A method for predicting knee torque based on plantar pressure characteristics

Abstract

This invention provides a method for predicting knee joint torque based on plantar pressure characteristics. By collecting and processing gait and plantar pressure data, gait information data is converted into knee joint torque data, thereby constructing a neural network model of plantar pressure information data and knee joint torque data. The model is trained using one of three methods: task-focused, subject-focused, or combined. This allows the predicted value of knee joint torque to be obtained simply by inputting plantar pressure information data. The prediction method of this application has achieved excellent prediction accuracy in experimental verifications at three months, six months, and one year postoperatively.

Description

Technical Field

[0001] This invention relates to medical informatics and digital healthcare in the biomedical industry, specifically to a method for predicting knee joint torque based on plantar pressure characteristics. Background Technology

[0002] The anterior cruciate ligament (ACL) is a crucial ligament in the human knee joint. It originates from the posterior aspect of the medial surface of the lateral femoral condyle, passes through the intercondylar fossa, and inserts onto the anterior surface of the intercondylar eminence of the tibia. Its function is to allow the knee joint to move flexibly within the body's reasonable range of motion and to limit excessive anterior tibial movement, maintaining joint stability. In sports involving running, jumping, changing direction, and landing, such as basketball, soccer, and skiing, the ACL plays a vital role. The ACL's tissue composition is similar to ligaments in other parts of the body, consisting primarily of fibrous tissue. It possesses only a limited range of stiffness and strength, and can be damaged by excessive stretching or improper impact, even rupturing in severe cases. In recent years, with the popularization of fitness for all, the proportion of people participating in various sports has gradually increased. However, scientific exercise theories and warm-up methods have not been widely disseminated among the sports population. This leads many enthusiasts to apply inappropriate intensity and movements to their knee joints during exercise, resulting in excessive load on the anterior cruciate ligament, causing damage, tearing, or even rupture.

[0003] For patients with ruptured anterior cruciate ligaments (ACLs), the primary treatment currently used in the medical field is anterior cruciate ligament reconstruction (ACLR). This surgery uses tendons from other parts of the patient's body to connect the ruptured ACL, thereby restoring knee joint stability. Although this surgery can alleviate symptoms to some extent, gait and movement abnormalities persist after ACLR surgery, affecting the patient's postoperative knee health. Current research on abnormal gait after ACLR mainly relies on complete optical motion capture systems and three-dimensional force tables for data collection and analysis. However, it is well known that this research method is severely limited by equipment, only usable in specific indoor environments, and cannot measure and assess gait characteristics in daily life. Furthermore, this equipment is expensive, only available in some top-tier hospitals in first-tier cities, making it difficult to widely disseminate scientifically quantified gait characteristic detection and targeted rehabilitation guidance. There is an urgent need for a simpler and more practical method to analyze the biomechanical changes of lower limb joints during gait.

[0004] Plantar pressure distribution refers to the pressure distribution at different locations on the ground caused by the contact between the foot and the ground during walking. It is a spatial mapping of the mechanical information of the interaction between the human body and the ground on the sole of the foot. This information is an important carrier of human gait and can characterize many dynamic features behind the gait process. Previous studies on plantar pressure signals have been limited to the analysis of its static geometric features, such as the center trajectory line of plantar pressure and the foot's gait angle, neglecting the spatial distribution of plantar pressure. The magnitude and distribution of plantar pressure at different locations are closely related to human posture. Measuring and analyzing these indicators can be used to assess the health of the subject's feet and the force line distribution of the lower limb near the ankle. However, the deeper spatial characteristics behind plantar pressure distribution still need to be explored, and the relationship between plantar pressure information and lower limb knee joint torque urgently needs to be clarified. Currently, no institution or individual has been able to predict knee joint torque by analyzing the dynamic distribution of plantar pressure during walking. Summary of the Invention

[0005] Based on the above background, this invention aims to propose a method for predicting the knee joint torque of the lower limbs during human movement based on the plantar pressure distribution during walking. The specific implementation steps are as follows:

[0006] 1. A method for predicting knee joint torque based on plantar pressure characteristics, comprising the following steps:

[0007] S1. Gait and plantar pressure data are collected and processed in the laboratory to obtain gait information data and plantar pressure information data;

[0008] S2. Convert gait information data into knee joint torque data, which will be used as the reference ground value in the neural network model training in step S4.

