Plantar shear force estimation method and device based on IMU and plantar pressure data

Abstract

The invention provides a plantar shear force estimation method and device based on IMU and plantar pressure data, and relates to the technical field of plantar stress and gait measurement, and the method comprises the steps: synchronously collecting plantar pressure data and IMU data in a running process; judging running modes according to the plantar pressure data, and determining input characteristics corresponding to the running modes; the input features are input into an estimation model based on a PatchTST model to predict shear force. According to the method, the plantar shear force can be accurately and rapidly estimated, the universality and practicability are high, the cost is low, and the dependence on the environment is low.

Description

Technical Field

[0001] This invention relates to the field of plantar force and gait measurement technology, and in particular to a method and apparatus for estimating plantar shear force based on IMU and plantar pressure data. Background Technology

[0002] During dynamic activities such as walking and running, the sole of the foot is subjected to multidimensional loads of vertical pressure and horizontal shear force (force parallel to the sole's contact surface). However, traditional foot biomechanical studies mostly focus on vertical pressure as the core indicator, which only reflects the load distribution in the vertical direction and cannot reveal the relative slippage trend and stress environment in the horizontal direction. Shear force, on the other hand, directly affects the load propagation, deformation patterns, and microcirculation perfusion of local tissues. Its coupling relationship with pressure (such as spatial offset and temporal peak synergy) alters tissue tolerance and is a key supplement to understanding the complete mechanism of gait mechanics. Therefore, without shear force data, it is difficult to fully explain the causes of foot tissue damage, resulting in blind spots in gait analysis and biomechanical assessment.

[0003] Existing methods for detecting plantar shear force have drawbacks such as high cost, high level of specialization, high dependence on the environment, long data calibration and calculation time, and poor practicality and versatility. Summary of the Invention

[0004] The main objective of this invention is to propose a method and apparatus for estimating plantar shear force based on IMU and plantar pressure data. This method can achieve accurate and rapid estimation of plantar shear force, and is highly versatile and practical with low cost and low environmental dependence.

[0005] This invention is achieved through the following technical solution:

[0006] The plantar shear force estimation method based on IMU and plantar pressure data includes the following steps:

[0007] Step S1: Synchronously collect plantar pressure data and IMU data during running, and preprocess the collected data to complete the data and remove outliers;

[0008] Step S2: Determine the running mode based on plantar pressure data, identify the input features corresponding to each running mode, and normalize the input features to eliminate dimensional differences. The input features consist of plantar pressure data and IMU data, and the IMU data includes acceleration data and angle data.

[0009] Step S3: Input Features. The input is based on the PatchTST model estimation model, which includes a projection layer, a mask layer, a partitioning layer, a conditional embedding layer, an encoder, and a shear force dimension mapping layer connected in sequence. The projection layer maps the dimensions of the input features corresponding to different running modes to a fixed intermediate dimension to obtain the mapped features. The mask layer generates a validity mask corresponding to each running mode based on the intermediate dimension. The partitioning layer divides the mapped features into multiple temporal segments to obtain a segment matrix. The conditional embedding layer combines the validity mask and the segment matrix to shield invalid channels to filter valid features. The encoder learns the temporal dependence of valid features through multi-head self-attention and uses a feedforward network to enhance the nonlinear fitting of the multi-head self-attention output. The channel structure is restored by anti-reconstruction to obtain the encoded features. The shear force dimension mapping layer converts the encoded features into shear force prediction values.

[0010] Furthermore, in step S1, data acquisition is performed using the insole. The insole is equipped with a pressure sensor for collecting plantar pressure data and an IMU module for collecting IMU data. During acquisition, the hardware clock of the IMU module is used as a reference to synchronize the pressure sensor. The insole is equipped with n pressure sensors to collect n channels of plantar pressure data.

[0011] Furthermore, in step S1, the data completion includes completing plantar pressure data using linear interpolation and completing IMU data using cubic spline interpolation. If the same type of data is missing for three consecutive time steps, the data is re-acquired. Outliers in plantar pressure data and IMU data are removed using the isolated forest algorithm.

[0012] Furthermore, step S2 specifically includes the following steps:

[0013] Step S21, Obtain At this moment and The value when and The running style is to land on the forefoot, when and The running style is to land on the heel, when and The running style involves landing on the entire foot, among which... , Indicates time t Pressure on the entire foot, This represents the sum of data collected by all pressure sensors on the forefoot at time t. This represents the sum of data collected by all pressure sensors on the heel at time t. This represents the basic pressure value for the forefoot. The base pressure value refers to the pressure data collected during takeoff.

