Lower limb exercise load assessment method and system based on multi-mode wearable device

Abstract

The invention discloses a lower limb exercise load assessment method and system based on multi-mode wearable equipment, and the method comprises the steps: 1) collecting exercise signals, three-dimensional track information and actual ground reaction force when a testee runs at different set speeds, and synchronizing the exercise signals, the three-dimensional track information and the actual ground reaction force through timestamps; 2) calculating the tibia force of the testee; 3) training a tibia force and vertical ground reaction force estimation model; 4) inputting the motion signals collected in real time into the tibia force and vertical ground reaction force prediction model to obtain predicted tibia force and vertical ground reaction force; and 5) calculating to obtain a lower limb exercise load evaluation index. The system comprises a pressure insole device, a pressure leg sleeve device, a surface myoelectricity device, an inertial sensor device, a training data acquisition module, a tibia force prediction model construction module, a vertical ground reaction force prediction model construction module, a vertical ground reaction force prediction module, a tibia force prediction module and a lower limb exercise load evaluation module. Compared with other methods, the method provided by the invention has higher systematicness and scientificity.

Description

Technical Field

[0001] This invention relates to the field of smart wearables, and more specifically to a method and system for assessing lower limb exercise load based on a multimodal wearable device. Background Technology

[0002] In daily training and mass fitness activities, sports injuries caused by high-intensity and repetitive impacts have become a significant health problem. Among these, musculoskeletal system injuries caused by unscientific training are dominant, affecting not only combat readiness but also significantly increasing the medical burden. The tibia, as the most vulnerable part to fatigue injuries, currently relies heavily on costly and complex biomechanical platforms for load monitoring, which are overly dependent on sophisticated instruments and specialized knowledge. The few existing methods for estimating exercise load based on wearable device data generally employ machine learning models. These methods place IMU devices on the anteromedial aspect of the right distal tibia to monitor vertical / synthetic tibial impacts, or place inertial sensors and insole pressure sensors on the shoe to calculate peak tibial force using regression algorithms.

[0003] However, current research only involves estimating peak tibial load or instantaneous tibial impact, failing to achieve real-time, full-process load monitoring during exercise. The dynamic changes in tibial load during exercise are thus lost, limiting our ability to identify and prevent tibial fatigue injuries in their early stages. Existing datasets for estimating lower limb load using laboratory equipment are scarce, lacking a complete data acquisition scheme, preprocessing, and musculoskeletal model-based estimation process, making it difficult to accurately obtain tibial load during exercise. Furthermore, existing wearable device methods for estimating tibial force fail to capture hidden patterns and dependencies in the signal, resulting in unsatisfactory overall prediction performance. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for assessing lower limb exercise load based on a multimodal wearable device.

[0005] The technical solution adopted to achieve the technical objective of this invention is as follows: a method for assessing lower limb exercise load based on a multimodal wearable device, comprising the following steps:

[0006] 1) Place cursor points at different joint points on the subjects, put on pressure insoles, pressure leg sleeves, inertial sensors, and surface electromyography devices, and have the subjects run at different paces on a treadmill.

[0007] 2) Use a pressure insole device to collect the plantar pressure of the subjects when they run at different paces.

[0008] The study used a pressure leg sleeve device to collect surface muscle force signals of the calf muscles when subjects ran at different paces.

[0009] Ankle acceleration signals and mid-tibial acceleration signals were collected from subjects running at different paces using inertial sensing devices.

[0010] The surface electromyography (EMG) device was used to collect surface EMG signals of the tibialis anterior muscle from the subjects while they were running at different paces.

[0011] An optical motion capture system was used to collect three-dimensional trajectory information of subjects at different joint points while they ran at different paces.

[0012] The actual ground reaction force of the subjects was collected using a treadmill at different paces.

[0013] 3) Synchronize three-dimensional trajectory information, actual ground reaction force, calf muscle surface force signal, ankle joint acceleration signal, mid-tibialis acceleration signal, tibialis anterior muscle surface electromyography signal and plantar pressure through timestamps.

[0014] 4) Calculate the tibial force of the subject based on the synchronized three-dimensional trajectory information and the actual ground reaction force.

[0015] 5) Based on the synchronous calf muscle surface force signal, ankle joint acceleration signal, tibial mid-shaft acceleration signal, tibialis anterior muscle surface electromyography signal, plantar pressure, and tibial force, a tibial force training sample set is constructed, and the model is trained using the tibial force training sample set to obtain a tibial force prediction model.

[0016] A training sample set for vertical ground reaction force is constructed based on synchronized ankle joint acceleration signals, plantar pressure, and actual ground reaction force. The model is then trained using this training sample set to obtain a vertical ground reaction force prediction model.

[0017] 6) Input the real-time collected surface muscle force signals of the calf muscles, ankle joint acceleration signals, mid-tibial acceleration signals, surface electromyography signals of the tibialis anterior muscle, and plantar pressure into the tibial force prediction model to obtain the predicted tibial force.

[0018] The real-time collected ankle joint acceleration signal and plantar pressure are input into the vertical ground reaction force prediction model to obtain the predicted vertical ground reaction force.

[0019] 7) Lower limb exercise load assessment indexes are obtained based on predicted tibial force and predicted vertical ground reaction force.

[0020] Furthermore, the steps for calculating the subject's tibial force based on synchronized three-dimensional trajectory information and actual ground reaction force are as follows:

[0021] 4.1) Scale the standard musculoskeletal model to match the subject's size to obtain the subject's musculoskeletal model.

[0022] 4.2) Input the synchronized three-dimensional trajectory information into the subject's musculoskeletal model to obtain the displacement information of different joints when the subject runs at different paces.

[0023] The joints from which the three-dimensional trajectory information is collected include the hip joint, knee joint, and ankle joint.

[0024] The displacement information at different joint points includes the joint's position, velocity, and acceleration.

[0025] 4.3) Based on the displacement information at different joint points and the actual ground reaction force, the net joint torque at each moment is decomposed into individual muscle forces, and the muscle force of the peripheral tibial muscle group is calculated by minimizing the sum of squares of muscle activation.

[0026] 4.4) Construct a finite element model of the tibia.

[0027] 4.5) Input the muscle force of the peripheral tibial muscle group and the actual ground reaction force into the tibial finite element model to calculate the Von Misesc stress of the tibia at different times.

[0028] Furthermore, the training process of the tibial force prediction model includes training in a laboratory environment and training in a real running environment.

[0029] The training process of the vertical ground reaction force prediction model includes training in a laboratory environment and training in a real running environment.

[0030] Furthermore, the tibial force prediction model trained in the laboratory environment includes a preprocessing module I, a TCN module I, a Transformer encoder module I, and a prediction head I.

[0031] The preprocessing module I is used to preprocess the synchronous calf muscle surface force signal, ankle joint acceleration signal, tibia mid-shaft acceleration signal, tibialis anterior muscle surface electromyography signal, and plantar pressure to obtain preprocessed data I.

[0032] The preprocessing includes filtering, downsampling, and dividing the data into time windows.

[0033] The TCN module I is used to extract features from the preprocessed data I to obtain temporal data I, and to process the temporal data I using a one-dimensional convolutional network architecture to obtain embedded features I.

[0034] The Transformer encoder module I processes the embedded feature I by stacking several encoders to obtain the encoded feature I.

[0035] The prediction head I aggregates and normalizes the encoded features I before feeding them into the linear layer to obtain the predicted tibial force.

[0036] The vertical ground reaction force prediction model trained in the laboratory environment includes a preprocessing module II, a TCN module II, a Transformer encoder module II, and a prediction head II.

[0037] The preprocessing module II is used to preprocess the synchronized ankle joint acceleration signal and plantar pressure to obtain preprocessed data II.

[0038] The preprocessing includes filtering, downsampling, and dividing the data into time windows.

[0039] The TCN module II is used to extract features from the preprocessed data II to obtain temporal data II, and to process the temporal data II using a one-dimensional convolutional network architecture to obtain embedded features II.

[0040] The Transformer encoder module II processes the embedded feature II by stacking several encoders to obtain the encoded feature II.

[0041] The prediction head II aggregates and normalizes the encoded features II before passing them into the linear layer to obtain the predicted vertical ground reaction force.

[0042] Furthermore, the step of processing the embedded features by stacking several encoders to obtain the encoded features is as follows:

[0043] b1 uses LayerNorm to normalize the embedded features, resulting in normalized feature X.

[0044] b2 uses three fully connected layers to map the normalized feature X into three vectors of the same shape: query vector Q, key vector K, and value vector V.

[0045] The query vector Q, key vector K, and value vector V are shown below:

[0046] Q = XW Q K = XW K V = XW V (1)

[0047] In the formula, W Q W K W V Both are weight matrices.

[0048] b3 calculates the attention weights based on the query vector Q and the key vector K.

[0049] The attention weights are as follows:

[0050]

[0051] In the formula, Attention(Q,K) represents the attention weights, and softmax is the normalization function. d represents the feature dimension, and K... T This is the transpose of the key vector K.

[0052] b4 uses attention weights to perform a weighted summation of the value vector V, obtaining the output of each encoder.

[0053] The output of each encoder is shown below:

[0054] output (h) =Attention(Q,K)·V (3)

[0055] In the formula, h is the encoder number, and output is... (h) This represents the output of the h-th encoder. Attention(Q,K) represents the attention weights.

[0056] The encoding features are as follows:

[0057] MultiHead(Q,K,V)=Concat(output (1) ,…,output (num_heads) W O (4)

[0058] In the formula, MultiHead(Q,K,V) represents the encoding features. O This is a linear projection matrix. `num_heads` is the total number of encoders. `Concat` is the concatenation function.

