretable plantar pressure prediction, which can be widely applied to gait assessment, rehabilitation training, and orthotic customization.

[0004] This invention provides a gait plantar pressure intelligent prediction system based on the Transformer structure, the system comprising:

[0005] A plantar pressure acquisition device collects raw plantar pressure data from localized areas of the foot.

[0006] Gait kinematics acquisition device to record gait kinematic data related to lower limb joint angles, stride length, stride frequency, and gait cycle;

[0007] The central processing server, deployed in the gait laboratory computer room, processes the raw plantar pressure data and gait kinematic data, generates a corrected plantar pressure distribution prediction matrix, and converts it into a pseudo-color pressure distribution map, regional pressure statistical indicators, and risk score results, which are then sent to the display and interactive terminal.

[0008] Display and interactive terminals, set up at doctors' workstations and rehabilitation training area terminals, include displays and human-computer interaction interfaces to present the predicted results of plantar pressure distribution.

[0009] Furthermore, the central processing server includes: a data preprocessing module, a feature extraction module, a gait biomechanical modeling module, a plantar contact mechanics solving module, a plantar pressure depth correction module, and a result generation module;

[0010] The data preprocessing module performs noise reduction filtering, time synchronization and coordinate system unification on the raw plantar pressure data and gait kinematic data, and completes gait cycle segmentation according to the walking event recognition strategy to obtain cleaned and aligned basic gait data.

[0011] The feature extraction module extracts plantar pressure distribution features, gait temporal features, joint kinematic features, and spatial posture features based on gait baseline data to construct structured gait feature data;

[0012] The gait biomechanical modeling module collects anthropometric parameters of the human lower limbs and feet. Based on the anthropometric parameters and structured gait feature data, it constructs a multi-rigid-body chain gait biomechanical model that includes the pelvis, thigh, calf and multi-segment foot, and calculates the reaction forces and moments of each joint to obtain the initial plantar load distribution characteristics.

[0013] The plantar contact mechanics solution module establishes a plantar contact mechanics model based on the joint reaction forces, segmental postures, and a set of biomechanical parameters including segmental mass, center of mass position, moment of inertia, and contact element stiffness and damping parameters output by the multi-rigid-body chain gait biomechanical model. In the plantar contact mechanics model, the rigid body segments of the hindfoot, midfoot, and forefoot are treated as independent force-bearing units. The spatial position and normal direction of each contact unit are determined by the segmental posture, and the instantaneous reaction forces of the hip, knee, and ankle joints are transmitted to the foot segments along the multi-rigid-body chain. According to the foot-ground contact equilibrium equation, the vertical force, shear force, and segmental rotational demand of the segments are distributed to the corresponding equivalent spring-damped contact elements. The contact force of each contact element is obtained by solving the normal deformation and tangential damping response of the contact elements. Based on the initial pressure distribution of each region of the foot obtained by the solution, it is restructured according to the discrete mesh of the foot to form an initial plantar pressure prediction matrix consistent with the geometry of the foot.

[0014] The plantar pressure depth correction module constructs a gated prefix memory sliding window Transformer model. This model performs temporal and spatial correlation modeling on the initial plantar pressure prediction matrix and structured gait feature data. A nonlinear mapping mechanism is used to correct the initial pressure distribution, outputting a corrected plantar pressure distribution prediction matrix. The construction process of the gated prefix memory sliding window Transformer model is as follows: based on the Transformer model, a prefix memory gating and sliding window bias modulation mechanism is introduced. Learnable temporal decay and enhancement are applied to the attention score, optimizing the temporal dependency modeling and local spatial sensitivity of the Transformer model, thus constructing the gated prefix memory sliding window Transformer model.

[0015] The results generation module is used to convert the corrected plantar pressure distribution prediction matrix into a pseudo-color pressure distribution map, regional pressure statistical indicators and risk score results, and send them to the display and interactive terminal for display.