[0009] S3. Construct a neural network model to map the relationship between plantar pressure data and knee joint torque data;

[0010] S4. Train and validate the neural network model in step S3, that is, input the plantar pressure information data collected in step S1 and the knee joint torque data obtained in step S2 into the neural network model in step S3.

[0011] S5. Input the plantar pressure information data into the neural network model trained in step S4 to obtain the predicted value of the knee joint torque.

[0012] Preferably, the acquisition laboratory in step S1 is configured as follows: the acquisition laboratory includes a three-dimensional optical motion capture system with 8 cameras and at least one 10-meter-long walkway, on which two three-dimensional force measuring platforms and one 2-meter-long plantar pressure measuring plate are placed in sequence; a Footscan plantar pressure acquisition system is installed.

[0013] Preferably, the steps for collecting knee joint dynamic change angle and torque data, gait data, and plantar pressure data in step S1 are as follows:

[0014] S1-C1, collecting static data, i.e., the subject stands in the data collection laboratory with his feet shoulder-width apart, his arms hanging naturally at his sides, maintaining the neutral position of the subtalar joint, and a total of three static data collections are performed.

[0015] S1-C2, Affix reflective markers, that is, fix 34 reflective markers on specific anatomical landmarks on both sides of the subject's body, including but not limited to: acromion, anterior superior iliac spine, posterior superior iliac spine, lateral thigh, lateral femur of mid-shaft, lateral femoral condyle, medial femoral condyle, tibial tuberosity, anterosuperior tibia, anteroinferior tibia, lateral lower leg, heel, lateral malleolus, medial malleolus, and the first, second, and fifth metatarsophalangeal joints. In addition, place one reflective marker on the subject's right scapula for subsequent calibration of the skeletal muscle model in the left-right direction.

[0016] S1-C3, Trial walking, which involves having the subject naturally step onto two force-measuring platforms placed in front of them at their normal cadence and stride, and marking the most suitable starting position during the trial walking;

[0017] S1-C4, Begin formal gait data collection, that is, the subject walks barefoot on the walking path at a self-selected speed, with each foot stepping on two force measurement platforms in the same gait cycle, and stepping on the plantar pressure collection plate placed behind in the next gait cycle.

[0018] S1-C5, Repeat steps S1-C1 to S1-C4 several times, changing the starting foot side during data acquisition.

[0019] The processing steps for plantar pressure data in step S1 are as follows:

[0020] S1-P1: Export the plantar pressure data acquired from the plantar pressure acquisition system in the format of the entire measurement plate image data.

[0021] S1-P2, each frame of the image is cropped. The cropping position is determined by the position of the footprints left by the left and right feet on the measuring plate during the gait cycle, so that the cropped pressure image contains exactly all the pixels that the left and right feet stepped on the measuring plate.

[0022] S1-P3. Zero-fill the cropped pressure image, adjust it to a uniform size, and stack it according to the different channels of the image to finally obtain a plantar pressure distribution image with appropriate size, fewer zero data, and storing the pressure data of the left and right feet in two channels respectively.

[0023] Preferably, the step of converting gait information data into knee joint torque data in step S2 is as follows:

[0024] S2-1. The spatial coordinate change data obtained by using reflective markers and the plantar pressure data obtained by the force measuring table are filtered and passed through low-pass filters to remove the influence of high-frequency noise.

