Load mass and model prediction fused exoskeleton control method and control system

Abstract

The invention discloses an exoskeleton control method and system fusing load mass and model prediction, and belongs to the technical field of exoskeleton system control. The method comprises the following steps: acquiring data, wherein the data comprises electromyographic signals of biceps brachii and triceps brachii; based on the calculated angle, angular velocity and angular acceleration of the elbow joint; the current load quality estimation value and the load quality estimation value at the previous moment are initially set to be zero; the exoskeleton on-demand auxiliary control comprises auxiliary state recognition, specifically, the current state is judged to be an excessive auxiliary state or an under-auxiliary state by analyzing the angular velocity and the electromyographic signal of the elbow joint in the sliding window; the auxiliary torque of the motor is generated according to the recognized auxiliary state; and updating the load mass estimation value: dynamically correcting the load mass estimation value according to the auxiliary state and the auxiliary torque, and updating the load mass estimation value to realize dynamic adjustment of the auxiliary torque of the joint motor. According to the method, the real-time sensing capability of the exoskeleton system to the unknown load is realized.

Description

Technical Field

[0001] This invention relates to the intersection of human-machine collaborative control, load recognition and biological signal processing technologies, specifically to an exoskeleton control method and control system that integrates load quality and model prediction, belonging to the field of exoskeleton system control technology. Background Technology

[0002] Currently, exoskeleton systems, as key devices for human-machine integration and enhancement, are widely used in fields such as rehabilitation medicine, industrial handling, and military support. Traditional exoskeleton control strategies largely rely on preset trajectory control, force sensor feedback, or simple kinematic modeling. While they offer some assistance under specific working conditions, they still have the following limitations:

[0003] 1. Lack of dynamic sensing capability for unknown loads: Most existing exoskeleton systems require pre-setting load parameters, making it difficult to adapt to varying load conditions.

[0004] In practical applications, where load conditions change frequently or cannot be predicted in advance, this static control method lacks the ability to provide "on-demand assistance." This static control approach can easily lead to insufficient or excessive assistance, affecting user comfort and energy efficiency.

[0005] 2. Insufficient utilization of electromyographic signals makes it difficult to achieve real-time intelligent adjustment: Although electromyographic signals can reflect the user's active intentions, traditional systems often only use them as start / stop triggers or rough identification criteria, failing to fully explore their time-leading characteristics and load correlation, resulting in system lag and insufficient control precision.

[0006] 3. The control model is too simple and difficult to adapt to complex dynamic environments: Most control strategies use simple linear mapping, single model prediction or single sensor data, which are insufficient for modeling nonlinear and time-varying human motion and load changes, thus limiting the adaptability and robustness of the system in actual operation.

[0007] 4. Antagonism exists during human-computer interaction, affecting control stability and user experience: During the assistance process, if the assist torque output by the system does not match the user's actual needs, it is easy to cause human-computer antagonism, resulting in increased muscle burden, uncoordinated movements, or even user resistance to use. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, this invention provides an exoskeleton control method and system that integrates load mass and model prediction. This method and system integrate mass estimation and model prediction control via electromyography (EMG) signals, and EMG-IMU fusion model sensing load changes to achieve adaptive assist control with "sensing-response," meaning the exoskeleton provides on-demand assistance to adapt to frequent changes or situations where load conditions cannot be sensed in advance during practical applications.

[0009] An exoskeleton control method that integrates load quality and model prediction includes the following steps:

[0010] S1. Data Collection

[0011] (1) Electromyographic signals of the biceps brachii and triceps brachii;

[0012] (2) Elbow joint angle, angular velocity and angular acceleration calculated based on IMU;

[0013] (3) The current load quality estimate and the previous load quality estimate are initially set to zero;

[0014] All data is aligned and segmented using a timestamped sliding window mechanism to ensure the synchronization and timing of data from different channels.