[0016] Furthermore, using a gated prefix memory sliding window Transformer model, temporal feature modeling and spatial correlation modeling are performed on the initial plantar pressure prediction matrix and structured gait feature data. The initial pressure distribution is corrected through a nonlinear mapping mechanism, and the process of outputting the corrected plantar pressure distribution prediction matrix includes the following steps:

[0017] Step S1: Slice the initial plantar pressure prediction matrix according to the gait cycle, perform regional statistical encoding on the two-dimensional distribution of plantar pressure corresponding to each time step to obtain a pressure feature vector representing the plantar pressure state; normalize, embed, map, and concatenate the structured gait feature data of the corresponding time step to form a gait feature vector of uniform length; weight and fuse the pressure feature vector and the gait feature vector in the feature dimension to construct a joint feature sequence of plantar pressure-gait arranged over time;

[0018] Step S2: For each time step feature vector of the plantar pressure-gait joint feature sequence, perform linear transformation and nonlinear activation using the gated feature extraction subnetwork to generate the gated pre-activation vector and amplitude modulation vector for the corresponding time step; compress the gated pre-activation vector into non-negative memory gate coefficients through a nonlinear gate function with amplitude modulation.

[0019] Step S3: Perform accumulation and prefix scan operations along the time dimension based on the non-negative memory gating coefficients to generate a memory decay prefix vector sequence that monotonically changes with the sequence position; and use the memory decay prefix vector in the sequence as the basis for constructing the global bias of the attention mechanism of the gated prefix memory sliding window Transformer model, and form a gating bias matrix for attention score modulation by calculating the prefix decay difference between any two time steps.

[0020] Step S4: The sliding window attention operation is performed in the gated prefix memory sliding window Transformer model to calculate the attention score between the query vector and the locally visible key vector at the current time step; the attention score is subjected to time-related decay and enhancement modulation through the gated bias matrix to achieve differentiated memory control for different time stages and different historical gait states, and to obtain an intermediate representation that integrates temporal-spatial information.

[0021] Step S5: Residual fusion is performed between the intermediate representation of the fused temporal-spatial information and the initial plantar pressure prediction matrix. The nonlinear coupling relationship of the pressure response in each region of the plantar surface and the abnormal pressure peak pattern are further modeled through a feedforward network containing a normalization layer and a nonlinear activation function. In the feedforward network, a corrected plantar pressure distribution prediction matrix is ​​generated by using multi-layer linear transformation, activation function and output mapping.

[0022] Furthermore, the central processing server is a rack-mount industrial server equipped with a 32-core CPU, a floating-point arithmetic unit, and a GPU accelerator card. It is connected to the plantar pressure acquisition device, gait kinematics acquisition device, and display and interactive terminal via a local area network, making it suitable for continuous operation in the gait laboratory of a rehabilitation hospital.

[0023] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0024] (1) This invention achieves a deep integration of biomechanical priors and the intelligent prediction capabilities of specific computational models, significantly improving the structural consistency and reliability of plantar pressure prediction. By constructing a multi-rigid-body chain biomechanical model of the lower limb and combining it with foot-ground contact mechanics solutions, the system can obtain an initial pressure distribution that conforms to the laws of kinematics before prediction, providing a physically interpretable basis for subsequent neural network modeling. Compared with existing methods that rely solely on deep learning, this invention effectively solves the problems of lack of theoretical support for pressure prediction and instability of results due to cross-individual differences, making a clear correspondence between pressure map and joint load and segmental posture, thereby enhancing the reliability of prediction results in gait assessment.

[0025] (2) This invention utilizes a gated prefix memory sliding window Transformer to perform deep correction on the initial pressure matrix, achieving dynamic memory modeling of gait phases and significantly improving the ability to represent temporal consistency and nonlinear pressure patterns. This Transformer model can simultaneously capture the coupled temporal characteristics of plantar pressure and joint motion, and enhances the sensitivity to gait cycle changes through memory gating coefficients and prefix decay bias, overcoming the shortcomings of traditional Transformers in accurately identifying gait phase transitions. Therefore, this invention can accurately model pressure changes in key phases such as heel strike, arch transition, and toe lift, ensuring stable and robust predictions under complex conditions such as rapid load changes and abnormal peaks, overcoming the problem of prediction jumps that easily occur in existing models during gait dynamics.