[0059] b5 concatenates the outputs of each encoder together and projects them back to the original embedding dimension through a linear transformation to obtain the encoded features.

[0060] Furthermore, the training process for the tibial force prediction model under real running conditions is as follows:

[0061] c1 acquires pseudo-labeled data generated in real running environments and training data in laboratory environments.

[0062] c2 is a tibial force prediction model trained in a laboratory environment. It uses pseudo-labeled data generated in a real running environment and data trained in a laboratory environment to perform semi-supervised optimization on the model, resulting in a tibial force prediction model trained in a real running environment.

[0063] The training process for the vertical ground reaction force prediction model in a real running environment is as follows:

[0064] d1 acquires pseudo-labeled data generated in a real running environment and training data in a laboratory environment.

[0065] d2 is a vertical ground reaction force prediction model trained in a laboratory environment. The model is semi-supervised by using pseudo-label data generated in a real running environment and data trained in a laboratory environment to obtain a vertical ground reaction force prediction model trained in a real running environment.

[0066] Furthermore, the steps for obtaining pseudo-label data generated in a real running environment are as follows:

[0067] e1 acquires real running data from different runners and then constructs different outdoor samples.

[0068] e2 leverages the randomness of Monte Carlo Dropout technology to perform multiple forward propagations on outdoor samples, thereby constructing the pseudo-label prediction distribution for different outdoor samples.

[0069] e3 calculates the pseudo-label mean and prediction variance for each outdoor sample, as shown below:

[0070]

[0071] In the formula, i is the forward propagation index, and I is the total number of samples. The pseudo-label prediction result for the i-th forward propagation. σ is the pseudo-label mean. 2 e4 calculates the coefficient of variation for each outdoor sample, as shown below:

[0072]

[0073] In the formula, σ is the standard deviation of the pseudo-label. Here, is the pseudo-label mean, and CV is the coefficient of variation. e5 maps the coefficient of variation of each outdoor sample to a confidence score, as shown below:

[0074]

[0075] In the formula, conf represents the confidence score, and CV represents the confidence score. low For the minimum coefficient of variation, CV high is the maximum value of the coefficient of variation, and clip is the limiting function.

[0076] e6 selects outdoor samples with confidence scores greater than a preset threshold as pseudo-label data generated in a real running environment.

[0077] Furthermore, the loss function used in the semi-supervised optimization is as follows:

[0078]

[0079] In the formula, This is the joint loss. α and λ are both hyperparameter weights.

[0080] Among them, the loss of laboratory data Loss of pseudo-label data Consistency loss As shown below:

[0081]

[0082] In the formula, t represents time. T represents the number of time intervals. y t This refers to the actual tibial force. This represents the predicted tibial force under laboratory conditions. pseudo These are pseudo-label data for tibial strength. It is a predicted tibia force under real running conditions. These are the pseudo-label outputs after adding two different perturbations to the same outdoor sample. MSE is the mean squared error function.

[0083] Wherein, weight w t As shown below:

[0084]

[0085] In the formula, t peak α1 represents the peak time. α1 represents the peak weight.

[0086] Furthermore, the lower limb exercise load assessment indicators include the predicted peak vertical ground reaction force and the predicted peak tibial force.

[0087] A lower limb exercise load assessment system applying the above method includes: a pressure insole device, a plantar pressure acquisition module, a pressure leg sleeve device, a surface muscle force acquisition module, a surface electromyography device, a surface electromyography acquisition module, an inertial sensor device, an acceleration signal acquisition module, a training data acquisition module, a tibial force prediction model construction module, a vertical ground reaction force prediction model construction module, a vertical ground reaction force prediction module, a tibial force prediction module, and a lower limb exercise load assessment module.

[0088] The pressure insole device is used to detect the plantar pressure of the test subject.

[0089] The plantar pressure acquisition module is used to collect plantar pressure data from subjects when they run at different paces.

[0090] The pressure leg sleeve device is used to detect the surface muscle strength of the calf of the test subject.

[0091] The surface muscle strength acquisition module is used to acquire surface muscle strength signals of the calf muscles when the subject runs at different paces.

[0092] The surface electromyography device is used to detect electromyographic signals of the calf muscles of the subject.

[0093] The surface electromyography (EMG) acquisition module is used to acquire surface EMG signals of the tibialis anterior muscle when the subject runs at different paces.

[0094] The inertial sensor device is used to detect acceleration signals in various parts of the subject's body.

[0095] The acceleration signal acquisition module is used to acquire ankle joint acceleration signals and mid-tibial acceleration signals when the subject runs at different paces.

[0096] The training data acquisition module includes an optical motion capture system, a force measurement treadmill, a data alignment module, and a tibial force calculation module.

[0097] The optical motion capture system is used to collect three-dimensional trajectory information of subjects at different joint points when they run at different paces.

[0098] The force-measuring treadmill is used to collect the actual ground reaction force when the subject runs at different paces.

[0099] The data alignment module synchronizes three-dimensional trajectory information, actual ground reaction force, and wearable device signals such as plantar pressure, calf muscle surface force signals, tibialis anterior muscle surface electromyography signals, and acceleration signals through timestamps.

[0100] The tibial force calculation module calculates the subject's tibial force based on synchronized three-dimensional trajectory information and actual ground reaction force.

[0101] The tibial force prediction model construction module constructs a tibial force training sample set based on synchronous wearable device signals and tibial force, and uses the tibial force training sample set to train the model to obtain the tibial force prediction model.

[0102] The tibial force prediction module inputs real-time collected plantar pressure, calf muscle surface force signals, tibialis anterior muscle surface electromyography signals, and acceleration signals into the tibial force prediction model to obtain the predicted tibial force.

[0103] The vertical ground reaction force prediction model construction module constructs a vertical ground reaction force training sample set based on synchronous plantar pressure, ankle joint acceleration signals, and actual ground reaction force, and uses the vertical ground reaction force training sample set to train the model to obtain the vertical ground reaction force prediction model.

[0104] The vertical ground reaction force prediction module inputs the real-time collected plantar pressure and ankle joint acceleration signals into the vertical ground reaction force prediction model to obtain the predicted vertical ground reaction force.

[0105] The lower limb exercise load assessment module obtains lower limb exercise load assessment indicators based on predicted tibial force and predicted vertical ground reaction force.

[0106] The technical effectiveness of this invention is undeniable. This invention provides a method for assessing lower limb exercise load based on multimodal wearable devices. This method includes a simultaneous data acquisition experimental scheme using dual-source data (one source is laboratory equipment, and the other source is wearable devices containing pressure insoles, inertial sensors, surface electromyography sensors, and surface muscle force sensors) and a method for calculating tibial load during exercise. Using the calculated tibial force as the gold standard, this invention further develops a machine learning model for estimating lower limb load (vertical ground reaction force and tibial force) using wearable device signals. This invention combines musculoskeletal models and finite element models to propose a systematic method for accurately calculating tibial load using laboratory equipment. Furthermore, a model for estimating tibial force based on multimodal wearable data has been developed, which combines the advantages of local feature extraction and global context modeling in biomechanical prediction, resulting in better prediction performance than baseline models. In addition, pseudo-labels obtained by screening wearable device data samples from real-world scenarios further improve the model's generalization ability.

[0107] This invention provides a complete set of acquisition schemes, preprocessing methods, and tibial force calculation methods based on musculoskeletal models and finite element models for simultaneous acquisition of dual-source data, which are more systematic and scientific than other methods.

[0108] The tibial force prediction model based on multimodal wearable devices proposed in this invention can combine the advantages of local feature extraction and global context modeling in biomechanical prediction, and its overall performance is better than the baseline model, showing obvious advantages.

[0109] Peak force is often a key factor in fatigue accumulation and bone damage. This invention proposes a new weight-based loss function, Weight_MSELoss, which pays more attention to the peak of biological force during training compared to the traditional MSELoss loss function.

[0110] The semi-supervised optimization method based on pseudo-label generation and consistency regularization proposed in this invention can further improve the generalization ability of the model in real application scenarios. Attached Figure Description

[0111] Figure 1 This is an overall flowchart of the present invention;

[0112] Figure 2This is a schematic diagram illustrating the tibial force calculation method based on the musculoskeletal model and the finite element model.

[0113] Figure 3 A schematic diagram of a two-stage lower limb exercise load prediction model;

[0114] Figure 4 This is a schematic diagram of the laboratory TCNformer model in the first phase;

[0115] Figure 5 A schematic diagram of the weighted loss function Weight_MSELoss; Figure 5 (a) is a schematic diagram of the ground reaction force; Figure 5 (b) is a diagram illustrating the weights;

[0116] Figure 6 This is a flowchart for the second stage of model fine-tuning. Detailed Implementation

[0117] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0118] Example 1:

[0119] See Figures 1 to 6 A method for assessing lower limb exercise load based on a multimodal wearable device includes the following steps:

[0120] 1) Place cursor points at different joint points on the subjects, put on pressure insoles, pressure leg sleeves, inertial sensors, and surface electromyography devices, and have the subjects run at different paces on a treadmill.

[0121] 2) Use a pressure insole device to collect the plantar pressure of the subjects when they run at different paces.

[0122] The study used a pressure leg sleeve device to collect surface muscle force signals of the calf muscles when subjects ran at different paces.

[0123] Ankle acceleration signals and mid-tibial acceleration signals were collected from subjects running at different paces using inertial sensing devices.

[0124] The surface electromyography (EMG) device was used to collect surface EMG signals of the tibialis anterior muscle from the subjects while they were running at different paces.

[0125] An optical motion capture system was used to collect three-dimensional trajectory information of subjects at different joint points while they ran at different paces.