[0015] S2, Exoskeleton On-Demand Assisted Control

[0016] (1) Assisted state recognition: By analyzing the elbow joint angular velocity and electromyographic signals within the sliding window, it is determined whether the current state is "over-assisted state" or "under-assisted state";

[0017] (2) Output auxiliary torque: Generate auxiliary torque for the motor based on the identified auxiliary state;

[0018] (3) Update the load mass estimate: Based on the auxiliary state and auxiliary torque, dynamically correct the load mass estimate and update the load mass estimate. Repeat this iterative cycle to achieve dynamic adjustment of the joint motor auxiliary torque.

[0019] Furthermore, when the identified state is an over-assisted state, the following steps are performed:

[0020] 1) Determine the current load estimation quality Overrated;

[0021] 2) Return to the previous load estimate quality That is, set ;

[0022] 3) Utilize Input the data calculated by the current IMU into the Transformer model, and output the predicted feedforward torque. ;

[0023] 4) The predicted feedforward torque The target torque is sent to the motor;

[0024] 5) Simultaneously using correction factors Adjust and update the current load estimate quality. ;

[0025] 6) Update history: Assigned ,Will Assigned .

[0026] Furthermore, when the identified state is under-assisted, the following steps are performed:

[0027] a) Estimate the current load quality The elbow joint angle and elbow joint angular acceleration calculated by the IMU are input into the Transformer model, and the output is the predicted feedforward torque. ;

[0028] b) Simultaneously input the electromyographic signals and the elbow joint angle and elbow joint angular acceleration calculated by the IMU into the Bi-LSTM model, and output the feedback torque. ;

[0029] c) Auxiliary torque The auxiliary torque is then sent to the motor via a fuzzy PID controller.

[0030] d) Input the feedback torque, elbow joint angle, and elbow joint angular acceleration into the Transformer model, and output the load mass correction value. and update the load quality estimate. The current load quality estimate Record as Then Assign to This iterative cycle enables dynamic adjustment of the joint motor's auxiliary torque by integrating both feedforward and feedback models.

[0031] An exoskeleton control system that integrates load mass and model prediction includes:

[0032] A dual-channel electromyography (EMG) signal acquisition module is used to acquire EMG signals.

[0033] The IMU module is used to collect and calculate elbow joint angle, angular velocity, and angular acceleration.

[0034] The combined drive and torque measurement module is used to drive forearm movement and collect elbow joint output torque.

[0035] The auxiliary status processing module is used for auxiliary status identification, outputting auxiliary torque, and updating load quality estimates.

[0036] The advantages of this invention compared to the prior art are:

[0037] 1. This application proposes a dynamic quality estimation scheme based on electromyographic signals, which realizes the real-time perception capability of exoskeleton systems for unknown loads. Unlike traditional control methods that rely on preset quality parameters, this application constructs a dynamic quality estimation model by analyzing the amplitude and response pattern of electromyographic signals and combining IMU measurement data, enabling the system to automatically identify and adjust when the load changes, and has good adaptability.

[0038] 2. In the under-assisted state, a dual-model collaborative control framework integrating Bi-LSTM and Transformer models was constructed to realize an on-demand assistance strategy in the under-assisted state that combines feedforward prediction and feedback compensation: The Bi-LSTM model is innovatively introduced to predict the instantaneous joint torque demand based on electromyography and IMU signals. At the same time, the Transformer model is designed to realize the bidirectional mapping between load mass and joint torque. Through multi-task learning, the modeling accuracy of nonlinear dynamic behavior is improved, and the response speed and stability of the control system to complex actions are significantly enhanced.

[0039] 3. Introducing a multi-task learning model with a task recognition mechanism in an under-assisted state to achieve unified modeling and collaborative training of joint torque and equivalent mass: This invention embeds task identifiers into the input of the Transformer model to distinguish between mass prediction and torque prediction tasks, enabling the model to learn the two regression tasks in parallel under a shared network structure, effectively improving training efficiency and model consistency.