[0025] S2-2. Generate a human skeletal muscle model using the processed spatial coordinate change data and plantar pressure data;

[0026] S2-3. Obtain the knee joint torque variation curve using a human skeletal muscle model;

[0027] S2-4. Normalize the knee joint torque data by dividing each knee joint torque by the product of the subject's weight and height at the time of the test to obtain the normalized knee joint torque data.

[0028] S2-5. The normalized knee joint torque data described above will be used as the reference ground truth in the subsequent training of the neural network model.

[0029] Preferably, the step of constructing the neural network model in step S3 is as follows:

[0030] S3-1. Expand the plantar pressure distribution image obtained in steps S1-P3 at the channel level, that is, calculate the difference between the plantar pressure distribution of the left and right feet at each moment and the previous frame, and add it to the channel of the input image in the form of stacking to obtain four-channel plantar pressure distribution image data.

[0031] S3-2. Perform data augmentation on the four-channel plantar pressure distribution image obtained in step S3-1, that is, identify the smallest rectangle where the footprint is located in the above plantar pressure distribution image, and translate the rectangle in the horizontal and vertical directions to achieve four times data augmentation.

[0032] S3-3. Predict the three components of flexion, adduction and internal rotation of the hip, knee and ankle joints on the left and right sides respectively. Build and train the corresponding joint torque prediction model for each joint. At this time, the three components of the joint torque are predicted and output by one model.

[0033] The torque prediction model for a specific joint consists of four modules. The first two modules are two-dimensional convolutional modules, which are responsible for preliminary spatial feature extraction and downsampling of the plantar pressure distribution image. The third module is a ConvLSTM layer, which performs spatiotemporal coupling feature extraction on the processed plantar pressure data. The fourth module consists of three multilayer perceptrons, which respectively complete the final fitting of the three components of the joint torque: flexion, adduction, and internal rotation. The input of this neural network model is several plantar pressure distribution image data, and the output is the three components of the joint torque.

[0034] The model described above, which derives knee joint torque data from plantar pressure distribution images, is the constructed neural network model.

[0035] Preferably, the neural network model training and validation described in step S4 involves three different training methods. The first is task-focused training, which focuses on the model learning the common gait characteristics of all subjects during the same postoperative period, thereby enabling the prediction from the subject's plantar pressure data to their knee joint torque data. The second is subject-focused training, which focuses on the specific gait patterns exhibited by the subject during their rehabilitation process. The training set for this method consists of all plantar pressure data and knee joint torque data of the subject before the rehabilitation time point, thereby enabling the learning of the subject's unique gait characteristics. The third is combined training, which combines the above two training methods, allowing the model to learn the common rehabilitation gait characteristics of the subject during a certain rehabilitation period, while simultaneously learning the subject's unique gait characteristics.

[0036] The innovations and beneficial effects of this invention are as follows:

[0037] 1. This paper proposes a novel paradigm for predicting lower limb joint torque using a sequence of plantar pressure distribution images as input. Furthermore, the correlation coefficient R for predicting knee adduction torque reaches 0.93, validating the rationality of the dataset (as shown in Table 1). Figure 1 As shown in the figure, this is a major innovation of the present invention.

[0038] 2. Through rigorous clinical verification, this invention can more conveniently predict knee joint torque in patients with abnormal gait after ACLR surgery, and can effectively provide patients with anterior cruciate ligament rupture with a scientific home rehabilitation plan with real-time feedback, which has application value and significance in the field of rehabilitation.

[0039] 3. It can help high-risk groups such as the elderly to assess the severity of their knee osteoarthritis and other diseases in a timely manner, thereby improving people's quality of life and health.

[0040] 4. The method of this invention bypasses certain foreign rehabilitation technology barriers, reduces the cost of scientific rehabilitation programs, and is more widely applicable.

[0041] 5. This is the first time that knee flexion moment, adduction moment and rotation moment can be predicted simultaneously. Most related works only focus on flexion moment and ignore the components in the other two directions.