[0014] Step S22: When the running style is forefoot landing, the corresponding input feature is represented as follows: When the running style is heel strike, the corresponding input feature is represented as: When the running style is full foot strike, the corresponding input feature is represented as: ,in, The pressure data collected by i pressure sensors located at the forefoot of the insole. Pressure data collected by ni pressure sensors located at the heel of the insole. These are the accelerations in the x-axis, y-axis, and z-axis directions collected by the IMU module, respectively. These are the roll angle, roll angle, and yaw angle collected by the IMU module, respectively.

[0015] Step S23: Apply Min-Max normalization to the input features Perform normalization to obtain the input features. .

[0016] Furthermore, in step S3, the projection layer adopts... Input features Mapping yields mapping features ,in, This is the projection weight matrix corresponding to the running style type. , This indicates landing on the ball of the foot. This indicates landing on the heel. This indicates that the entire foot lands on the ground. For dimension The set of real numbers, For the dimension corresponding to running type, As the intermediate dimension, This is the projection bias vector corresponding to the running style.

[0017] Furthermore, in step S3, the mask layer is determined according to the formula... Generate binary validity mask , For indicator functions, This is the set of valid channels corresponding to the input feature of running style type.

[0018] Furthermore, in step S3, the partitioning layer first maps the features. Convert to channel priority format Then set the number of segments per channel to Finally, the time series is divided using a sliding window to obtain the fragment matrix. ,in, For tensor dimension permutation operations, For training batch size, For mapping features The timing length, This is the sliding window splitting function. The length of the segment obtained by segmentation. This is the sliding step size.

[0019] Furthermore, in step S3, the effective features selected by the conditional embedding layer are represented as follows: ,in, , , , It represents the Hadamah accumulation. This indicates a dimension expansion operation. To map the low-dimensional features of each segment to the dimensions of the Transformer model, , To share projection parameters, The full-channel embedding matrix is ​​used; the encoded features obtained by the encoder are represented as follows: ,in, , , For layer normalization operations, the encoder's feedforward network , This represents the activation function. This is the weight matrix corresponding to Multi-Head Attention (MHA). for The weight matrix corresponding to the activation function, This is the bias vector corresponding to Multi-Head Attention (MHA). for Bias vectors corresponding to activation functions, multi-head attention , Here, h represents the number of attention heads, and the concatenation function is used. , , , , , , , For attention projection parameters, Attention mask for invalid combinations Indicates will The attention weight is reset to a minimum value. , , For tensor dimension reshaping operations, For tensor dimension expansion operations, These are parameters for multi-head fusion.

[0020] Furthermore, in step S3, the predicted shear force value obtained by the shear force dimension mapping layer is expressed as: ,in, , , , These represent the weights and biases of the shear force dimension mapping layer, respectively. The mapping weights and biases correspond to the running style. This is the physical limit value normalized to Min-Max.

[0021] This invention is also achieved through the following technical solutions:

[0022] A plantar shear force estimation device based on IMU and plantar pressure data is used to implement the estimation method described in any of the preceding claims. It includes an insole, an input feature determination module, and an estimation module. The insole is equipped with a pressure sensor for collecting plantar pressure data and an IMU module for collecting IMU data. The input feature determination module determines the running mode based on the plantar pressure data and determines the input features corresponding to each running mode. It then normalizes the input features to eliminate dimensional differences. These input features consist of plantar pressure data and IMU data, including acceleration and angle data. The estimation module is based on the PatchTST model and includes a projection layer, a mask layer, a partitioning layer, and so on, connected sequentially. The system consists of a conditional embedding layer, an encoder, and a shear force dimension mapping layer. The projection layer maps the dimensions of the input features corresponding to different running styles to a fixed intermediate dimension to obtain the mapped features. The masking layer generates a validity mask corresponding to each running style based on the intermediate dimension. The partitioning layer divides the mapped features into multiple temporal segments to obtain a segment matrix. The conditional embedding layer combines the validity mask and the segment matrix to shield invalid channels and filter valid features. The encoder learns the temporal dependence of valid features through multi-head self-attention and uses a feedforward network to enhance the nonlinear fitting of the multi-head self-attention output. The channel structure is restored by anti-reconstruction to obtain the encoded features. The shear force dimension mapping layer converts the encoded features into shear force prediction values.

[0023] As can be seen from the above description of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0024] This invention first synchronously collects plantar pressure data and IMU data during running. Then, it determines the running mode based on the plantar pressure data, thereby determining the input features corresponding to each running mode. The input features are input into an estimation model based on the PatchTST model. The projection layer of the estimation model maps the dimensions of the input features corresponding to different running modes to a fixed intermediate dimension, obtaining mapped features. The masking layer generates a validity mask corresponding to each running mode based on the intermediate dimension. The partitioning layer divides the mapped features into multiple temporal segments to obtain a segment matrix. The conditional embedding layer combines the validity mask and the segment matrix to shield invalid channels to filter valid features. The encoder learns the temporal dependence of valid features through multi-head self-attention and uses a feedforward network to enhance the nonlinear fitting of the multi-head self-attention output. The channel structure is restored by anti-reconstruction to obtain encoded features. The shear force dimension mapping layer converts the encoded features into shear force prediction values. In this way, multimodal data of wearable insoles are combined with deep learning to estimate plantar shear force, effectively improving estimation accuracy and speed, enhancing versatility and practicality, and reducing cost and environmental dependence. Attached Figure Description

[0025] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] Figure 1 This is a flowchart of the present invention.