[0126] The actual ground reaction force of the subjects was collected using a treadmill at different paces.

[0127] 3) Synchronize three-dimensional trajectory information, actual ground reaction force, calf muscle surface force signal, ankle joint acceleration signal, mid-tibialis acceleration signal, tibialis anterior muscle surface electromyography signal and plantar pressure through timestamps.

[0128] 4) Calculate the tibial force of the subject based on the synchronized three-dimensional trajectory information and the actual ground reaction force.

[0129] 5) Based on the synchronous calf muscle surface force signal, ankle joint acceleration signal, tibial mid-shaft acceleration signal, tibialis anterior muscle surface electromyography signal, plantar pressure, and tibial force, a tibial force training sample set is constructed, and the model is trained using the tibial force training sample set to obtain a tibial force prediction model.

[0130] A training sample set for vertical ground reaction force is constructed based on synchronized ankle joint acceleration signals, plantar pressure, and actual ground reaction force. The model is then trained using this training sample set to obtain a vertical ground reaction force prediction model.

[0131] 6) Input the real-time collected surface muscle force signals of the calf muscles, ankle joint acceleration signals, mid-tibial acceleration signals, surface electromyography signals of the tibialis anterior muscle, and plantar pressure into the tibial force prediction model to obtain the predicted tibial force.

[0132] The real-time collected ankle joint acceleration signal and plantar pressure are input into the vertical ground reaction force prediction model to obtain the predicted vertical ground reaction force.

[0133] 7) Lower limb exercise load assessment indexes are obtained based on predicted tibial force and predicted vertical ground reaction force.

[0134] Example 2:

[0135] A method for assessing lower limb exercise load based on a multimodal wearable device, the main technical content of which is described in Example 1, further includes the following steps for calculating the subject's tibial force based on synchronized three-dimensional trajectory information and actual ground reaction force:

[0136] 4.1) Scale the standard musculoskeletal model to match the subject's size to obtain the subject's musculoskeletal model.

[0137] 4.2) Input the synchronized three-dimensional trajectory information into the subject's musculoskeletal model to obtain the displacement information of different joints when the subject runs at different paces.

[0138] The joints from which the three-dimensional trajectory information is collected include the hip joint, knee joint, and ankle joint.

[0139] The displacement information at different joint points includes the joint's position, velocity, and acceleration.

[0140] 4.3) Based on the displacement information at different joint points and the actual ground reaction force, the net joint torque at each moment is decomposed into individual muscle forces, and the muscle force of the peripheral tibial muscle group is calculated by minimizing the sum of squares of muscle activation.

[0141] The net joint torque at each moment is achieved through the static optimization function of the OpenSIM software.

[0142] The muscle strength of the peripheral tibial muscles was calculated directly using OpenSIM software.

[0143] 4.4) Construct a finite element model of the tibia.

[0144] The finite element model of the tibia was created using COMSOL software, as detailed below:

[0145] A standard adult male tibial model, manufactured by GOM in Germany, was imported into COMSOL software. This model was created using CT scans and calibrated; the solid model was provided as a SolidWorks file. Based on relevant literature, the tibial model was set with a bone mineral density of 1300 kg / m³, a Young's modulus of 7 × 10⁹ Pa, and a Poisson's ratio of 0.3. A simple fixed constraint was then applied to the distal tibia to completely restrict the surface near the tibia-talus interface in terms of translation and rotation, and the ground reaction force was equated to ankle joint forces applied to the contact surface between the ankle and tibia. The origin and insertion coordinates of the major lower limb muscles were obtained from human anatomy literature to determine the force application locations of different muscles in the model.

[0146] 4.5) Input the muscle force of the peripheral tibial muscle group and the actual ground reaction force into the tibial finite element model to calculate the Von Misesc stress of the tibia at different times.

[0147] The Qualisys 3D motion acquisition and analysis system generates C3D files, which need to be separated into TRC files containing trajectory information and MOT files containing ground reaction forces using Matlab (MathWorks, R2022a, US) software before further processing. In step 1, based on the subject's height and weight, and with minor adjustments to some bony marker positions, the original gait2392 model is scaled to obtain a model (Scaled model.osim) that matches the subject, ensuring that the root mean square error of all points does not exceed 0.02m and the maximum error does not exceed 0.04m. In step 2, the trajectory information of the marker points during motion is input into the scaled model, and Inverse Kinematics (IK) is used to solve for the displacement and rotation angles of the hip, knee, and ankle joints during motion (IKResult.mot). Step 3, Static Optimization (STO), decomposes the net joint torque at each moment into individual muscle forces based on the position, velocity, acceleration, and ground reaction force of each joint. Muscle forces are calculated by minimizing the sum of the squares of muscle activation. Next, a standard adult male tibia model, manufactured by GOM in Germany, was imported into COMSOL software. This model was created using CT scans and calibrated; the solid model was provided as a SolidWorks file. Referring to relevant literature, the tibia model was set with a bone density of 1300 kg / m³, a Young's modulus of 7 × 10⁹ Pa, and a Poisson's ratio of 0.3. A simple fixed constraint was then applied to the distal tibia to completely restrict the surface near the tibia-talus interface in terms of translation and rotation, and the ground reaction force was equated to ankle joint forces applied to the contact surface between the ankle and tibia. The origin and insertion coordinates of the major lower limb muscles were obtained from human anatomy literature to determine the force application locations of different muscles in the model. By applying muscle force and ground reaction force frame by frame to the tibia model in step 4, the Von Misesc stress of each steady-state tibia can be calculated, in N / m². It can be visually observed that the maximum body stress occurs in the distal third of the tibia, which is consistent with the fact that this is a common site for fatigue fractures.

[0148] Example 3:

[0149] A method for assessing lower limb exercise load based on a multimodal wearable device, the main technical contents of which are described in any one of Examples 1 to 2. Furthermore, the training process of the tibial force prediction model includes training in a laboratory environment and training in a real running environment.

[0150] The training process of the vertical ground reaction force prediction model includes training in a laboratory environment and training in a real running environment.

[0151] Example 4:

[0152] A method for assessing lower limb exercise load based on a multimodal wearable device, the main technical contents of which are described in any one of Examples 1 to 3. Further, the tibial force prediction model trained in the laboratory environment includes a preprocessing module I, a TCN module I, a Transformer encoder module I, and a prediction head I.

[0153] The preprocessing module I is used to preprocess the synchronous calf muscle surface force signal, ankle joint acceleration signal, tibia mid-shaft acceleration signal, tibialis anterior muscle surface electromyography signal, and plantar pressure to obtain preprocessed data I.

[0154] The preprocessing includes filtering, downsampling, and dividing the data into time windows.

[0155] The TCN module I is used to extract features from the preprocessed data I to obtain temporal data I, and to process the temporal data I using a one-dimensional convolutional network architecture to obtain embedded features I.

[0156] The TCN module employs a stacked convolutional architecture. The first three convolutional layers have dilation factors of 1, 2, and 3, respectively, and use 40 convolutional filters of size 5 to gradually expand the receptive field to extract long temporal dependencies. Next, a convolutional layer with a kernel size of (4,1) is used to compress the feature dimension, reducing the temporal dimension. After the convolutional operations, BatchNorm2d is used to normalize the output, ensuring that the distribution of outputs from each layer remains stable during training. Then, the ELU activation function is applied, followed by pooling layers and a Dropout layer with a dropout rate of 0.5. Finally, a projection layer rearranges the multidimensional feature maps into a flattened embedding vector. The TCN module acts as a feature extractor throughout the model, using dilated convolutions to gradually expand the receptive field, allowing the network to capture dependencies at different temporal scales at shallower layers without needing to over-deepen the network structure.

[0157] The Transformer encoder module I processes the embedded feature I by stacking several encoders to obtain the encoded feature I.

[0158] The Transformer Encoder module consists of multiple Transformer Encoder Blocks stacked together. Each encoder block includes a multi-head self-attention mechanism and a feedforward layer, and uses residual connections and Layer Normalization to enable the model to better capture complex long-term dependencies in time series.

[0159] The prediction head I aggregates and normalizes the encoded features I before feeding them into the linear layer to obtain the predicted tibial force.

[0160] The vertical ground reaction force prediction model trained in the laboratory environment includes a preprocessing module II, a TCN module II, a Transformer encoder module II, and a prediction head II.

[0161] The preprocessing module II is used to preprocess the synchronized ankle joint acceleration signal and plantar pressure to obtain preprocessed data II.

[0162] The preprocessing includes filtering, downsampling, and dividing the data into time windows.

[0163] The TCN module II is used to extract features from the preprocessed data II to obtain temporal data II, and to process the temporal data II using a one-dimensional convolutional network architecture to obtain embedded features II.

[0164] The Transformer encoder module II processes the embedded feature II by stacking several encoders to obtain the encoded feature II.

[0165] The prediction head II aggregates and normalizes the encoded features II before passing them into the linear layer to obtain the predicted vertical ground reaction force.

[0166] Example 5:

[0167] A method for assessing lower limb exercise load based on a multimodal wearable device, the main technical contents of which are described in any one of Embodiments 1 to 4, further comprising the following step of processing the embedded features by stacking several encoders to obtain the encoded features:

[0168] b1 uses LayerNorm to normalize the embedded features, resulting in normalized feature X.

[0169] b2 uses three fully connected layers to map the normalized feature X into three vectors of the same shape: query vector Q, key vector K, and value vector V.

[0170] The query vector Q, key vector K, and value vector V are shown below:

[0171] Q = XW Q K = XW K V = XW V (1)

[0172] In the formula, W Q W K W V Both are weight matrices.

[0173] b3 calculates the attention weights based on the query vector Q and the key vector K.