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments: Attached Figure Description

[0041] Figure 1 This is a flowchart of the exoskeleton control method that integrates load mass and model prediction according to the present invention;

[0042] Figure 2 This is the exoskeleton control logic diagram for the present invention, which integrates load quality and model prediction.

[0043] Figure 3 This is a schematic diagram of the Bi-LSTM model structure;

[0044] Figure 4 This is a schematic diagram of the Transformer model structure;

[0045] Figure 5 A graph showing the performance evaluation metrics for a complete training run of the model;

[0046] Figure 6 This is a flowchart illustrating the experimental process of on-demand assist control in the embodiments.

[0047] Figure 7 for Figure 6Static load test and dynamic load test curves of the experimental setup;

[0048] Figure 8 This is a graph showing the changes in dual-channel electromyography signals during bicep curls with and without exoskeleton assistance, as illustrated in the example.

[0049] Figure 9 This is a graph showing the average electromyographic amplitude of participants under three load conditions in the example. Detailed Implementation

[0050] The embodiments of the technical solutions of this application will now be described in detail with reference to the accompanying drawings. Unless otherwise stated, the technical or scientific terms used in this application have the ordinary meanings understood by those skilled in the art.

[0051] Example 1: This example proposes an on-demand auxiliary control method for exoskeleton that integrates electromyography (EMG) signals for mass estimation and model-based collaborative control. This method estimates the auxiliary torque required by the current joint motors and integrates dynamic real-time estimation and correction of the load mass using EMG signals. The final estimated joint torque is sent to the motors via a fuzzy PID controller, thereby improving the overall system's target torque tracking performance. The control flow is as follows: Figure 1 As shown.

[0052] Control signal acquisition and processing

[0053] The exoskeleton loading control system collects the following signal data in real time:

[0054] Two electromyographic signals: biceps brachii emg1_t and triceps brachii emg2_t, arranged according to SENIAM;

[0055] Elbow joint angle θ, angular velocity ω, and angular acceleration α calculated based on IMU;

[0056] Current load estimate Compared with the previous time estimate Initially, m = 0.

[0057] All data is aligned and segmented using a timestamped sliding window mechanism (window length 60 ms) to ensure the synchronization and timing of data from different channels.

[0058] The entire control strategy can be divided into the following three steps:

[0059] 1. Assisted state recognition: By analyzing the elbow joint angular velocity and electromyographic signals within the sliding window, it is determined whether the current state is "over-assisted" or "under-assisted".

[0060] 2. Output auxiliary torque: Generate auxiliary torque for the motor based on the identified auxiliary status. ;

[0061] 3. Update load quality estimate: Dynamically correct the load quality estimate based on auxiliary status and auxiliary torque. And update the load quality estimate.

[0062] State recognition logic

[0063] Under-assisted state: The user actively applies force to drive the exoskeleton movement, which is characterized by increased biceps activity and relatively weak triceps.

[0064] Over-assisted state: The system applies too much force, and the user needs to actively resist it, which is manifested as increased activity of the triceps brachii and even antagonism.

[0065] The control strategy branch logic employs a combination of feedforward prediction based on a multi-task learning (MTL) model and feedback torque regulation based on a bidirectional long short-term memory (Bi-LSTM) neural network. This strategy estimates the auxiliary torque required by the joint motor while simultaneously correcting for the weight of the load in real time. The final estimated auxiliary torque is transmitted to the motor via a PID controller, improving the system's tracking performance. The specific control logic diagram is shown below. Figure 2 As shown, its core lies in estimating and dynamically tracking the mass of an unknown object using electromyography (EMG) signals, while simultaneously combining inertial measurement unit (IMU) data to predict the joint assist torque required at the next moment. The essence of this strategy is to fuse mass estimation and motion state information to generate the assist torque required for motor output in real time.