[0027] Figure 2 This is a schematic diagram of the insole of the present invention.

[0028] Figure 3 This is a force diagram illustrating different running methods according to the present invention.

[0029] Figure 4 This is a comparison chart of the total plantar shear force predicted by this invention with the benchmark value.

[0030] Figure 5 This is a comparison of partial segments of the plantar shear force in the forward and backward directions and the baseline value corresponding to different running styles according to the present invention. Detailed Implementation

[0031] The present invention will be further described below through specific embodiments.

[0032] like Figure 1 As shown, the plantar shear force estimation method based on IMU and plantar pressure data includes the following steps:

[0033] Step S1: Synchronously collect plantar pressure data and IMU data during running, and preprocess the collected data to complete the data and remove outliers;

[0034] Specifically, using, such as Figure 2The insole shown is used for data acquisition. It is equipped with pressure sensors to collect plantar pressure data and an IMU module to collect IMU data. The insole has n pressure sensors to collect n channels of plantar pressure data. It also has circuit boards connected to each pressure sensor, and the IMU module is mounted on the circuit board. The specific structure of the circuit board, pressure sensors, IMU module, and their connection structure are all existing technologies. The pressure sensors and IMU modules of the left and right insoles are symmetrically arranged. The following description uses the left insole as an example.

[0035] The generation of plantar shear force is directly related to the dynamics of running (acceleration) and foot posture (angle). Plantar pressure data alone cannot fully characterize the running scenario. Therefore, an IMU module is used to collect acceleration and angle. Acceleration refers to the acceleration along the x / y / z axes, and angle refers to roll angle, yaw angle, and yaw angle. Acceleration reflects the impact intensity and speed changes of the foot during running, such as the vertical / horizontal impact acceleration upon landing, and captures "dynamic impact characteristics" that are difficult to cover by pressure data. For example, the impact acceleration when landing on the forefoot is significantly different from that when landing on the heel. Speed ​​reflects the spatial posture of the foot during running, such as the foot tilt angle and posture stability upon landing, supplementing the "postural context" corresponding to the plantar pressure distribution. For example, the yaw angle when landing on the entire foot is different from that when landing on the forefoot, which directly affects the distribution pattern of shear force.

[0036] To ensure consistent timestamps, the IMU module's hardware clock is used as a reference during data acquisition to synchronize the pressure sensors. All sensor data acquisitions include a timestamp 't' in a uniform format, measured in milliseconds, formatted as "cumulative milliseconds since system startup".

[0037] Data completion includes: if pressure data is missing at a certain time t, such as due to sensor transmission delay, linear interpolation is used to complete the data at time t based on the pressure values ​​at (t-10) ms and (t+10) ms; if IMU data is missing, cubic spline interpolation is used to complete it to preserve the temporal smoothness of acceleration and angle; if the same type of data is missing for three consecutive time steps (times), local data resampling is triggered, specifically resampling only the most recent 20 ms data of that sensor to avoid overall data loss.

[0038] Outliers were removed from the completed plantar pressure and IMU data using the Isolation Forest algorithm. The specific process is based on existing technology.

[0039] Left foot n-channel plantar pressure data is represented as The unit is N; based on the distribution of pressure sensors on the insole, the data from the 1st to the ith channel represents the forefoot pressure data, and the data from the (i+1)th to the nth channel represents the heel pressure data.

[0040] Forefoot pressure and its representation ,in, This represents the forefoot pressure at the current time t. This indicates the pressure on the forefoot at the previous moment. Indicates the pressure on the forefoot at the next moment;

[0041] Heel pressure and its representation are as follows , This indicates the pressure on the heel at the current moment. This indicates the pressure on the heel of the foot at the previous moment. Indicates the pressure on the heel of the foot at the next moment;

[0042] The pressure on the whole foot is expressed as , This indicates the pressure on the entire foot at the current moment. This indicates the pressure on the entire foot at the previous moment. This indicates the pressure on the entire foot at the next moment;

[0043] Forefoot base pressure value The basic pressure value of the heel is The base pressure value refers to the pressure exerted by the foot on the insole when the foot is off the ground after wearing the insole.

[0044] IMU acceleration data is in the x-axis direction. y-axis direction z-axis direction The unit is m / s 2 ;

[0045] IMU angular velocity data is roll angle Roll angle Yaw angle The unit is degrees.