[0026] (3) This invention enhances cross-individual generalization ability and practical application adaptability through a hybrid framework of "biomechanical prior + Transformer intelligent correction," making plantar pressure prediction more practical in rehabilitation assessment and orthotic design. Compared with pure mechanical models that are difficult to solve in real time and traditional data models that are highly dependent on training samples, this invention effectively balances physical realism, prediction accuracy, and computational efficiency, enabling the system to adapt to different populations, gait speeds, and foot morphological conditions while maintaining stable output. The corrected pressure distribution map, regional pressure statistical indicators, and risk scores generated by the system can be directly used to identify abnormal pressure, track the rehabilitation process, and guide the setting of individualized orthotic parameters, significantly improving the operability and decision support capabilities of clinical applications. Attached Figure Description

[0027] Figure 1 A schematic diagram of the system framework of an intelligent prediction system for gait plantar pressure based on the Transformer structure provided by the present invention;

[0028] Figure 2This is a flowchart illustrating the gated prefix memory sliding window Transformer model provided in Example 3. Detailed Implementation

[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0030] Example 1, according to Figure 1 This invention provides a gait plantar pressure intelligent prediction system based on the Transformer structure, applicable to gait assessment and orthotic design scenarios in rehabilitation hospital gait laboratories, orthopedic departments, and sports medicine clinics. The system includes:

[0031] The gait acquisition platform, located in the gait laboratory, is used to provide test subjects with a walking route of fixed length and clear direction to ensure the stability and repeatability of the walking environment during gait testing.

[0032] The plantar pressure acquisition device, arranged on the gait acquisition platform, consists of a plantar pressure sensing array composed of multi-point piezoelectric sensors and resistive pressure sensors, and is used to continuously collect raw plantar pressure data of local areas of the sole during the test subject's walking process.

[0033] The gait kinematics acquisition device is set in the lower limb position of the subject in the gait acquisition platform. It includes an optical motion capture camera array arranged in a fixed position in space and an inertial measurement unit (IMU) worn on the lower leg and dorsum of the foot of the subject. It is used to record gait kinematic data related to lower limb joint angles, stride length, stride frequency and gait cycle.

[0034] The central processing server, deployed in the gait laboratory computer room, processes the raw plantar pressure data and gait kinematic data, generates a corrected plantar pressure distribution prediction matrix, and converts it into a pseudo-color pressure distribution map, regional pressure statistical indicators, and risk score results, which are then sent to the display and interactive terminal.

[0035] Display and interactive terminals, set up at doctors' workstations and rehabilitation training area terminals, include displays and human-computer interaction interfaces to present the predicted results of plantar pressure distribution.

[0036] Example 2, based on Example 1, in which the central processing server includes: a data preprocessing module, a feature extraction module, a gait biomechanical modeling module, a plantar contact mechanics solving module, a plantar pressure depth correction module, and a result generation module;

[0037] The data preprocessing module performs noise reduction filtering, time synchronization and coordinate system unification on the raw plantar pressure data and gait kinematic data, and completes gait cycle segmentation according to the walking event recognition strategy to obtain cleaned and aligned basic gait data.

[0038] The feature extraction module extracts plantar pressure distribution features, gait temporal features, joint kinematic features, and spatial posture features based on gait baseline data to construct structured gait feature data;

[0039] The gait biomechanical modeling module collects anthropometric parameters of the human lower limbs and feet. Based on the anthropometric parameters and structured gait feature data, it constructs a multi-rigid-body chain gait biomechanical model that includes the pelvis, thigh, calf and multi-segment foot, and calculates the reaction forces and moments of each joint to obtain the initial plantar load distribution characteristics.