[0042] 6. Two special training set data partitioning methods, subject-focused and joint training, are proposed, which improve the prediction accuracy of the model in certain situations. Attached Figure Description

[0043] Figure 1 : Knee joint torque prediction curves at six months post-surgery using the combined model;

[0044] Figure 2 : Flowchart of the main steps of the prediction scheme of this invention;

[0045] Figure 3 Schematic diagram of the data collection laboratory;

[0046] Figure 4 : Schematic diagram of collecting static data and pasting reflective markers;

[0047] Figure 5 Gait acquisition motion breakdown diagram and example diagram of detection data;

[0048] Figure 6 : Schematic diagram of plantar pressure data processing;

[0049] Figure 7 Human skeletal muscle model;

[0050] Figure 8 : Schematic diagram of joint angle changes before anterior cruciate ligament reconstruction;

[0051] Figure 9 Schematic diagram of joint torque changes before anterior cruciate ligament reconstruction;

[0052] Figure 10 Schematic diagram of joint angle changes 3 months after anterior cruciate ligament reconstruction surgery;

[0053] Figure 11 Schematic diagram of joint torque changes 3 months after anterior cruciate ligament reconstruction;

[0054] Figure 12 Schematic diagram of joint angle changes 6 months after anterior cruciate ligament reconstruction surgery;

[0055] Figure 13 Schematic diagram of joint torque changes 6 months after anterior cruciate ligament reconstruction;

[0056] Figure 14Schematic diagram of joint angle changes 1 year after anterior cruciate ligament reconstruction surgery;

[0057] Figure 15 Schematic diagram of joint torque changes 1 year after anterior cruciate ligament reconstruction;

[0058] Figure 16 (a) Structure diagram of the knee joint torque prediction model; (b) Structure diagram of the Conv2D module; (c) Schematic diagram of the MLP structure.

[0059] Figure 17 : Schematic diagram of task-focused training method;

[0060] Figure 18 : Schematic diagram of subject-focused training methods;

[0061] Figure 19 : Schematic diagram of joint training method. Detailed Implementation

[0062] The solution of the present invention will be further described below. Unless otherwise specified, the embodiments and features of the present invention can be combined with each other.

[0063] The main steps of the prediction method of this invention are as follows: Figure 2 As shown, the knee joint torque curve obtained using the knee joint torque prediction method in this application is shown in the figure. Figure 1 The predicted values ​​of knee joint torque obtained by the joint model are shown in the table below. The table shows the changes in knee joint torque at different postoperative periods. The closer the mean R value (e.g., 0.76, 0.92, 0.94) is to 1, the smaller the deviation between the verified value and the true value. The smaller the standard deviation (e.g., ±0.07, ±0.06, ±0.02), the smaller the data dispersion.

[0064] variable 3 months after surgery 6 months after surgery 1 year after surgery Left hip flexion moment 0.76 ± 0.07 0.92 ± 0.06 0.94 ± 0.02 Left hip adduction moment 0.95 ± 0.02 0.97 ± 0.01 0.98 ± 0.01 Left knee flexion moment 0.8 ± 0.13 0.9 ± 0.04 0.9 ± 0.05 Left knee adduction moment 0.92 ± 0.06 0.93 ± 0.05 0.92 ± 0.06 Left knee rotational torque 0.79 ± 0.08 0.9 ± 0.04 0.9 ± 0.04 Left ankle dorsiflexion moment 0.98 ± 0.01 0.99 ± 0.01 0.99 ± 0 Right hip flexion moment 0.91 ± 0.04 0.93 ± 0.02 0.94 ± 0.03 Right hip adduction moment 0.96 ± 0.02 0.97 ± 0.02 0.98 ± 0.01 Right knee flexion moment 0.79 ± 0.12 0.88 ± 0.03 0.9 ± 0.04 Right knee adduction moment 0.83 ± 0.13 0.91 ± 0.05 0.87 ± 0.08 Right knee rotational torque 0.8 ± 0.11 0.88 ± 0.05 0.89 ± 0.04 Right ankle dorsiflexion moment 0.98 ± 0.02 0.99 ± 0.01 0.99 ± 0.00