[0046] Step S2: Determine the running mode based on plantar pressure data, identify the input features corresponding to each running mode, and normalize the input features to eliminate dimensional differences. The input features consist of plantar pressure data and IMU data, and the IMU data includes acceleration data and angle data.

[0047] Specifically, the steps include the following:

[0048] Step S21, Obtain At this moment and The value when and The running style is to land on the forefoot, when and The running style is to land on the heel, when and The running style involves landing on the entire foot, among which... , This represents the sum of data collected by all pressure sensors on the forefoot at time t. This represents the sum of data collected by all pressure sensors on the heel at time t; force diagrams for different running styles are shown below. Figure 3 As shown;

[0049] Step S22: When the running style is forefoot landing, the corresponding input feature is represented as follows: When the running style is heel strike, the corresponding input feature is represented as: When the running style is full foot strike, the corresponding input feature is represented as: ,in, The pressure data collected by i pressure sensors located at the forefoot of the insole. Pressure data collected by ni pressure sensors located at the heel of the insole. These are the accelerations in the x-axis, y-axis, and z-axis directions collected by the IMU module, respectively. These are the roll angle, roll angle, and yaw angle collected by the IMU module, respectively.

[0050] Step S23: Apply Min-Max normalization to the input features Perform normalization to obtain the input features. ;

[0051] To eliminate the differences in the dimensions of different sensors while preserving the data distribution trend, avoid the influence of numerical range during model training, and improve the convergence speed, Min-Max normalization is used for the input features. Perform normalization, mapping the normalization range to [0,1].

[0052] Step S3: Input Features. The input is based on the PatchTST model estimation model, which includes a projection layer, a mask layer, a partitioning layer, a conditional embedding layer, an encoder, and a shear force dimension mapping layer connected in sequence. The projection layer maps the dimensions of the input features corresponding to different running styles to a fixed intermediate dimension to obtain the mapped features. The mask layer generates a validity mask corresponding to each running style based on the intermediate dimension. The partitioning layer divides the mapped features into multiple temporal segments to obtain a segment matrix. The conditional embedding layer combines the validity mask and the segment matrix to shield invalid channels to filter valid features. The encoder learns the temporal dependence of valid features through multi-head self-attention and uses a feedforward network to enhance the nonlinear fitting of the multi-head self-attention output. The channel structure is restored by anti-reconstruction to obtain the encoded features. The shear force dimension mapping layer converts the encoded features into shear force prediction values.

[0053] As can be seen from step S2, the input features corresponding to forefoot strike, heel strike, and full foot strike are not the same; that is, different running styles correspond to different dimensions of input features. To predict plantar shear force and avoid subsequent module adaptation problems caused by different input feature dimensions, a projection layer is designed at the input of the estimation model to uniformly map input features of different dimensions to a fixed intermediate dimension. Thus, the PatchTST model structure can be adapted to three running modes without modifying it.

[0054] Specifically, the projection layer adopts Input features Mapping yields mapping features ,in, For training batch size, For mapping features The timing length, This is the projection weight matrix corresponding to the running style type. , This indicates landing on the ball of the foot. This indicates landing on the heel. This indicates that the entire foot lands on the ground. For dimension The set of real numbers, Input features corresponding to running style type Dimensions As the intermediate dimension, In this embodiment, the projection bias vector corresponding to the running style is taken as... To cover the largest dimension of all scenarios, The input feature dimension corresponding to a full-foot strike is... Corresponding to the value when type=A, n is the total number of pressure sensors on the insole, corresponding to n channels of plantar pressure data, and 6 is the fixed-dimensional data collected by the IMU module, including x-axis acceleration, y-axis acceleration, z-axis acceleration, roll angle, yaw angle, and yaw angle, for a total of 6 channels.

[0055] For the projection weight matrix, a sparse initialization strategy is adopted, that is, the projection weights of the original input features that have no corresponding dimension are initialized to 0, in order to avoid interference from invalid parameters, such as when landing on the forefoot. The column corresponding to the midfoot and rearfoot channel is initialized to 0.

[0056] To shield "invalid channels after unification of dimensions" (such as when the forefoot lands), during the conditional embedding layer and encoder stages. (The location corresponding to the midfoot and rearfoot channel) ensures that the estimation model only focuses on effective features, avoiding invalid information from diluting the learning efficiency of the estimation model. The mask layer is based on the formula... Generate binary validity mask , This is an indicator function; the mask value is 1 for valid channels and 0 for invalid channels. For the set of valid channels corresponding to running type, "Channel" is an abbreviation for "channel," representing a unique identifier for a single data dimension in the input features. Each channel corresponds to a specific sensor data or feature dimension, such as a specific acquisition channel of a forefoot pressure sensor or a specific dimension of data from an IMU module. If a channel belongs to the valid channel set for the current running mode... (Right now The indicator function I(⋅) outputs 1, corresponding to a mask position of 1, indicating that the channel data is valid; if a channel does not belong to Ctype, the indicator function outputs 0, corresponding to a mask position of 0, indicating that the channel data is invalid and will be blocked later.