[0174] The attention weights are as follows:

[0175]

[0176] In the formula, Attention(Q,K) represents the attention weights, and softmax is the normalization function. d represents the feature dimension, and K... T This is the transpose of the key vector K.

[0177] b4 uses attention weights to perform a weighted summation of the value vector V, obtaining the output of each encoder.

[0178] The output of each encoder is shown below:

[0179] output (h) =Attention(Q,K)·V (3)

[0180] In the formula, h is the encoder number, and output is... (h) This represents the output of the h-th encoder. Attention(Q,K) represents the attention weights.

[0181] The encoding features are as follows:

[0182] MultiHead(Q,K,V)=Concat(output (1) ,...,output (num_heads) W O (4)

[0183] In the formula, MultiHead(Q,K,V) represents the encoding features. O This is a linear projection matrix. `num_heads` is the total number of encoders. `Concat` is the concatenation function.

[0184] b5 concatenates the outputs of each encoder together and projects them back to the original embedding dimension through a linear transformation to obtain the encoded features.

[0185] Example 6:

[0186] A method for assessing lower limb exercise load based on a multimodal wearable device, the main technical contents of which are described in any one of Examples 1 to 5. Furthermore, the training process of the tibial force prediction model in a real running environment is as follows:

[0187] c1 acquires pseudo-labeled data generated in real running environments and training data in laboratory environments.

[0188] c2 is a tibial force prediction model trained in a laboratory environment. It uses pseudo-labeled data generated in a real running environment and data trained in a laboratory environment to perform semi-supervised optimization on the model, resulting in a tibial force prediction model trained in a real running environment.

[0189] The training process for the vertical ground reaction force prediction model in a real running environment is as follows:

[0190] d1 acquires pseudo-labeled data generated in a real running environment and training data in a laboratory environment.

[0191] d2 is a vertical ground reaction force prediction model trained in a laboratory environment. The model is semi-supervised by using pseudo-label data generated in a real running environment and data trained in a laboratory environment to obtain a vertical ground reaction force prediction model trained in a real running environment.

[0192] Example 7:

[0193] A method for assessing lower limb exercise load based on a multimodal wearable device, the main technical contents of which are described in any one of Examples 1 to 6, further wherein the step of obtaining pseudo-label data generated in a real running environment is as follows:

[0194] e1 acquires real running data from different runners and then constructs different outdoor samples.

[0195] Because the wearable device signals collected during running are continuous, they are divided into time windows of the same size, and the data in each time window is a sample. The size of the time window can be customized.

[0196] e2 leverages the randomness of Monte Carlo Dropout technology to perform multiple forward propagations on outdoor samples, thereby constructing the pseudo-label prediction distribution for different outdoor samples.

[0197] e3 calculates the pseudo-label mean and prediction variance for each outdoor sample, as shown below:

[0198]

[0199] In the formula, i is the forward propagation index, and I is the total number of samples. The pseudo-label prediction result for the i-th forward propagation. σ is the pseudo-label mean. 2 e4 calculates the coefficient of variation for each outdoor sample, as shown below:

[0200]

[0201] In the formula, σ is the standard deviation of the pseudo-label. Here, is the pseudo-label mean, and CV is the coefficient of variation. e5 maps the coefficient of variation of each outdoor sample to a confidence score, as shown below:

[0202]

[0203] In the formula, conf represents the confidence score, and CV represents the confidence score. low For the minimum coefficient of variation, CV high is the maximum value of the coefficient of variation, and clip is the limiting function.

[0204] e6 selects outdoor samples with a confidence score greater than 0.8 as pseudo-label data generated in the real running environment.

[0205] Example 8:

[0206] A method for assessing lower limb exercise load based on a multimodal wearable device, the main technical contents of which are described in any one of Examples 1 to 7. Furthermore, the loss function used in the semi-supervised optimization is as follows:

[0207]

[0208] In the formula, This is the joint loss. α and λ are both hyperparameter weights.

[0209] Among them, the loss of laboratory data Loss of pseudo-label data Consistency loss As shown below:

[0210]

[0211] In the formula, t represents time. T represents the number of time intervals. y t This refers to the actual tibial force. This represents the predicted tibial force under laboratory conditions. pseudo These are pseudo-label data for tibial strength. It is a predicted tibia force under real running conditions. These are the pseudo-label outputs after adding two different perturbations to the same outdoor sample. MSE is the mean squared error function.

[0212] and All results are predicted by the model, but in the entire process of joint loss optimization, it is necessary to consider that the accuracy of the model trained in the laboratory environment should not decrease (loss from using laboratory data). To measure this, we also need to consider subsequent optimization using data from real running environments (using the loss from pseudo-labeled data). (Measured by the model), which is the result predicted by the model in both processes.

[0213] Wherein, weight w t As shown below:

[0214]

[0215] In the formula, t peak α1 represents the peak time. α1 represents the peak weight.

[0216] Example 9:

[0217] A method for assessing lower limb exercise load based on a multimodal wearable device, the main technical contents of which are described in any one of Examples 1 to 8. Further, the lower limb exercise load assessment index includes the predicted peak vertical ground reaction force and the predicted peak tibial force.

[0218] Example 10:

[0219] A lower limb exercise load estimation system using the method described in any one of embodiments 1-9 includes: a pressure insole device, a plantar pressure acquisition module, a pressure leg sleeve device, a surface muscle force acquisition module, a surface electromyography device, a surface electromyography acquisition module, an inertial sensor device, an acceleration signal acquisition module, a training data acquisition module, a tibial force prediction model construction module, a vertical ground reaction force prediction model construction module, a vertical ground reaction force prediction module, a tibial force prediction module, and a lower limb exercise load assessment module.

[0220] The pressure insole device is used to detect the plantar pressure of the test subject.

[0221] The plantar pressure acquisition module is used to collect plantar pressure data from subjects when they run at different paces.

[0222] The pressure leg sleeve device is used to detect the surface muscle strength of the calf of the test subject.

[0223] The surface muscle strength acquisition module is used to acquire surface muscle strength signals of the calf muscles when the subject runs at different paces.

[0224] The surface electromyography device is used to detect electromyographic signals of the calf muscles of the subject.

[0225] The surface electromyography (EMG) acquisition module is used to acquire surface EMG signals of the tibialis anterior muscle when the subject runs at different paces.

[0226] The inertial sensor device is used to detect acceleration signals in various parts of the subject's body.

[0227] The acceleration signal acquisition module is used to acquire ankle joint acceleration signals and mid-tibial acceleration signals when the subject runs at different paces.

[0228] The training data acquisition module includes an optical motion capture system, a force measurement treadmill, a data alignment module, and a tibial force calculation module.

[0229] The optical motion capture system is used to collect three-dimensional trajectory information of subjects at different joint points when they run at different paces.

[0230] The force-measuring treadmill is used to collect the actual ground reaction force when the subject runs at different paces.

[0231] The data alignment module synchronizes three-dimensional trajectory information, actual ground reaction force, and wearable device signals such as plantar pressure, calf muscle surface force signals, tibialis anterior muscle surface electromyography signals, and acceleration signals through timestamps.

[0232] The tibial force calculation module calculates the subject's tibial force based on synchronized three-dimensional trajectory information and actual ground reaction force.

[0233] The tibial force prediction model construction module constructs a tibial force training sample set based on synchronous wearable device signals and tibial force, and uses the tibial force training sample set to train the model to obtain the tibial force prediction model.

[0234] The tibial force prediction module inputs real-time collected plantar pressure, calf muscle surface force signals, tibialis anterior muscle surface electromyography signals, and acceleration signals into the tibial force prediction model to obtain the predicted tibial force.

[0235] The vertical ground reaction force prediction model construction module constructs a vertical ground reaction force training sample set based on synchronous plantar pressure, ankle joint acceleration signals, and actual ground reaction force, and uses the vertical ground reaction force training sample set to train the model to obtain the vertical ground reaction force prediction model.

[0236] The vertical ground reaction force prediction module inputs the real-time collected plantar pressure and ankle joint acceleration signals into the vertical ground reaction force prediction model to obtain the predicted vertical ground reaction force.

[0237] The lower limb exercise load assessment module obtains lower limb exercise load assessment indicators based on predicted tibial force and predicted vertical ground reaction force.

[0238] Example 11:

[0239] See Figures 1 to 6 A method and system for assessing lower limb exercise load based on a multimodal wearable device, the main technical contents of which include:

[0240] A method for assessing lower limb exercise load based on a multimodal wearable device, comprising:

[0241] 1) Place cursor points at different joint points on the subjects, put on pressure insoles, pressure leg sleeves, inertial sensing devices (IMU), and surface electromyography (sEMG) devices, and have the subjects run at different paces on a treadmill.

[0242] 2) Collect plantar pressure signals from subjects running at different paces using a pressure insole device;

[0243] The surface muscle strength signals of the calf muscle group were collected from the subjects when they ran at different paces using a pressure leg sleeve device.

[0244] Ankle acceleration signals and mid-tibial acceleration signals were collected from subjects running at different paces using inertial sensing devices.

[0245] Surface electromyography (EMG) devices were used to collect surface EMG signals of the tibialis anterior muscle from subjects running at different paces.

[0246] The optical motion capture system was used to collect three-dimensional trajectory information of different bony landmarks at different paces when the subjects ran at different speeds;

[0247] The actual three-dimensional ground reaction force of the subjects was collected using a force-measuring treadmill when they ran at different paces.

[0248] 3) Synchronize three-dimensional trajectory information, actual ground reaction force, surface electromyography, acceleration, surface muscle force, and plantar pressure through timestamps;

[0249] 4) Calculate the tibial force of the subject based on synchronized three-dimensional trajectory information and actual ground reaction force;

[0250] 5) Construct a tibial force training sample set based on synchronized surface electromyography, acceleration, surface muscle force, plantar pressure, and tibial force, and use the tibial force training sample set to train the model to obtain a tibial force prediction model.