[0065] Example 1: Implementation of the setup of a data collection laboratory

[0066] Gait and plantar pressure data were collected in the laboratory for subjects at four time points: before anterior cruciate ligament surgery, three months after surgery, six months after surgery, and one year after surgery. Figure 3 The data acquisition laboratory is equipped with a 3D optical motion capture system with 8 cameras (100Hz, Vicon). Four cameras are shown in the image, with four more positioned opposite them. A 10-meter-long walkway is set up on the floor of the acquisition laboratory. Figure 3As shown on the ground, two three-dimensional force platforms (1000Hz, AMTI) and a 2-meter-long plantar pressure measurement plate were placed side by side on the pedestrian walkway. The plantar pressure data was collected using the Footscan system (126Hz, RSscan International).

[0067] Example 2: Implementation of Gait and Plantar Pressure Data Acquisition Procedures

[0068] like Figure 4 As shown, firstly, static data was collected. Subjects stood in the center of the data collection laboratory with their feet shoulder-width apart and arms naturally placed at their sides, maintaining a neutral subtalar joint position. Three static tests were conducted to collect static data used to define the coordinate system of the skeletal segments. Secondly, reflective markers were affixed. Thirty-four reflective markers were bilaterally fixed to specific anatomical landmarks on the subject, including: acromion, anterior superior iliac spine, posterior superior iliac spine, lateral thigh, mid-thigh lateral femur, lateral femoral condyle, medial femoral condyle, tibial tuberosity, anterosuperior tibia, anteroinferior tibia, lateral lower leg, heel, lateral malleolus, medial malleolus, and the first, second, and fifth metatarsophalangeal joints. In addition, an extra reflective marker was placed on the subject's right scapula for subsequent calibration of the skeletal muscle model in the left-right direction. Thirdly, a trial walk was conducted. Subjects, under guidance, were required to naturally step onto two adjacent force-measuring platforms within their normal stride frequency and stride length. The most suitable starting position for the subject during the trial walk was determined for subsequent reference. Next, the formal gait data collection began, such as... Figure 5 As shown, subjects walked barefoot on a walking path at a self-selected speed, placing each foot on two adjacent force measurement platforms within the same gait cycle, and then on a plantar pressure measurement plate placed behind them in the next gait cycle. Finally, subjects repeated the walking test ten times, varying the starting foot to increase the diversity of the dataset.

[0069] Example 3: Implementation of Plantar Pressure Data Processing

[0070] like Figure 6As shown, firstly, the plantar pressure data is exported from the Footscan system. The export format is the entire measurement plate. Secondly, the entire measurement plate pressure distribution image of each frame is cropped into two smaller pressure images. The cropping position is determined by the footprint positions left by the left and right feet on the measurement plate during that gait cycle, so that these two smaller pressure images respectively contain all the pixels where the left and right feet stepped on the measurement plate. The footprint images in this process can be obtained by taking the maximum value of each pixel in the plantar pressure data of the measurement plate during the gait cycle. Thirdly, the cropped images are zero-paddingd to a uniform length and width, and stacked according to the different channels of the images, finally obtaining a pressure distribution image of appropriate size, fewer zero data, and storing the pressure distribution of the left and right feet in two separate channels.