[0057] The partitioning layer performs temporal segmentation (Patch) on the unified-dimensional features, splitting only the temporal dimension while preserving the channel structure. The specific process is as follows:

[0058] First, map the features. Convert to channel priority format ,in, For tensor dimension permutation operations, that is... The first and second dimensions are replaced because the original storage format of wearable sensor data is "batch × number of channels × time series length", and the sliding window needs to split the time series dimension (seqlen). Therefore, the "time series length" dimension is first adjusted to a position that the sliding window can operate by dimension replacement.

[0059] Then let the Patch length be... Step size is Then the number of segments in a single channel is ,in, To round down;

[0060] Finally, the time series is divided using a sliding window to obtain the fragment matrix. ,in, , This is a sliding window function that operates according to the segment length. Extract timing segments with a step size of [missing information]. , For the sliding window splitting function, Segmenting according to the set patchlen and stride physically means dividing continuous wearable time-series data into several local segments. The length of the segment obtained by segmentation. The sliding step size is the time step interval between two adjacent segments.

[0061] This invention assigns dedicated embedding vectors to different modal features (pressure, acceleration, angular velocity) through basic projection and dedicated channel embedding. At the same time, it combines validity masking to filter valid modalities, which strengthens the physical correlation of multimodal features. For example, it embeds and binds forefoot pressure with the corresponding tilt angle, avoiding information confusion caused by simply splicing multimodal features.

[0062] The specific process of conditional embedding layer is as follows:

[0063] First, the low-dimensional features of each patch are mapped to the Transformer model dimension of the PatchTST model. ,Right now ,in, , To share projection parameters;

[0064] Then, a unique embedding vector is assigned to each channel: , ,in, For full-channel embedding matrix, This represents a dimension expansion operation, which adds a new dimension of length 1 to the 0th dimension, ensuring that the embedding matrix is ​​consistent with the previous dimension. Dimension matching, To map the low-dimensional features of each segment to the dimensions of the Transformer model;

[0065] Finally, valid features are selected using a validity mask. ,in, It represents the Hadamardi (or Hadama) stack.

[0066] In the encoder stage, features are first reassembled to adapt to the Transformer input, then the temporal dependencies of the patch are learned through multi-head self-attention, FFN is used to enhance nonlinear fitting, residual connections and layer normalization are used to ensure training stability, and finally the channel structure is restored by anti-reassembly.

[0067] The specific process is as follows:

[0068] First, merge the "Batch-Channel" dimensions to adapt to the Transformer input format, resulting in... ,in, For tensor dimension reshaping operations, It is the target shape, and its core function is to redefine the length of the dimensions without changing the total number of tensor elements, thereby changing the "dimensional perspective" of the tensor (the order of elements remains unchanged).

[0069] Then, for the "Batch-Channel" combination corresponding to the invalid channel, the masker attention is calculated: , , For dimensional reshaping operations, For tensor dimension expansion operations, It represents the number of times the dimension is copied. Its core function is to copy tensor elements according to a specified dimension, thereby expanding the dimensional length of the tensor.

[0070] Next, the temporal dependency of the effective channels is calculated using multi-head self-attention. ,in, Here, h represents the number of attention heads, and the expression for calculating a single attention head is: , , , , , , For attention projection parameters, , Attention mask for invalid combinations Indicates will The attention weights are reset to a minimum value, and after Softmax, they approach 0. For multi-head fusion parameters;

[0071] Next, the feedforward network FFN enhances the nonlinear fitting function. ,in, This represents the activation function, specifically the modified linear unit. Its core function is to introduce non-linearity into the model to address the problem that linear models cannot fit complex patterns. It is also computationally efficient and alleviates the vanishing gradient problem. This is the weight matrix corresponding to Multi-Head Attention (MHA), and it's the first-layer weight matrix. Its function is to map the low-dimensional features output by MHA to a higher-dimensional space, achieving feature upsizing and transformation. for The weight matrix corresponding to the activation function is the second-layer weight matrix. Its function is to map the high-dimensional features after ReLU activation back to the hidden layer dimension of the model, thereby completing the dimensionality reduction of the features and ensuring that the output dimension of FFN is consistent with the input (i.e., MHA output). Here, represents the bias vector corresponding to Multi-Head Attention (MHA), and represents the first-layer bias vector. As additive bias parameters, their function is to compensate for the shift after the linear transformation of the weight matrix, thereby increasing the model's fitting ability. for The bias vector corresponding to the activation function is the second-layer bias vector, which is also an additive bias parameter to compensate for the offset of the second-layer linear transformation. The final output dimension is dmodel, which matches the encoder input dimension and meets the requirements of subsequent layer normalization and residual connection.