[0040] The plantar contact mechanics solution module establishes a plantar contact mechanics model based on the joint reaction forces, segmental postures, and a set of biomechanical parameters including segmental mass, center of mass position, moment of inertia, and contact element stiffness and damping parameters output by the multi-rigid-body chain gait biomechanical model. In the plantar contact mechanics model, the rigid body segments of the hindfoot, midfoot, and forefoot are treated as independent force-bearing units. The spatial position and normal direction of each contact unit are determined by the segmental posture, and the instantaneous reaction forces of the hip, knee, and ankle joints are transmitted to the foot segments along the multi-rigid-body chain. According to the foot-ground contact equilibrium equation, the vertical force, shear force, and segmental rotational demand of the segments are distributed to the corresponding equivalent spring-damped contact elements. The contact force of each contact element is obtained by solving the normal deformation and tangential damping response of the contact elements. Based on the initial pressure distribution of each region of the foot obtained by the solution, it is restructured according to the discrete mesh of the foot to form an initial plantar pressure prediction matrix consistent with the geometry of the foot.

[0041] Foot-ground contact equilibrium equation formula:

[0042] To ensure overall foot balance, the force and moment synthesis segment of all contact units should satisfy the following:

[0043] Overall force balance:

[0044] ;

[0045] in, Indicates the subscript of the foot contact unit. Indicates the first The normal unit vector of each contact element. Indicates the first The magnitude of the normal contact force of each contact unit Indicates the first The tangential unit vector of each contact element Indicates the first The magnitude of the tangential damping force of each contact unit; Indicates the first The resultant force vector of the contact elements; Represents the ankle joint reaction force vector;

[0046] Overall torque balance formula:

[0047] ;

[0048] in, Indicates the first The position vector of each contact unit relative to the ankle joint reference point Indicates the first The torque generated by each contact unit on the ankle joint; This represents the ankle joint reaction torque vector;

[0049] The plantar pressure depth correction module constructs a gated prefix memory sliding window Transformer model. This model performs temporal and spatial correlation modeling on the initial plantar pressure prediction matrix and structured gait feature data. A nonlinear mapping mechanism is used to correct the initial pressure distribution, outputting a corrected plantar pressure distribution prediction matrix. The construction process of the gated prefix memory sliding window Transformer model is as follows: based on the Transformer model, a prefix memory gating and sliding window bias modulation mechanism is introduced. Learnable temporal decay and enhancement are applied to the attention score, optimizing the temporal dependency modeling and local spatial sensitivity of the Transformer model, thus constructing the gated prefix memory sliding window Transformer model.

[0050] The results generation module is used to convert the corrected plantar pressure distribution prediction matrix into a pseudo-color pressure distribution map, regional pressure statistical indicators and risk score results, and send them to the display and interactive terminal for display.

[0051] Example 3, according to Figure 2 This embodiment is based on Embodiment 2. In this embodiment, a gated prefix memory sliding window Transformer model is used to perform temporal feature modeling and spatial correlation modeling on the initial plantar pressure prediction matrix and structured gait feature data. The initial pressure distribution is corrected through a nonlinear mapping mechanism, and the process of outputting the corrected plantar pressure distribution prediction matrix includes the following steps:

[0052] Step S1: Feature Joint Encoding: The initial plantar pressure prediction matrix is ​​sliced ​​according to the gait cycle. The two-dimensional distribution of plantar pressure corresponding to each time step is statistically encoded in the region to obtain a pressure feature vector representing the plantar pressure state. The structured gait feature data of the corresponding time step (including the sagittal and coronal angles of the hip, knee, and ankle joints, single step length, step frequency corresponding to the number of steps per unit time, and gait parameters marked by gait stages from heel strike to toe lift) is normalized, embedded, mapped, and concatenated to form a gait feature vector of uniform length. The pressure feature vector and the gait feature vector are weighted and fused in the feature dimension to construct a plantar pressure-gait joint feature sequence arranged over time.