[0071] Example 4: Implementation of the steps for converting gait information data into knee joint torque data

[0072] First, the spatial coordinate change data of the reflective markers output by the optical motion capture system and the data obtained from the force measuring table are both filtered after export, passing through Butterworth low-pass filters with cutoff frequencies of 12Hz and 100Hz respectively, to eliminate the influence of some high-frequency noise. For example... Figure 7 As shown, secondly, the size of the generalized human skeletal muscle model is determined and scaled using the coordinate data of reflective markers at a specific moment through the processed spatial coordinate change data and plantar pressure data, generating the skeletal muscle model of the subject in this data. Thirdly, the joint angle and joint torque change curves are then calculated using inverse kinematics and inverse dynamics modules, such as... Figures 8-15 As shown. Fourth, to ensure the comparability of joint torque data between different individuals, the joint torques are normalized. This involves dividing each joint torque by the product of the subject's body weight (BW) and height (BH) at the time of testing, resulting in a normalized joint torque value. Fifth, the joint torques calculated at this stage will serve as the reference ground truth in subsequent model training.

[0073] Example 5: Implementation of the steps for constructing a neural network model

[0074] First, the preprocessed plantar pressure distribution images of the left and right feet stacked together are augmented at the channel level. The difference between the plantar pressure distribution of each foot at each moment and the previous frame is calculated and added to the input image channels in a stacked form, resulting in four-channel plantar pressure distribution image data. Second, data augmentation is performed on the image, that is, the smallest rectangle containing the footprint is identified in the above plantar pressure distribution image, and the rectangle is translated in the horizontal and vertical directions, ultimately achieving four times data augmentation. Third, the three components of flexion, adduction, and internal rotation of the hip, knee, and ankle joints on the left and right sides are predicted, totaling 18 prediction indicators. A corresponding joint torque prediction model is built and trained for each joint, and at this time, the three components of the joint torque are predicted and output by a single model.

[0075] like Figure 16 As shown in (a), the torque prediction model for a specific joint mainly consists of four modules:

[0076] The first two modules are two-dimensional convolutional modules (Conv2D Block), which are responsible for preliminary spatial feature extraction and downsampling of the plantar pressure distribution image, and then maintain the sequence independence of each test by Split by Test.

[0077] The third module is the ConvLSTM layer, which extracts spatiotemporal coupled features from the processed foot pressure video. ConvLSTM combines the characteristics of convolution and LSTM, and can process spatial features (from Conv2D Block) and capture the dynamic changes of plantar pressure over time (such as the movement of the pressure center during movement and the temporal process of pushing off the ground), outputting a feature sequence that integrates spatiotemporal information.

[0078] The final module consists of three Multi-Layer Perception (MLP) machines, each performing a final fitting of the three components of the joint moment: flexion, adduction, and internal rotation. The model input consists of plantar pressure distribution images from various time points across several test batches. After processing by the first two modules, the resulting images are segmented, with all time steps from the same test grouped together and stacked to obtain several video datasets. These batches of video data are then input into a ConvLSTM layer, and the output of this layer is copied three times. These copies are then fitted using three different MLP machines to obtain the three components of the joint moment.

[0079] in Figure 16(b) demonstrates the spatial feature extraction and dimensionality reduction of a single-frame plantar pressure map, providing high-quality spatial features for subsequent temporal modeling. In (b), Conv2D (two-dimensional convolution) captures local patterns of pressure distribution (such as key areas like heel strike and forefoot push-off); BN (Batch Normalization) stabilizes the training process and accelerates convergence; ELU (Exponential Linear Unit Activation Function) introduces nonlinearity and enhances the model's expressive power; and MaxPooling 2D (Max Pooling) reduces dimensionality and retains the most significant spatial features.

[0080] Figure 16 (c) demonstrates how high-dimensional features are flattened and mapped to the target output space (such as the torque dimension) through a fully connected layer, achieving the mapping from spatiotemporal features to torque curves. Three independent MLPs (Multilayer Perceptrons) correspond to the torque prediction of the three degrees of freedom of the knee joint, respectively. Each MLP maps the spatiotemporal features output by ConvLSTM to the corresponding torque curves (Flexion, Adduction, Rotation), enabling simultaneous prediction of multiple tasks.