[0072] These parameters are trainable parameters that the Transformer model learns automatically during training, not fixed values ​​set manually. The model continuously adjusts the values ​​of these parameters based on the training data, allowing FFN to better capture the non-linear relationships of the input features.

[0073] Finally, each encoder layer is trained to improve stability through residual connections and layer normalization. After stacking L encoder layers, the encoded features are obtained. ,in, , , , This is a layer normalization operation.

[0074] The shear force dimension mapping layer converts encoded features into predicted shear force values ​​through "temporal prediction + dedicated dimension mapping". The specific process is as follows:

[0075] First, convert the Patch dimension to the prediction time series dimension. ,get , ,in, , The weight and bias parameters of the shear force dimension mapping layer convert the high-dimensional encoded features output by the encoder from the temporal segment dimension to the shear force prediction temporal dimension. Both parameters are trainable parameters of the model and are optimized through gradient descent during training to adapt to the feature mapping requirements of shear force prediction.

[0076] Then, the intermediate features are mapped to the shear force dimension, retaining only the contribution of the effective channels: ,in, , The mapping weights (set to 0 for rows corresponding to invalid channels) and biases correspond to the running mode;

[0077] Finally, inverse normalization is performed to restore the original physical unit N of shear force, resulting in the predicted shear force value. ,in, The physical limit value is the Min-Max normalized value, specifically, This refers to the maximum shear force. This refers to a non-negative shear force.

[0078] The estimation device for implementing the above-mentioned shear force estimation method includes an insole, an input feature determination module, and an estimation module. The insole is equipped with pressure sensors for collecting plantar pressure data and an IMU module for collecting IMU data. More specifically, multiple pressure sensor arrays are arranged on the forefoot and heel of the insole, and the pressure sensors are numbered sequentially from the forefoot to the heel. A PCB board is also mounted on the insole, connected to each pressure sensor and transmitting its collected data to a computing terminal. The IMU module is mounted on the PCB board. The PCB board setup, the specific setup of the pressure sensors, the numbering process, and the communication process with the computing terminal are all existing technologies.

[0079] Both the input feature determination module and the estimation module are performed on a computing terminal, which can be a CPU. The input feature determination module determines the running mode based on plantar pressure data and determines the corresponding input features for each running mode. It then normalizes the input features to eliminate dimensional differences. The input features consist of plantar pressure data and IMU data, including acceleration and angle data. The estimation module is based on the PatchTST model and includes a projection layer, a masking layer, a partitioning layer, a conditional embedding layer, an encoder, and a shear force dimension mapping layer connected in sequence. The projection layer maps the dimensions of the input features corresponding to different running modes to a fixed intermediate dimension to obtain the mapped features. The masking layer generates a validity mask corresponding to each running mode based on the intermediate dimension. The partitioning layer divides the mapped features into multiple temporal segments to obtain a segment matrix. The conditional embedding layer combines the validity mask and the segment matrix to filter invalid channels and select valid features. The encoder learns the temporal dependence of valid features through multi-head self-attention and uses a feedforward network to enhance the nonlinear fitting of the multi-head self-attention output. It then uses inverse recombination to recover the channel structure and obtain the encoded features. The shear force dimension mapping layer converts the encoded features into predicted shear force values.

[0080] Figure 4This is a time-series comparison chart demonstrating the overall performance of the method of this invention in predicting the performance during macroscopic continuous running. The chart records the complete dynamic process of plantar shear force changes within a 2000-millisecond time window. This interval covers approximately five complete continuous gait cycles performed by the subject at a normal jogging pace. The vertical axis of the chart displays values ​​reaching thousands of Newtons (peak values ​​exceeding 2500 N) and remaining positive throughout, representing the total amplitude of the composite shear force vector on the horizontal plane of the foot, rather than a component force in a specific direction. It macroscopically reflects the intensity of the overall horizontal frictional effect of the ground on the foot. A comparison of the blue solid line representing the calculation results of this invention with the red dashed line representing the measured values ​​from high-precision equipment shows a very high degree of agreement between the two over the entire 2-second span. Whether in the ground-touch push-off phase where shear force rises sharply or in the airborne phase where it rapidly drops to zero, this invention accurately tracks every peak, trough, and rapid change trend of the baseline value. This fully demonstrates the strong robustness and excellent overall accuracy of this invention in capturing the total shear force amplitude over a high dynamic range during long-term continuous monitoring.

[0081] Figure 5 It provides a microscopic comparison of three typical running landing patterns (forefoot strike, heel strike, and full foot strike), capturing a segment of a single gait cycle (approximately 150 milliseconds) to demonstrate the invention's ability to capture detailed shear force characteristics. Figure 4 In stark contrast, Figure 5 The numerical range of the middle vertical axis is small (-150 N to 300 N) and shows significant alternation between positive and negative values, indicating that the graph shows the shear force corresponding to the braking force (negative value area) and driving force (positive value area) generated in the forward and backward directions during running.