[0251] A training sample set for vertical ground reaction force is constructed based on synchronized ankle joint acceleration, plantar pressure, and actual ground reaction force. The model is then trained using the training sample set to obtain a vertical ground reaction force prediction model.

[0252] 6) Input the real-time collected surface electromyography, acceleration, surface muscle force, and plantar pressure into the tibial force prediction model to obtain the predicted tibial force;

[0253] The real-time collected ankle joint acceleration and plantar pressure are input into the vertical ground reaction force prediction model to obtain the predicted vertical ground reaction force.

[0254] 7) Lower limb exercise load assessment indexes are obtained based on predicted tibial force and predicted vertical ground reaction force.

[0255] The steps for calculating the subject's tibial force based on synchronized three-dimensional trajectory information and actual ground reaction force are as follows:

[0256] a1 scales the standard musculoskeletal model to match the subject's musculoskeletal model, thus obtaining the subject's musculoskeletal model.

[0257] a2 inputs the synchronized three-dimensional trajectory information into the subject's musculoskeletal model to obtain the displacement information at different joint points when the subject runs at different paces.

[0258] The joints from which the three-dimensional trajectory information is collected include the hip joint, knee joint, and ankle joint;

[0259] The displacement information at different joint points includes the joint's position, velocity, and acceleration;

[0260] Based on displacement information at different joint points and actual vertical ground reaction force, a3 decomposes the net joint torque at each moment into individual muscle forces, and calculates the muscle force of the peripheral tibial muscle group by minimizing the sum of squares of muscle activation.

[0261] A4 is used to construct a finite element model of the tibia;

[0262] a5 inputs the muscle force of the peripheral tibial muscles and the actual vertical ground reaction force into the tibial finite element model to calculate the Von Misesc stress of the tibia at different times.

[0263] The lower limb load prediction model includes two stages of model training: the first stage is laboratory-scale model training, and the second stage is generalization to training in a real running environment.

[0264] The first stage of the tibial force prediction model involves training the TCNformer deep learning model using surface electromyography, acceleration, surface muscle force, and plantar pressure signals collected in a laboratory setting, thereby constructing a tibial force prediction model under laboratory conditions.

[0265] The second stage of the tibial force prediction model is based on the load estimation model trained in the laboratory scenario. Labeled data from the laboratory scenario and pseudo-labeled data generated in the real running scenario are input to perform semi-supervised optimization of the model, resulting in a tibial force prediction model with better generalization ability.

[0266] The first stage of the vertical ground reaction force prediction model is to train the TCNformer deep learning model by collecting ankle joint acceleration and plantar pressure signals in a laboratory setting, and then construct the vertical ground reaction force prediction model in a laboratory environment.

[0267] The second stage of the vertical ground reaction force prediction model is based on the vertical ground reaction force prediction model trained in the laboratory scenario. Labeled data from the laboratory scenario and pseudo-labeled data generated in the real running scenario are input to perform semi-supervised optimization of the model, so as to obtain a vertical ground reaction force prediction model with better generalization ability.

[0268] The tibial force prediction model under laboratory conditions includes a preprocessing module I, a TCN module I, a Transformer encoder module I, and a prediction head I.

[0269] The preprocessing module I is used to preprocess synchronized surface electromyography, acceleration, surface muscle force, and plantar pressure signals.

[0270] The TCN module I is used to extract features from the preprocessed data to obtain temporal data, and to process the temporal data using a one-dimensional convolutional network architecture to obtain embedded features;

[0271] The Transformer encoder module I processes the embedded features by stacking several encoders to obtain encoded features;

[0272] The prediction head I aggregates and normalizes the encoded features before feeding them into the linear layer to obtain the predicted tibial force.

[0273] The vertical ground reaction force prediction model in the laboratory environment includes a preprocessing module II, a TCN module II, a Transformer encoder module II, and a prediction head II.

[0274] The preprocessing module II is used to preprocess the synchronized ankle joint acceleration and plantar pressure signals;

[0275] The TCN module II is used to extract features from the preprocessed data to obtain temporal data, and to process the temporal data using a one-dimensional convolutional network architecture to obtain embedded features;

[0276] The Transformer encoder module II processes the embedded features by stacking several encoders to obtain coded features;

[0277] The prediction head II aggregates and normalizes the encoded features before passing them into the linear layer to obtain the predicted vertical ground reaction force.

[0278] The preprocessing includes filtering, downsampling, and dividing the data into time windows.

[0279] The steps for processing the embedded features by stacking several encoders to obtain the encoded features are as follows:

[0280] b1 uses LayerNorm to normalize the embedded features, obtaining the normalized feature X;

[0281] b2 uses three fully connected layers to map the normalized feature X into three vectors of the same shape: query vector Q, key vector K, and value vector V.

[0282] b3 calculates the attention weights based on the query vector Q and the key vector K;

[0283] b4 uses attention weights to perform a weighted summation of the value vector V to obtain the output of each encoder;

[0284] b5 concatenates the outputs of each encoder together and projects them back to the original embedding dimension through a linear transformation to obtain the encoded features.

[0285] The query vector Q, key vector K, and value vector V are shown below:

[0286] Q = XW Q K = XW K V = XW V (1)

[0287] In the formula, W Q W K W V Both are weight matrices.

[0288] The attention weights are as follows:

[0289]

[0290] In the formula, Attention(Q,K) represents the attention weights, softmax is the normalization function, d is the feature dimension, and K is the number of features. T This is the transpose of the key vector K.

[0291] The output of each encoder is shown below:

[0292] output (h) =Attention(Q,K)·V(3)

[0293] In the formula, h is the encoder number, and output is... (h) Let Q be the output of the h-th encoder; Attention(Q,K) represents the attention weights.

[0294] The encoding features are as follows:

[0295] MultiHead(Q,K,V)=Concat(output (1) ,...,output(num_heads) W O (4)

[0296] In the formula, MultiHead(Q,K,V) represents the encoding features; W O is the linear projection matrix; num_heads is the total number of encoders; Concat is the concatenation function.

[0297] The pseudo-label data generated in the real running scenario is produced through the following steps:

[0298] c1 introduces the Monte Carlo Dropout technique to retain Dropout activation during the testing phase and uses its randomness to perform multiple forward propagations on the same sample input, thereby constructing the predicted distribution of the output;

[0299] c2 calculates the pseudo-label mean and prediction variance for each sample;

[0300] c3 calculates the coefficient of variation for each sample;

[0301] c3 maps the coefficient of variation to a confidence score;

[0302] c4 filters out pseudo-label data with high confidence based on confidence level.

[0303] The pseudo-label mean and prediction variance are shown below:

[0304]

[0305] In the formula, The result of the i-th forward propagation is given, where T is the total number of samples. σ is the pseudo-label mean. 2 To predict variance;

[0306] The coefficient of variation is shown below:

[0307]

[0308] In the formula, σ is the standard deviation of the pseudo-label. is the pseudo-label mean, and CV is the coefficient of variation.

[0309] The confidence scores are shown below:

[0310]

[0311] In the formula, conf represents the confidence score, and CV represents the confidence score. low The low-variance threshold is set to 5%. When the predicted coefficient of variation is below this value, the model is considered highly certain about the prediction, and the confidence level is considered to be 1.0. (CV)high The high variance threshold is set at 15%. When the predicted coefficient of variation is higher than this value, the model is considered to be highly uncertain about the predicted value, and the confidence level is considered to be 0.0. That is, the confidence level varies linearly between 0.0 and 1.0.

[0312] The high-confidence pseudo-label data refers to sample data with a confidence level greater than 0.8.

[0313] The semi-supervised optimization process comprehensively considers supervised and unsupervised information, and the joint loss function used includes consistency loss, laboratory data loss, and pseudo-label loss.

[0314] The joint loss function is as follows:

[0315]

[0316] In the formula, It is a joint loss. It is a loss of laboratory data. It's the loss of pseudo-labels. It is the consistency loss, where α is the hyperparameter weight set to 0.5, and λ is the hyperparameter weight set to 0.5.

[0317] The consistency loss is as follows:

[0318]

[0319] In the formula, To show the output corresponding to adding two slight perturbations to the same sample. This results in a loss of consistency.

[0320] The loss of the laboratory data was calculated using the weighted loss function Weight_MSELoss, as shown below:

[0321]

[0322] In the formula, t is time; T is the number of time intervals; y t Actual tibial force; For predicted tibial force;

[0323] Wherein, weight w t As shown below:

[0324]

[0325] In the formula, α is the weight of the peak point; t peak This is the peak time.

[0326] The loss of the pseudo-label is as follows:

[0327]

[0328] In the formula, y pseudo It's a false label for tibial strength. It is the predicted tibial strength. It is the loss of pseudo-labels.

[0329] The lower limb exercise load assessment indicators include the predicted peak vertical ground reaction force and the predicted peak tibial force.

[0330] A lower limb exercise load assessment system applying the above method includes: a pressure insole device, a plantar pressure acquisition module, a pressure leg sleeve device, a surface muscle force acquisition module, a surface electromyography device, a surface electromyography acquisition module, an inertial sensor device, an acceleration signal acquisition module, a training data acquisition module, a tibial force prediction model construction module, a vertical ground reaction force prediction model construction module, a vertical ground reaction force prediction module, a tibial force prediction module, and a lower limb exercise load assessment module;

[0331] The pressure insole device is used to detect the plantar pressure of the test subject;

[0332] The plantar pressure acquisition module is used to collect plantar pressure data of subjects when they run at different paces.