[0053] Step S2: Gated Memory Extraction: For each time-step feature vector of the plantar pressure-gait joint feature sequence, a gated feature extraction subnetwork is used to perform linear transformation and nonlinear activation to generate a gated pre-activation vector and amplitude modulation vector for the corresponding time step. A nonlinear gating function with amplitude modulation is used to compress the gated pre-activation vector into non-negative memory gating coefficients, which characterize the degree of retention and attenuation of historical memory in the current gait state. The non-negative memory gating coefficients are adaptively calculated from the plantar pressure-gait joint features of the current time step, achieving differentiated memory control for different time phases and pressure patterns. The formula used is as follows:

[0054] Gated preactivation vector formula:

[0055] ;

[0056] in, Indicates the current time step. Indicates the first Plantar pressure-gait joint feature vector at each time step; This represents the weight matrix used to generate the gated pre-activation vector in the gated feature extraction subnetwork. Indicates corresponding to The bias vector; Indicates time step The gated preactivation vector;

[0057] Amplitude modulation vector formula:

[0058] ;

[0059] in, This represents the weight matrix used to generate the amplitude modulation vector. This represents the activation function of the Exponential Linear Unit; Indicates time step The amplitude modulation vector;

[0060] Formula for nonnegative memory gating coefficient:

[0061] ;

[0062] in, This represents element-wise multiplication (Hadamard product). Indicates time step The non-negative memory-gated coefficient vector; Represents a smooth, nonlinear function;

[0063] This memory gating coefficient It adaptively adjusts the memory intensity for different gait phases (such as heel strike, ball of the foot transition, and toe lift), achieving "gait phase perception-based memory adjustment" which is not available in traditional Transformers, and can learn the phased patterns of pressure changes.

[0064] Step S3: Prefix decay construction: Based on the non-negative memory gating coefficients, accumulation and prefix scan operations are performed along the time dimension to generate a sequence of memory decay prefix vectors that monotonically change with the sequence position; and the memory decay prefix vectors in the sequence are used as the basis for global bias construction of the attention mechanism of the gated prefix memory sliding window Transformer model. By calculating the prefix decay difference between any two time steps, a gating bias matrix for attention score modulation is formed, which is used to perform learnable decay modulation on the attention score; the prefix scan operation is completed in the on-chip cache using a streaming scan algorithm divided by time blocks, and the prefix results are only retained in external storage, avoiding the storage of large-scale intermediate tensors, reducing the GPU memory access overhead, and making it suitable for long-term stable operation in GPU or AI acceleration servers;

[0065] Step S4: Sliding Window Attention Modulation: The sliding window attention operation is performed in the gated prefix memory sliding window Transformer model to calculate the attention score between the query vector and the locally visible key vector at the current time step; through the gated bias matrix, the attention score is subjected to time-related attenuation and enhancement modulation to achieve differentiated memory control for different time phases and different historical gait states, and to obtain an intermediate representation that integrates temporal-spatial information, providing a basis for nonlinear correction of plantar pressure distribution;

[0066] Specifically, the time-related attenuation and enhancement modulation includes: incorporating the gated bias matrix into the calculation of the sliding window attention normalization weights to construct gated local attention weights, thereby achieving dynamic weighting based on time-space coupling characteristics and generating intermediate representations with better temporal consistency and spatial sensitivity. The formula used is as follows:

[0067] ;

[0068] in, Indicates querying the location index. Indicates the position index of the locally visible key. This represents the local attention weights after gating modulation. Indicates the width of the sliding window. Indicates at time step Query vector and time step The basic attention score (logit) between key vectors. Indicates at time step (Query side) and time step The gating bias term between the (key sides) is used to perform time-related decay or enhancement modulation on the basic attention score; Represents an exponential function. This represents the sliding window indicator function, used to ensure that only key vectors within the window range participate in the attention calculation; Indicates at time step Query vector and time step within the sliding window Basic attention scoring between key vectors; Indicates attention link The gated bias term is used to perform consistent dynamic modulation on the positions of all candidate keys within the window;