[0082] Figure 5 The three sub-figures (a), (b), and (c) clearly present the distinctly different force waveform fingerprints under different landing modes:

[0083] Figure 5 (a) The forefoot landing exhibits a rapid change in force. A large peak positive driving force is rapidly generated upon landing, followed by a sharp decrease into negative braking force, demonstrating the elasticity and rapid transition of forefoot landing. This invention accurately captures this extremely rapid and significant positive-negative transition.

[0084] Figure 5(b) The heel strike exhibits a typical bimodal structure (which may appear as a plateau phase under a specific directional projection). Initially, there is a relatively wide positive force plateau (or a small peak), followed by a rapid decline to a significant negative braking peak. This is typical of heel-first strike followed by forefoot push-off. This invention fits well the plateau phase fluctuations and the subsequent steep descent.

[0085] Figure 5 (c) The waveform of the full-foot landing is the most complex, showing an alternating trend of "negative-positive-negative", reflecting the complex force balance process when the entire foot lands at the same time. Even in the face of such complex non-monotonic changes, the prediction line still closely follows the baseline.

[0086] In all these complex and significantly different scenarios, the blue prediction curve can still accurately reproduce every subtle fluctuation and positive-to-negative transition point of the red baseline curve. This strongly proves that the model of the present invention not only has the ability to distinguish different time-state features, but also can reconstruct the dynamic details of the shear force component in a specific direction with high precision within an extremely short time window.

[0087] Figure 4 This invention demonstrates that it can accurately estimate the magnitude (amplitude) of the total shear force during running and its overall variation over time; and Figure 5 This further demonstrates that the present invention can delve into details, accurately distinguish different landing modes, and accurately predict the positive and negative conversions and waveform characteristics of shear force components (such as braking force and driving force) in specific directions. The combination of these two aspects fully demonstrates the high performance of the present invention.

[0088] In this invention, the terms "first," "second," and "third," etc., are used only to distinguish similar objects and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. The use of terms such as "upper," "lower," "left," "right," "front," and "rear" to indicate orientation or positional relationships is based on the orientation or positional relationships shown in the accompanying drawings and is only for the convenience of describing the invention, not to indicate or imply that the device referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation on the scope of protection of this invention. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0089] Furthermore, in the description of this application, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0090] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.

Claims

1. A method for plantar shear force estimation based on IMU and plantar pressure data, characterized in that: The method comprises the following steps: Step S1, synchronously collecting foot pressure data and IMU data during running, and pre-processing the collected data to complete the data and remove outliers; Step S2, determining the running mode according to the foot pressure data, determining the input features corresponding to each running mode, and normalizing the input features to eliminate dimensional differences, wherein the input features are composed of foot pressure data and IMU data, and the IMU data includes acceleration data and angle data; Step S3, inputting the input features into an estimation model based on the PatchTST model, wherein the estimation model comprises a projection layer, a mask layer, a division layer, a conditional embedding layer, an encoder and a shear force dimension mapping layer connected in turn, the projection layer uniformly maps the dimensions of the input features corresponding to different running modes to a fixed intermediate dimension to obtain mapping features, the mask layer generates validity masks corresponding to each running mode based on the intermediate dimension, the division layer divides the mapping features into multiple time sequence segments to obtain a segment matrix, the conditional embedding layer combines the validity masks and the segment matrix to shield invalid channels to screen effective features, the encoder learns the time sequence dependence of the effective features through multi-head self-attention, and uses its feedforward network to enhance the nonlinear fitting of the multi-head self-attention output to recover the channel structure through anti-reorganization to obtain encoded features, and the shear force dimension mapping layer converts the encoded features into shear force prediction values.

2. The method of plantar shear force estimation based on IMU and plantar pressure data according to claim 1, characterized in that: In the step S1, data is collected by using an insole, the insole is provided with a pressure sensor for collecting foot pressure data and an IMU module for collecting IMU data, and when collecting data, the pressure sensor is synchronously time-granted by taking the hardware clock of the IMU module as a reference, wherein the insole is provided with n pressure sensors to collect n-channel foot pressure data.

3. The method of plantar shear force estimation based on IMU and plantar pressure data according to claim 2, characterized in that: In the step S1, the data completion includes using a linear interpolation method to complete the foot pressure data and using a cubic spline interpolation method to complete the IMU data, and if the same type of data is missing for three consecutive time steps, the data is re-collected; the isolated forest algorithm is used to remove outliers from the foot pressure data and the IMU data.