[0333] The pressure leg sleeve device is used to detect the surface muscle strength of the calf of the test subject;

[0334] The surface muscle strength acquisition module is used to acquire surface muscle strength signals of the calf muscles when the subject runs at different paces.

[0335] The surface electromyography device is used to detect electromyographic signals of the calf muscles of the subject.

[0336] The surface electromyography acquisition module is used to acquire surface electromyography signals of the tibialis anterior muscle when the subject runs at different paces.

[0337] The inertial sensor device is used to detect acceleration signals of various parts of the subject's body;

[0338] The acceleration signal acquisition module is used to acquire ankle joint acceleration signals and mid-tibial acceleration signals when the subject runs at different paces.

[0339] The training data acquisition module includes an optical motion capture system, a force measurement treadmill, a data alignment module, and a tibial force calculation module.

[0340] The optical motion capture system is used to collect three-dimensional trajectory information of subjects at different joint points when running at different paces;

[0341] The force-measuring treadmill is used to collect the actual ground reaction force when the subject runs at different paces;

[0342] The data alignment module synchronizes three-dimensional trajectory information, actual ground reaction force, and plantar pressure, surface muscle force, surface electromyography, and acceleration signals from wearable devices via timestamps.

[0343] The tibial force calculation module calculates the subject's tibial force based on synchronized three-dimensional trajectory information and actual ground reaction force.

[0344] The tibial force prediction model building module constructs a tibial force training sample set based on synchronous wearable device signals and tibial force, and uses the tibial force training sample set to train the model to obtain the tibial force prediction model.

[0345] The tibial force prediction module inputs real-time collected plantar pressure, surface muscle force, surface electromyography, and acceleration into the tibial force prediction model to obtain the predicted tibial force.

[0346] The vertical ground reaction force prediction model construction module constructs a vertical ground reaction force training sample set based on synchronous plantar pressure, ankle joint acceleration signals, and actual vertical ground reaction force, and uses the vertical ground reaction force training sample set to train the model to obtain the vertical ground reaction force prediction model.

[0347] The vertical ground reaction force prediction module inputs the real-time collected plantar pressure and ankle joint acceleration signals into the vertical ground reaction force prediction model to obtain the predicted vertical ground reaction force.

[0348] The lower limb exercise load assessment module obtains lower limb exercise load assessment indicators based on predicted tibial force and predicted vertical ground reaction force.

[0349] Example 12:

[0350] See Figures 1 to 6 A method and system for assessing lower limb exercise load based on a multimodal wearable device, the main technical contents of which include:

[0351] Before the experiment begins, the optical motion capture system will be calibrated to determine the spatial coordinate system of the camera system's field of view, and the calibration will be performed strictly according to the manufacturer's calibration procedure. Detailed experimental steps are as follows:

[0352] (1) The subjects filled out the "Experimental Personnel Information Form", read the "Instructions for Data Collection Personnel", and did some simple warm-up exercises.

[0353] (2) Testing the connectivity of laboratory equipment: Referring to the gait2392 model point map, paste all the cursor points on the test subjects. Stand on the force-measuring platform and test whether all the cursor points can be captured by the camera and whether the force-measuring platform can continuously and accurately display the ground reaction force. Save a signal and test whether it can be stored normally and exported in the specified format.

[0354] (3) Wearable device connectivity: Subjects were equipped with pressure insoles, pressure leg sleeves, an inertial measurement unit (IMU), and a surface electromyography (sEMG) device. The pressure insoles were worn on the soles of both feet, the pressure leg sleeves on the calf muscles, the inertial sensor was placed on the lateral ankle and mid-tibia, and the sEMG device was placed on the tibialis anterior muscle. Subjects performed simple movements, and all devices were observed to exhibit significant signal fluctuations. Tests were also conducted to ensure a complete and continuous signal could be displayed on a computer, and data was saved to ensure proper storage.

[0355] (4) Baseline test: The subject stands still on the treadmill with his arms outstretched, and data is collected for 30 seconds.

[0356] (5) Signal acquisition: The subjects stood on the treadmill and ran for 10 minutes at a pace of 3 m / s (about 11 km / h), while the laboratory equipment and wearable sensor signals were collected simultaneously.

[0357] (6) Remove the reflective markers and wearable devices from the subjects, and save and organize the experimental data.

[0358] After the experimental data collection is complete, we will synchronize the signals from the laboratory equipment and the wearable device using timestamps. First, we will downsample the high-sampling-frequency laboratory equipment to the same frequency as the low-sampling-frequency wearable device, and then align the data according to the sampling time of each sample. Further calculation of tibial force using the laboratory equipment will require a musculoskeletal model and finite element analysis. (See attached...) Figure 2 As shown, the first three steps of building the musculoskeletal model are implemented using Opensim v.4.4 software, a free musculoskeletal modeling and simulation application and library. The input to the musculoskeletal model is the 3D trajectory information of the whole-body marker points and the ground reaction force; the output is the muscle force of the peripheral tibial muscle group. The final step uses COMSOL... (COMSOL This software (v.6.0.cn.comsol.com.COMSOL AB, Stockholm, Sweden) is used for the construction and analysis of a finite element model of the tibia. Muscle forces and ground reaction forces are input into the finite element model to calculate the Von Misesc stress of the tibia.

[0359] The following are the detailed steps of the musculoskeletal model and finite element analysis: The Qualisys 3D motion acquisition and analysis system obtains a c3d file, which needs to be separated into a trc file containing trajectory information and a mot file containing ground reaction force using Matlab (MathWorks, R2022a, US) software before further operations. In step (1), based on the subject's height and weight, and after fine-tuning the positions of some bony markers, the original gait2392 model is scaled to obtain a model Scaled model.osim that matches the subject, so that the root mean square error of all points does not exceed 0.02m and the maximum error does not exceed 0.04m. The trajectory information of the marker points during the motion is input into the scaled model in step (2) and Inverse Kinematics (IK) is performed to obtain the displacement and rotation angles of the hip, knee, and ankle joints during the motion, IKResult.mot. Step (3) Static Optimization (STO) decomposes the net joint torque at each moment into individual muscle forces based on the position, velocity, acceleration, and ground reaction force of each joint. The muscle force is calculated by minimizing the sum of the squares of muscle activation. Next, a standard adult male tibia model was imported into COMSOL software. This model was manufactured by GOM in Germany, modeled using CT scans, and calibrated. The solid model was provided as a SolidWorks file. Based on relevant literature, the bone density of the tibia model was set to 1300 kg / m3, Young's modulus to 7*10^9 Pa, and Poisson's ratio to 0.3. A simple fixed constraint was then used at the distal end of the tibia to completely restrict the surface near the tibia-talar interface in terms of translation and rotation, and the ground reaction force was equivalent to the ankle joint force applied to the contact surface between the ankle joint and the tibia. The origin and insertion coordinates of the major muscles of the lower limb were obtained by consulting human anatomy literature to determine the force application positions of different muscles in the model. By applying muscle force and ground reaction force frame by frame to the tibia model in step (4), the Von Misesc stress of each steady-state tibia can be calculated, with units of N / m2. It can be intuitively observed that the maximum body stress occurs at the distal third of the tibia, which is consistent with the fact that this is a common site for fatigue fractures.

[0360] The above steps were used to calculate the tibial force during exercise using laboratory equipment, while the vertical ground reaction force can be directly monitored using a force-measuring treadmill. The next goal is to use the pressure insole signals collected simultaneously to predict these two bioforces.

[0361] As attached Figure 3 As shown, we constructed a two-stage model. The two stages respectively achieve the estimation of biomechanical forces using signals from multimodal wearable devices and optimize the model to improve its generalization ability in real-world scenarios. In the first stage, two biomechanical forces—tibial force and vertical ground reaction force—are estimated using multimodal wearable device signals collected in the laboratory. In the second stage, pseudo-labels are generated using signals from real-world running scenarios, and the model is semi-supervised and fine-tuned by combining this with data from the laboratory scenario.

[0362] As attached Figure 4 The first-stage model, named TCNformer, is a fusion of TCN and Transformer encoders used to extract multi-scale features from time series data and perform complex temporal dependency modeling. It ultimately achieves regression prediction through a prediction head. The first module preprocesses the collected data, including filtering, downsampling, and dividing the data into time windows. The processed data is then shuffled and fed into the model in batches using a DataLoader for training. The second module is the TCN module, which extracts features from the input time series data using a one-dimensional convolutional network architecture. The third module is the Transformer Encoder module, the core encoder of the model. It further processes the embedded features output by the TCN by stacking multiple encoder blocks, thus forming a rich model of the time series. The last module is the model's prediction head, which aggregates and normalizes the encoded features before feeding them into a linear layer, outputting the predicted vertical ground reaction force or tibial force.

[0363] The model is described in detail below:

[0364] The TCN module employs a stacked convolutional architecture. The first three convolutional layers have dilation factors of 1, 2, and 3, respectively, and use 40 convolutional filters of size 5 to gradually expand the receptive field to extract long temporal dependencies. Next, a convolutional layer with a kernel size of (4,1) is used to compress the feature dimension, reducing the temporal dimension. After the convolutional operations, BatchNorm2d is used to normalize the output, ensuring that the distribution of outputs from each layer remains stable during training. Then, the ELU activation function is applied, followed by pooling layers and a Dropout layer with a dropout rate of 0.5. Finally, a projection layer rearranges the multidimensional feature maps into a flattened embedding vector. The TCN module acts as a feature extractor throughout the model, using dilated convolutions to gradually expand the receptive field, allowing the network to capture dependencies at different temporal scales at shallower layers without needing to over-deepen the network structure.