[0069] Step S5: Residual Fusion Correction: The intermediate representation of the fused temporal-spatial information and the initial plantar pressure prediction matrix are residually fused. Through a feedforward network containing a normalization layer and a nonlinear activation function, the nonlinear coupling relationship of the pressure response in each region of the plantar surface and the abnormal pressure peak pattern are further modeled. In the feedforward network, multi-layer linear transformation, activation function and output mapping are used to generate a corrected plantar pressure distribution prediction matrix. The corrected plantar pressure distribution prediction matrix is ​​consistent with the original gait sequence in the temporal dimension and corresponds one-to-one with the plantar anatomical regions in the spatial dimension. It can be used for subsequent pressure risk assessment and insole / orthotic parameter optimization.

[0070] In conventional technical fields, the process of performing temporal feature modeling and spatial correlation modeling on the initial plantar pressure prediction matrix and structured gait feature data using the Transformer model, correcting the initial pressure distribution through a nonlinear mapping mechanism, and outputting a corrected plantar pressure distribution prediction matrix specifically includes the following steps:

[0071] Step B1: Feature Construction and Assembly Processing: Expand the initial plantar pressure prediction matrix in chronological order and extract pressure distribution features for each time step; standardize the structured gait features (joint angles, gait frequency, stride length) and directly assemble them with the pressure features to form sequential gait-pressure feature data;

[0072] Step B2: Perform temporal feature modeling: Input the serialized gait-stress feature data into the standard Transformer encoder, and use its global self-attention mechanism to model the correlation between different time steps to obtain an intermediate representation of temporal dependence;

[0073] Step B3: Perform spatial correlation modeling and feature transformation: Based on the intermediate representation representing temporal dependence, the spatial correlation between different regions of the sole is modeled through a conventional feedforward network structure to obtain potential feature data for pressure correction;

[0074] Step B4: Perform multi-layer nonlinear mapping on the latent feature data to restore the model output to a prediction matrix consistent with the original plantar pressure distribution spatial dimension, thereby achieving overall correction of the initial pressure distribution and obtaining the corrected plantar pressure distribution prediction matrix; output the corrected plantar pressure distribution prediction matrix in chronological order as the final plantar pressure distribution prediction result for subsequent gait assessment or auxiliary orthopedic analysis.

[0075] Example 4, based on Example 3, uses a rack-mounted industrial server equipped with a 32-core CPU, a floating-point unit, and a GPU accelerator card. It is connected to the plantar pressure acquisition device, the gait kinematics acquisition device, and the display and interactive terminal via a local area network, making it suitable for continuous operation in the gait laboratory of a rehabilitation hospital.