[0081] Example 6: Implementation of Neural Network Model Training and Validation

[0082] To better understand the relationship between plantar pressure and joint torque in different subjects and at different levels of rehabilitation, this invention proposes three different training methods: such as Figure 17 As shown, Task-Focused Training, such as Figure 18 As shown, subject-focused training and such Figure 19 As shown, there are two training methods: Combined and Combined Training. Taking the rehabilitation gait six months post-surgery as an example, Task-Focused Training emphasizes learning the common gait characteristics of all subjects' rehabilitation gait six months post-surgery. It uses data from all other subjects (excluding the test subject) six months post-surgery as the training set, thus achieving predictions from plantar pressure to joint torque in the test subject. Subject-Focused Training focuses only on the rehabilitation progress of the test subject and their specific gait abnormalities. In this case, the training set consists of all test data prior to this rehabilitation time point, including pre-operative plantar pressure and joint torque data. For example, using six-month post-surgery rehabilitation data, the training set comprises data collected from three months post-surgery and pre-surgery, enabling the learning of the subject's unique gait abnormality patterns. Combined Training combines the two training sets, allowing the model to learn the common characteristics of all subjects' six-month post-surgery rehabilitation gait while also considering the subject's specific gait abnormality trends.

[0083] The above embodiments are merely representative examples of specific implementations of the present invention, and are only used to provide detailed illustrative examples of the technical solutions. They are not intended to limit the scope of protection of the present invention.

Claims

1. A method for predicting knee joint torque based on plantar pressure characteristics, comprising the following steps: S1. Gait and plantar pressure data are collected and processed in the laboratory to obtain gait information data and plantar pressure information data; S2. Convert gait information data into knee joint torque data, which will be used as the reference ground value in the neural network model training in step S4. S3. Construct a neural network model to map the relationship between plantar pressure data and knee joint torque data; S4. Train and validate the neural network model in step S3, that is, input the plantar pressure information data collected in step S1 and the knee joint torque data obtained in step S2 into the neural network model in step S3. S5. Input the plantar pressure information data into the neural network model trained in step S4 to obtain the predicted value of the knee joint torque.

2. The knee joint torque prediction method as described in claim 1, characterized in that, The data acquisition laboratory in step S1 is set up as follows: The data acquisition laboratory includes a three-dimensional optical motion capture system with 8 cameras and at least one 10-meter-long walkway. Two three-dimensional force measuring platforms and one 2-meter-long plantar pressure measuring plate are placed on the walkway in sequence; the Footscan plantar pressure acquisition system is installed.

3. The knee joint torque prediction method as described in claim 2, characterized in that, The gait and plantar pressure data acquisition steps in step S1 are as follows: S1-C1, collecting static data, i.e., the subject stands in the data collection laboratory with his feet shoulder-width apart, his arms hanging naturally at his sides, maintaining the neutral position of the subtalar joint, and a total of three static data collections are performed. S1-C2, Affix reflective markers, that is, fix 34 reflective markers on specific anatomical landmarks on both sides of the subject's body, including but not limited to: acromion, anterior superior iliac spine, posterior superior iliac spine, lateral thigh, lateral femur of mid-shaft, lateral femoral condyle, medial femoral condyle, tibial tuberosity, anterosuperior tibia, anteroinferior tibia, lateral lower leg, heel, lateral malleolus, medial malleolus, and the first, second, and fifth metatarsophalangeal joints. In addition, place one reflective marker on the subject's right scapula for subsequent calibration of the skeletal muscle model in the left-right direction. S1-C3, Trial walking, which involves having the subject naturally step onto two force-measuring platforms placed in front of them at their normal cadence and stride, and marking the most suitable starting position during the trial walking; S1-C4, Begin formal gait data collection, that is, the subject walks barefoot on the walking path at a self-selected speed, with each foot stepping on two force measurement platforms in the same gait cycle, and stepping on the plantar pressure collection plate placed behind in the next gait cycle. S1-C5, Repeat steps S1-C1 to S1-C4 several times, changing the starting foot side during data acquisition; The processing steps for plantar pressure data in step S1 are as follows: S1-P1: Export the plantar pressure data acquired from the plantar pressure acquisition system in the format of the entire measurement plate image data. S1-P2, each frame of the image is cropped. The cropping position is determined by the position of the footprints left by the left and right feet on the measuring plate during the gait cycle, so that the cropped pressure image contains exactly all the pixels that the left and right feet stepped on the measuring plate. S1-P3. Zero-fill the cropped pressure image, adjust it to a uniform size, and stack it according to the different channels of the image to finally obtain a plantar pressure distribution image with appropriate size, fewer zero data, and storing the pressure data of the left and right feet in two channels respectively.