[0365] The Transformer Encoder module consists of multiple stacked Transformer Encoder Blocks. Each encoder block includes a multi-head self-attention mechanism and a feedforward layer, and uses residual connections and Layer Normalization to enable the model to better capture complex long-term dependencies in time series. Each encoder block processes the embedded features independently. First, the input features are normalized using LayerNorm and denoted as X. Then, we set the embsize to 40. Then, three fully connected layers are used to map the normalized features into three vectors of the same shape: query(Q), key(K), and value(V), as shown in Equation (1).

[0366] Q = XW Q K = XW K V = XW V (1)

[0367] Among them, W Q W K W V ∈R em_size*em_size This is a learnable weight matrix. We compute the dot product between each query and all keys to obtain an attention score, representing the similarity between each position in the query sequence and each position in the key sequence. To avoid excessively large dot product values ​​leading to numerical instability, we scale the results by dividing by the square root of the feature dimension. The attention weights are then obtained by softmax normalization, as shown in formula (2). During the calculation, multiple attention heads calculate the weights in parallel, and each head focuses on different feature information.

[0368]

[0369] Then, the values ​​are weighted and summed using attention weights to obtain the weighted representation of each query position. Each value in the attention weight matrix represents the degree of attention that query position pays to each key position. The elements of the value matrix are weighted according to these weights to form a new output, as shown in formula (3).

[0370] output (h) =Attention(Q,K)·V (3)

[0371] Here, `output` is the weighted sum, representing the output of each attention head after processing the input; `h` is the number of the attention head, and we set up a total of 8 attention heads. Finally, the outputs of all attention heads are concatenated together and transformed by a linear transformation W. o Projecting back to the original embedding dimension size results in the final output of the multi-head attention, as shown in Equation (4).

[0372] MultiHead(Q,K,V)=Concat(output (1) ,…,output (num_heads) W O (4)

[0373] Among them W o It is a linear projection matrix, and the concatenated dimensions are (batch_size, query_len, emb_size), which is consistent with the embedding dimension of the input. batch_size represents the size of the input batch, and query_len is the length of the query sequence.

[0374] Furthermore, during running, more attention is often paid to the peak values ​​of vertical ground reaction force and tibial force. These peak values ​​represent instantaneous high loads and are key factors contributing to fatigue accumulation and bone damage. When training the multi-channel time series prediction model, we did not use the traditional MSE loss function, but instead proposed a weighted loss function, Weight_MSELoss. By detecting the peak values ​​of vertical ground reaction force and tibial force and assigning higher weights to the peak positions, the model's focus on these key peak values ​​is enhanced, improving the prediction accuracy of peak value changes. As shown in Equation (5), y t It is a real vertical ground reaction force. These are the model predictions, and the weight vector is w = [w1, w2, ..., w...]. T This includes weight adjustments for each time step t.

[0375]

[0376] After trying different width conditions, the maximum value of the vertical ground reaction force with a minimum width greater than 12 was identified as the peak value, and the maximum value of the tibial force with a minimum width greater than 8 was identified as the peak value. The base weight of all data was set to 1, and the weight of all data with time steps within 3 to the left and right of the peak points was increased, as shown in formula (6).

[0377]

[0378] As attached Figure 5 The weighting is intuitively represented. The principle of the weighted loss function Weight_MSELoss is to increase the weight of the peak value in the loss function, while the weight of other regions is set to the default value of 1.

[0379] As attached Figure 6 The diagram shows an optimization method for further fine-tuning the model, called a semi-supervised optimization method based on pseudo-label generation and consistency regularization.

[0380] A semi-supervised fine-tuning method based on pseudo-label generation and consistency regularization aims to enhance the model's adaptability to the target domain using unlabeled real running data and further fine-tune a previously trained model for estimating tibial force using wearable device data. The core of the method consists of two stages: first, generating high-confidence pseudo-label samples through Monte Carlo Dropout; and second, performing two-stage fine-tuning training using both laboratory data and pseudo-label samples.

[0381] In the pseudo-label generation stage, Monte Carlo Dropout is introduced to model the uncertainty of the prediction results. In traditional testing, the Dropout layer is usually turned off, while the Monte Carlo Dropout method retains Dropout activation during the testing phase, utilizing its randomness to perform multiple forward propagations on the same input, thereby constructing the predicted distribution of the output. For each outdoor sample, the prediction result of the i-th forward propagation is denoted as... A total of T samples were taken, and the mean of the pseudo-labels was... and prediction variance σ 2 Defined as formulas respectively:

[0382]

[0383] Furthermore, the ratio of the standard deviation to the mean of the pseudo-labels is used as the coefficient of variation (CV):

[0384]

[0385] To filter out high-quality pseudo-labels, CVs are mapped to confidence scores:

[0386]

[0387] Among them, CV low is the low mutation threshold, which is set to 5%. When the predicted coefficient of variation is lower than this value, it is considered that the model is very certain about this prediction, and the confidence level is considered to be 1.0; CV high is the high mutation threshold, which is set to 15%. When the predicted coefficient of variation is higher than this value, it is considered that the model is highly uncertain about this predicted value, and the confidence level is considered to be 0.0. When 5% < CV < 15%, the confidence level changes linearly between 0.0 and 1.0. And the confidence threshold θ = 0.5 is set, and only the samples with conf > θ and their corresponding pseudo-labels are retained as unsupervised training data.

[0388] In the model fine-tuning stage, a two-stage strategy is adopted to alleviate the catastrophic forgetting phenomenon, with a total of 200 training cycles. In the first stage (about 67 rounds), the Transformer encoder and the regression prediction head are frozen, and only the shallow feature extraction network is trained to adjust the response ability to the low-level signals in the new environment; in the second stage (about 133 rounds), all parameters are unfrozen, and the learning rate is reduced to one-fifth of the original, further adapting to the target domain distribution. The model uses the Adam optimizer, with the initial learning rate set to 1e-5 and reduced to 2e-6 when entering the second stage. This progressive learning rate adjustment strategy helps prevent overfitting and achieve fine parameter adjustment. In addition, to enhance the model's ability to model the structure of unlabeled data, a consistency regularization term is introduced. Two slight perturbations are added to the same sample Its corresponding output is The consistency loss is given by the formula:

[0389]

[0390] Taking into account both supervised and unsupervised information, the total loss function is defined as in the formula.

[0391]

[0392] Among them, is the Weight-MSELoss of the laboratory data, is the MSELoss of the pseudo-label, and α = 0.5, λ = 0.5 are hyperparameter weights.

[0393] The entire semi-supervised fine-tuning method based on pseudo-label generation and consistency regularization can effectively alleviate the impact of data domain inconsistency on the model performance and significantly enhance the model's application ability in real scenarios.

Claims

1. A method for assessing lower limb exercise load based on a multimodal wearable device, characterized in that, Includes the following steps: 1) Place cursor points at different joint points on the subjects, put on pressure insoles, pressure leg sleeves, inertial sensors, and surface electromyography devices, and have the subjects run at different paces on a treadmill. 2) Collect plantar pressure data from subjects running at different paces using a pressure insole device; The surface muscle strength signals of the calf muscle group were collected from the subjects when they ran at different paces using a pressure leg sleeve device. Ankle acceleration signals and mid-tibial acceleration signals were collected from subjects running at different paces using inertial sensing devices. Surface electromyography (EMG) devices were used to collect surface EMG signals of the tibialis anterior muscle from subjects running at different paces. The optical motion capture system was used to collect three-dimensional trajectory information of subjects at different joint points while running at different paces; The actual ground reaction force of the subjects was collected using a force-measuring treadmill when they ran at different paces. 3) Synchronize three-dimensional trajectory information, actual ground reaction force, calf muscle surface force signal, ankle joint acceleration signal, mid-tibialis acceleration signal, tibialis anterior muscle surface electromyography signal and plantar pressure through timestamps; 4) Calculate the tibial force of the subject based on synchronized three-dimensional trajectory information and actual ground reaction force; 5) Based on the synchronous calf muscle surface force signal, ankle joint acceleration signal, tibial midshaft acceleration signal, tibialis anterior muscle surface electromyography signal, plantar pressure, and tibial force, a tibial force training sample set is constructed, and the model is trained using the tibial force training sample set to obtain a tibial force prediction model. A training sample set for vertical ground reaction force is constructed based on synchronized ankle joint acceleration signals, plantar pressure, and actual ground reaction force. The model is then trained using the training sample set to obtain a vertical ground reaction force prediction model. 6) Input the real-time collected calf muscle surface force signal, ankle joint acceleration signal, mid-tibialis acceleration signal, tibialis anterior muscle surface electromyography signal, and plantar pressure into the tibial force prediction model to obtain the predicted tibial force; The real-time collected ankle joint acceleration signal and plantar pressure are input into the vertical ground reaction force prediction model to obtain the predicted vertical ground reaction force. 7) Lower limb exercise load assessment indexes are obtained based on predicted tibial force and predicted vertical ground reaction force.

2. The method for assessing lower limb exercise load based on a multimodal wearable device according to claim 1, characterized in that, The steps for calculating the subject's tibial force based on synchronized three-dimensional trajectory information and actual ground reaction force are as follows: 4.1) Scale the standard musculoskeletal model to match the subject's size to obtain the subject's musculoskeletal model. 4.2) Input the synchronized three-dimensional trajectory information into the musculoskeletal model of the subject to obtain the displacement information of different joints when the subject runs at different paces; The joints from which the three-dimensional trajectory information is collected include the hip joint, knee joint, and ankle joint; The displacement information at the different joint points includes the joint's position, velocity, and acceleration; 4.3) Based on the displacement information at different joint points and the actual ground reaction force, the net joint torque at each moment is decomposed into individual muscle forces, and the muscle force of the peripheral tibial muscle group is calculated by minimizing the sum of squares of muscle activation. 4.4) Construct a finite element model of the tibia; 4.5) Input the muscle force of the peripheral tibial muscle group and the actual ground reaction force into the tibial finite element model to calculate the Von Misesc stress of the tibia at different times.