[0076] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A gait plantar pressure intelligent prediction system based on a Transformer structure, characterized in that, The system comprises: a plantar pressure acquisition device for acquiring plantar pressure raw data; a gait kinematics acquisition device for recording gait kinematics data; a central processing server for processing the plantar pressure raw data and the gait kinematics data, generating a corrected plantar pressure distribution prediction matrix, and converting the corrected plantar pressure distribution prediction matrix into a pseudo-color pressure distribution map, regional pressure statistical indicators, and a risk score result, and sending the result to a display and interaction terminal; the display and interaction terminal is arranged at a doctor workstation and a rehabilitation training area terminal, and comprises a display and a human-computer interaction interface, and is used for presenting the plantar pressure distribution prediction result; the central processing server comprises a data preprocessing module, a feature extraction module, a gait biomechanics modeling module, a plantar contact mechanics solving module, and a plantar pressure depth correction module; the feature extraction module constructs structured gait feature data; the plantar contact mechanics solving module calculates an initial pressure distribution to form an initial plantar pressure prediction matrix; the plantar pressure depth correction module constructs a gated prefix memory sliding window Transformer model, performs time series feature modeling and spatial correlation modeling on the initial plantar pressure prediction matrix and the structured gait feature data through the gated prefix memory sliding window Transformer model, corrects the initial pressure distribution through a nonlinear mapping mechanism, and outputs a corrected plantar pressure distribution prediction matrix; the construction process of the gated prefix memory sliding window Transformer model is that, based on a Transformer model, a prefix memory gating and a sliding window bias modulation mechanism are introduced to optimize time series dependency modeling and local spatial sensitivity of the Transformer model, and the gated prefix memory sliding window Transformer model is constructed; the process of outputting the corrected plantar pressure distribution prediction matrix through the gated prefix memory sliding window Transformer model specifically comprises the following steps: Step S1: slice and regionally statistically encode the initial plantar pressure prediction matrix according to a gait cycle to obtain a pressure feature vector; normalize, embed, and splice the structured gait feature data to form a gait feature vector; and weight and fuse the pressure feature vector and the gait feature vector in the feature dimension to obtain a plantar pressure-gait joint feature sequence; Step S2: for each time step feature vector of the plantar pressure-gait joint feature sequence, perform linear transformation and nonlinear activation to generate a gated pre-activation vector and an amplitude modulation vector; and through a nonlinear gating function with amplitude modulation, compress the gated pre-activation vector into a non-negative memory gating coefficient; Step S3: based on the non-negative memory gating coefficient, perform accumulation and prefix scan operation along the time dimension to generate a memory decay prefix vector; and take the memory decay prefix vector as a global bias construction basis of an attention mechanism of the gated prefix memory sliding window Transformer model, and form a gating bias matrix by calculating a prefix decay difference between any two time steps. Step S4: the sliding window attention operation performed in the gated prefix memory sliding window Transformer model, calculates the attention score between the query vector of the current time step and the locally visible key vector; through the gating bias matrix, the attention score is time-dependent decay and enhancement modulation to obtain the intermediate representation of fusion time-space information; Step S5: residual fusion is performed on the intermediate representation of fusion time-space information and the initial plantar pressure prediction matrix, and the non-linear coupling relationship and abnormal pressure peak pattern of the pressure response of each region of the foot are further modeled through the feedforward network containing the normalization layer and the nonlinear activation function; the multi-layer linear transformation, activation function and output mapping in the feedforward network are used to generate the corrected plantar pressure distribution prediction matrix.

2. The intelligent gait plantar pressure prediction system based on the Transformer structure according to claim 1, characterized in that: The data preprocessing module specifically includes: denoising filtering, time synchronization and coordinate system processing of the original plantar pressure data and gait kinematics data to obtain gait basic data; The feature extraction module specifically includes: extracting features based on the gait basic data to construct structured gait feature data; The gait biomechanical modeling module specifically includes: collecting anthropometric parameters, and constructing a gait biomechanical model containing the pelvis, thigh, lower leg and multi-segment foot based on the anthropometric parameters and structured gait feature data; The plantar contact mechanics solving module specifically includes: outputting a set of biomechanical parameters based on the gait biomechanical model, establishing a plantar contact mechanics model, and calculating the initial pressure distribution in the plantar contact mechanics model and restructuring it to form an initial plantar pressure prediction matrix.

3. The intelligent gait plantar pressure prediction system based on the Transformer structure according to claim 2, characterized in that: The feature extraction module specifically extracts plantar pressure distribution features, gait timing features, joint kinematics features and spatial posture features based on the gait basic data to construct structured gait feature data.

4. The intelligent gait plantar pressure prediction system based on the Transformer structure according to claim 1, characterized in that: The central processing server is a rack-mounted industrial server configured with a 32-core CPU, a floating point operation unit and a GPU acceleration card, and is connected to the plantar pressure acquisition device, the gait kinematics acquisition device and the display and interaction terminal through a local area network.

Images