4. The knee joint torque prediction method as described in claim 3, characterized in that, The step of converting gait information data into knee joint torque data in step S2 is as follows: S2-1. The spatial coordinate change data obtained by using reflective markers and the plantar pressure data obtained by the force measuring table are filtered and passed through low-pass filters to remove the influence of high-frequency noise. S2-2. Generate a human skeletal muscle model using the processed spatial coordinate change data and plantar pressure data; S2-3. Obtain the knee joint torque variation curve using a human skeletal muscle model; S2-4. Normalize the knee joint torque data by dividing each knee joint torque by the product of the subject's weight and height at the time of the test to obtain the normalized knee joint torque data. S2-5. The normalized knee joint torque data described above will be used as the reference ground truth in the subsequent training of the neural network model.

5. The knee joint torque prediction method as described in claim 4, characterized in that, The steps for constructing the neural network model in step S3 are as follows: S3-1. Expand the plantar pressure distribution image obtained in steps S1-P3 at the channel level, that is, calculate the difference between the plantar pressure distribution of the left and right feet at each moment and the previous frame, and add it to the channel of the input image in the form of stacking to obtain four-channel plantar pressure distribution image data. S3-2. Perform data augmentation on the four-channel plantar pressure distribution image obtained in step S3-1, that is, identify the smallest rectangle where the footprint is located in the above plantar pressure distribution image, and translate the rectangle in the horizontal and vertical directions to achieve four times data augmentation. S3-3. Predict the three components of flexion, adduction and internal rotation of the hip, knee and ankle joints on the left and right sides respectively. Build and train the corresponding joint torque prediction model for each joint. At this time, the three components of the joint torque are predicted and output by one model. The torque prediction model for a specific joint consists of four modules. The first two modules are two-dimensional convolutional modules, which are responsible for preliminary spatial feature extraction and downsampling of the plantar pressure distribution image. The third module is a ConvLSTM layer, which performs spatiotemporal coupling feature extraction on the processed plantar pressure data. The fourth module consists of three multilayer perceptrons, which respectively complete the final fitting of the three components of the joint torque: flexion, adduction, and internal rotation. The input of this neural network model is several plantar pressure distribution image data, and the output is the three components of the joint torque. The model described above, which derives knee joint torque data from plantar pressure distribution images, is the constructed neural network model.

6. The knee joint torque prediction method as described in claim 5, characterized in that, The neural network model training and verification described in step S4 has three different training methods. One of them is task-focused training, which focuses on the model learning the common gait characteristics of all subjects' rehabilitation gait during the same period after surgery, thereby realizing the prediction from the subject's plantar pressure information data to their knee joint torque data. The second method is subject-focused training, which focuses on the specific gait patterns exhibited by the subject during their rehabilitation process. The training set for this method consists of all plantar pressure and knee torque data of the subject before the rehabilitation time point, thereby learning the subject's unique gait characteristics. The third method is combined training, which combines the above two training methods to enable the model to learn the common rehabilitation gait characteristics of the subject during a certain rehabilitation period, while also learning the subject's unique gait characteristics.

Images