3. The method for assessing lower limb exercise load based on a multimodal wearable device according to claim 1, characterized in that, The training process of the tibial force prediction model includes training in a laboratory environment and training in a real running environment. The training process of the vertical ground reaction force prediction model includes training in a laboratory environment and training in a real running environment.

4. The method for assessing lower limb exercise load based on a multimodal wearable device according to claim 3, characterized in that, The tibial force prediction model trained in the laboratory environment includes a preprocessing module I, a TCN module I, a Transformer encoder module I, and a prediction head I. The preprocessing module I is used to preprocess the synchronous calf muscle surface force signal, ankle joint acceleration signal, mid-tibialis acceleration signal, tibialis anterior muscle surface electromyography signal, and plantar pressure to obtain preprocessed data I; The preprocessing includes filtering, downsampling, and dividing the data into time windows; The TCN module I is used to extract features from the preprocessed data I to obtain temporal data I, and to process the temporal data I using a one-dimensional convolutional network architecture to obtain embedded features I. The Transformer encoder module I processes the embedded feature I by stacking several encoders to obtain the encoded feature I; The prediction head I aggregates and normalizes the encoded feature I and then feeds it into the linear layer to obtain the predicted tibial force. The vertical ground reaction force prediction model trained in the laboratory environment includes a preprocessing module II, a TCN module II, a Transformer encoder module II, and a prediction head II. The preprocessing module II is used to preprocess the synchronized ankle joint acceleration signal and plantar pressure to obtain preprocessed data II; The preprocessing includes filtering, downsampling, and dividing the data into time windows; The TCN module II is used to extract features from the preprocessed data II to obtain temporal data II, and to process the temporal data II using a one-dimensional convolutional network architecture to obtain embedded features II; The Transformer encoder module II processes the embedded feature II by stacking several encoders to obtain the encoded feature II; The prediction head II aggregates and normalizes the encoded features II before passing them into the linear layer to obtain the predicted vertical ground reaction force.

5. The method for assessing lower limb exercise load based on a multimodal wearable device according to claim 4, characterized in that, The steps for processing the embedded features by stacking several encoders to obtain the encoded features are as follows: b1 uses LayerNorm to normalize the embedded features, obtaining the normalized feature X; b2 uses three fully connected layers to map the normalized feature X into three vectors of the same shape: query vector Q, key vector K, and value vector V. The query vector Q, key vector K, and value vector V are shown below: Q=XW Q ,K=XW K ,V=XW V (1) In the formula, W Q W K W V Both are weight matrices; b3 calculates the attention weights based on the query vector Q and the key vector K; The attention weights are as follows: In the formula, Attention(Q,K) represents the attention weights, softmax is the normalization function, and d represents the feature dimension. This is the transpose of the key vector K; b4 uses attention weights to perform a weighted summation of the value vector V to obtain the output of each encoder; The output of each encoder is shown below: output (h) =Attention(Q,K)·V (3) In the formula, h is the encoder number, and output is... (h) Let Q be the output of the h-th encoder; Attention(Q,K) represents the attention weights. The encoding features are as follows: MultiHead(Q,K,V)=Concat(output (1) ,...,output (num_heads) )W O (4) In the formula, MultiHead(Q,K,V) represents the encoding features; W O This is the linear projection matrix; num_heads is the total number of encoders; Concat is the concatenation function; b5 concatenates the outputs of each encoder together and projects them back to the original embedding dimension through a linear transformation to obtain the encoded features.

6. The method for assessing lower limb exercise load based on a multimodal wearable device according to claim 3, characterized in that, The training process for the tibial force prediction model in a real running environment is as follows: c1 acquires pseudo-labeled data generated in real running environments and training data in laboratory environments; c2 is a tibial force prediction model trained in a laboratory environment. It uses pseudo-labeled data generated in a real running environment and data trained in a laboratory environment to perform semi-supervised optimization of the model, resulting in a tibial force prediction model trained in a real running environment. The training process for the vertical ground reaction force prediction model in a real running environment is as follows: d1 acquires pseudo-labeled data generated in a real running environment and training data in a laboratory environment; d2 is a vertical ground reaction force prediction model trained in a laboratory environment. The model is semi-supervised by using pseudo-label data generated in a real running environment and data trained in a laboratory environment to obtain a vertical ground reaction force prediction model trained in a real running environment.

7. The method for assessing lower limb exercise load based on a multimodal wearable device according to claim 6, characterized in that, The steps for obtaining pseudo-label data generated in a real running environment are as follows: e1 acquires real running data from different runners and then constructs different outdoor samples; e2 leverages the randomness of Monte Carlo Dropout technology to perform multiple forward propagations on outdoor samples, thereby constructing the pseudo-label prediction distribution for different outdoor samples. e3 calculates the pseudo-label mean and prediction variance for each outdoor sample, as shown below: In the formula, i is the forward propagation index, and I is the total number of samples. The pseudo-label prediction result for the i-th forward propagation. σ is the pseudo-label mean. 2 The prediction variance for pseudo-labels; e4 calculates the coefficient of variation for each outdoor sample, as shown below: In the formula, σ is the standard deviation of the pseudo-label, y is the mean of the pseudo-label, and CV is the coefficient of variation; e5 maps the coefficient of variation of each outdoor sample to a confidence score, as shown below: In the formula, conf represents the confidence score, and CV represents the confidence score. low For the minimum coefficient of variation, CV high The maximum value of the coefficient of variation is denoted by `clip`, which is the limiting function. e6 selects outdoor samples with confidence scores greater than a preset threshold as pseudo-label data generated in a real running environment.

8. The method for assessing lower limb exercise load based on a multimodal wearable device according to claim 6, characterized in that, The loss function used in the semi-supervised optimization is as follows: In the formula, This is the joint loss; α and λ are both hyperparameter weights. Among them, the loss of laboratory data Loss of pseudo-label data Consistency loss As shown below: In the formula, t is time; T is the number of time intervals; y t Actual tibial force; y is the predicted tibial force under laboratory conditions; pseudo These are pseudo-label data for tibial strength. It is a predicted tibia force under real running conditions; These are the pseudo-label outputs after adding two different perturbations to the same outdoor sample; MSE is the mean squared error function. Wherein, weight w t As shown below: In the formula, t peak α1 represents the peak time; α1 represents the peak weight.

9. The method for assessing lower limb exercise load based on a multimodal wearable device according to claim 1, characterized in that, The lower limb exercise load assessment indicators include the predicted peak vertical ground reaction force and the predicted peak tibial force.

10. A lower limb exercise load assessment system applying the method of any one of claims 1 to 9, characterized in that, include: Pressure insole device, foot pressure acquisition module, pressure leg sleeve device, surface muscle force acquisition module, surface electromyography device, surface electromyography acquisition module, inertial sensor device, acceleration signal acquisition module, training data acquisition module, tibial force prediction model construction module, vertical ground reaction force prediction model construction module, vertical ground reaction force prediction module, tibial force prediction module, lower limb exercise load assessment module; The pressure insole device is used to detect the plantar pressure of the test subject; The plantar pressure acquisition module is used to collect plantar pressure data of subjects when they run at different paces. The pressure leg sleeve device is used to detect the surface muscle strength of the calf of the test subject; The surface muscle strength acquisition module is used to acquire surface muscle strength signals of the calf muscles when the subject runs at different paces. The surface electromyography device is used to detect electromyographic signals of the calf muscles of the subject. The surface electromyography acquisition module is used to acquire surface electromyography signals of the tibialis anterior muscle when the subject runs at different paces. The inertial sensor device is used to detect acceleration signals of various parts of the subject's body; The acceleration signal acquisition module is used to acquire ankle joint acceleration signals and mid-tibial acceleration signals when the subject runs at different paces. The training data acquisition module includes an optical motion capture system, a force measurement treadmill, a data alignment module, and a tibial force calculation module. The optical motion capture system is used to collect three-dimensional trajectory information of subjects at different joint points when running at different paces; The force-measuring treadmill is used to collect the actual ground reaction force when the subject runs at different paces; The data alignment module synchronizes three-dimensional trajectory information, actual ground reaction force and wearable device signals of plantar pressure, calf muscle surface force signal, tibialis anterior muscle surface electromyography signal and acceleration signal through timestamps. The tibial force calculation module calculates the subject's tibial force based on synchronized three-dimensional trajectory information and actual ground reaction force. The tibial force prediction model building module constructs a tibial force training sample set based on synchronous wearable device signals and tibial force, and uses the tibial force training sample set to train the model to obtain the tibial force prediction model. The tibial force prediction module inputs real-time collected plantar pressure, calf muscle surface force signals, tibialis anterior muscle surface electromyography signals, and acceleration signals into the tibial force prediction model to obtain the predicted tibial force. The vertical ground reaction force prediction model construction module constructs a vertical ground reaction force training sample set based on synchronous plantar pressure, ankle joint acceleration signals, and actual ground reaction force, and uses the vertical ground reaction force training sample set to train the model to obtain the vertical ground reaction force prediction model. The vertical ground reaction force prediction module inputs the real-time collected plantar pressure and ankle joint acceleration signals into the vertical ground reaction force prediction model to obtain the predicted vertical ground reaction force. The lower limb exercise load assessment module obtains lower limb exercise load assessment indicators based on predicted tibial force and predicted vertical ground reaction force.

Images