4. The IMU and plantar pressure data based plantar shear force estimation method of claim 2 or 3, wherein: The step S2 specifically comprises the following steps: Step S21, Obtain At this moment and The value when and The running style is to land on the forefoot, when and The running style is to land on the heel, when and The running style involves landing on the entire foot, among which... , Indicates time t Pressure on the entire foot, This represents the sum of data collected by all pressure sensors on the forefoot at time t. This represents the sum of data collected by all pressure sensors on the heel at time t. This represents the basic pressure value for the forefoot. This refers to the baseline pressure value of the heel, which is the pressure data collected during takeoff. Step S22, when the running mode is forefoot landing, the corresponding input feature is represented as When the running mode is rearfoot landing, the corresponding input feature is represented as When the running mode is flatfoot landing, the corresponding input feature is represented as Wherein, is the pressure data collected by the i pressure sensors located on the forefoot of the insole, is the pressure data collected by the n-i pressure sensors located on the rearfoot of the insole, are the x-axis acceleration, y-axis acceleration and z-axis acceleration collected by the IMU module respectively, are the roll angle, roll angle and yaw angle collected by the IMU module respectively; Step S23, normalizing the input features by Min-Max normalization normalizing the input features .

5. The IMU and plantar pressure data based plantar shear force estimation method of claim 4, wherein: In the step S3, the projection layer adopts The input features are mapped to get mapped features , wherein is a projection weight matrix corresponding to the running type type, , represents the forefoot landing, represents the rearfoot landing, represents the full-foot landing, is a real number set with a dimension of , is a dimension of the input features corresponding to the running type type, is an intermediate dimension, is a projection bias vector corresponding to the running type.

6. The IMU and plantar pressure data based plantar shear force estimation method of claim 5, wherein: In the step S3, the mask layer is generated according to the formula generate a binary validity mask , is an indicator function, is a set of valid channels corresponding to the running mode type.

7. The IMU and plantar pressure data based plantar shear force estimation method of claim 6, wherein: In the step S3, the division layer first converts the mapping feature into a channel-first format , then sets the number of segments of a single channel as , and finally divides the time sequence by a sliding window to obtain a segment matrix , where is a tensor dimension permutation operation, is a training batch size, is a time sequence length of the mapping feature , is a sliding window segmentation function, is a segment length obtained by segmentation, is a sliding step.

8. The IMU and plantar pressure data based plantar shear force estimation method of claim 7, wherein: The effective feature filtered by the conditional embedding layer in the step S3 is represented as wherein, , , , represents Hadamard product, represents dimension expansion operation, is the dimension after mapping the low-dimensional feature of each segment to the Transformer model, , is the shared projection parameter, is the full-channel embedding matrix; the encoded feature obtained by the encoder is represented as wherein, , , is the layer normalization operation, and the feedforward network of the encoder is , represents the activation function, is the weight matrix corresponding to the multi-head attention MHA, is the weight matrix corresponding to the activation function, is the bias vector corresponding to the multi-head attention MHA, is the bias vector corresponding to the activation function, and the multi-head attention , is the concatenation function, and h represents the number of attention heads, , , , , , , , is the attention projection parameter, is the invalid combination attention mask, represents setting the attention weight of to a minimum value, , , is the tensor dimension reshaping operation, is the tensor dimension expansion operation, is the multi-head fusion parameter.

9. The IMU and plantar pressure data based plantar shear force estimation method of claim 8, wherein: The shear force prediction value obtained by the shear force dimension mapping layer in the step S3 is expressed as wherein, , , , are the weight and bias of the shear force dimension mapping layer, respectively, are the mapping weight and bias corresponding to the running mode, are the physical limit values of the Min-Max normalization.

10. Apparatus for plantar shear force estimation based on IMU and plantar pressure data for implementing the estimation method according to any one of claims 1 to 9, characterized in that: The insole, the input feature determination module and the estimation module are included, the insole is provided with a pressure sensor collecting plantar pressure data and an IMU module collecting IMU data, the input feature determination module judges the running mode according to the plantar pressure data, determines the input feature corresponding to each running mode, and normalizes the input feature to eliminate the dimensional difference, the input feature is composed of plantar pressure data and IMU data, the IMU data includes acceleration data and angle data, the estimation module is based on PatchTST model, including projection layer, mask layer, division layer, conditional embedding layer, encoder and shear force dimension mapping layer connected in turn, the projection layer uniformly maps the dimension of the input feature corresponding to different running modes to a fixed intermediate dimension to obtain a mapping feature, the mask layer generates an effectiveness mask corresponding to each running mode based on the intermediate dimension, the division layer divides the mapping feature into multiple time sequence segments to obtain a segment matrix, the conditional embedding layer combines the effectiveness mask and the segment matrix to shield invalid channels to screen effective features, the encoder learns the time sequence dependence of the effective features through multi-head self-attention, uses a feedforward network to enhance the nonlinear fitting of the multi-head self-attention output, and recovers the channel structure through anti-reorganization to obtain an encoded feature, and the shear force dimension mapping layer converts the encoded feature into a shear force prediction value.

Images