[0066] Over-Assist handling Figure 2 (left side)

[0067] 1) Determine the quality of the current estimate Overrated;

[0068] 2) Revert to the previous estimate That is, set ;

[0069] 3) Utilize Inputting current IMU observations into the Transformer model, predicting feedforward torque. ;

[0070] 4) The target torque is sent to the motor via a fuzzy PID controller;

[0071] 5) Simultaneously using correction factors Adjust and update the current valuation. ;

[0072] 6) Update history: Assigned ,Will Assigned .

[0073] Under-Assist handling Figure 2 (Right side): This state is based on a control strategy that combines feedforward predictions from a multi-task learning (MTL) model with feedback regulation based on a bidirectional long short-term memory network (Bi-LSTM).

[0074] a) Enter the current load quality estimate The elbow joint angle θ and elbow joint angular acceleration α are input into the Transformer model, and the output is the predicted feedforward torque. ;

[0075] b) Simultaneously input the EMG signal and the elbow joint angle θ and elbow joint angular acceleration α calculated by the IMU into the Bi-LSTM model, and output the feedback torque. ;

[0076] c) Auxiliary torque The auxiliary torque is sent to the motor via a fuzzy PID controller;

[0077] d) Combination θ, α, Transformer model output quality correction values And update the quality estimate: Current quality Record as Then Assign to .

[0078] Specifically: First, determine the current assist state: The signal acquisition module aligns the data from each channel using timestamps and processes it using a sliding window (one window every 60 milliseconds). By utilizing elbow joint angular velocity and dual-channel electromyography (EMG) signals, it's possible to determine whether the exoskeleton is underactuated or overacted, thus helping to infer whether the load weight estimate is overestimated or underestimated. When the exoskeleton is underactuated, the user actively applies force to move the exoskeleton; at this time, biceps activity is high, while triceps activity is relatively low. When the exoskeleton is overacted, to counteract excessive external force, the user may actively engage triceps activity, leading to increased triceps activity in the dual-channel EMG signal and potentially causing significant antagonism. After determining the state, targeted processing is performed:

[0079] When the system is determined to be in an over-assisted state, it means that the newly obtained quality estimate is higher than the actual value after the last quality correction. At this point, the current quality estimate should be restored to the previous estimate. Using the quality estimate from the previous iteration Based on the collected IMU data, the Transformer model outputs the feedforward torque. Then The target torque is sent to the joint motor via the PID controller. Simultaneously, the load estimate is updated. ,right conduct The correction factor is obtained. , will Assigned ,Will Assigned This completes one iteration cycle.

[0080] When it is determined that the exoskeleton is in an underassisted state, the following steps are performed: First, the elbow joint angle θ, elbow joint angular acceleration α, and estimated mass calculated by the IMU at the current moment are used. Input is fed into the Transformer architecture to predict feedforward torque. Simultaneously, the collected dual-channel electromyographic signals, along with θ and α calculated by the IMU, are input into a BiLSTM neural network to output the feedback torque that needs to be adjusted. The target torque is obtained by adding the feedforward torque to the feedback torque. This data is then sent to the motor via a PID controller for control. Furthermore, feedback torque is utilized. In addition to the angle θ and angular acceleration α obtained through IMU calculations, the Transformer architecture calculates the required mass correction error. Finally, the quality estimate was updated: the new quality value. and the current quality Record as Then will Assign to This completes one iteration cycle. Through the above control process, this design can dynamically adjust the auxiliary torque of the joint motor while simultaneously correcting the load weight in real time. Fuzzy PID control does not rely on a precise mathematical model of the system and has advantages in handling control problems of nonlinear systems. In this scheme, fuzzy PID control is used to improve the tracking performance of the target torque.

[0081] The on-demand assisted control method for exoskeleton that integrates electromyography signals for quality estimation and model-based collaborative control in Embodiment 1 has the following beneficial effects:

[0082] 1. Real-time estimation and dynamic tracking of unknown loads are achieved. By sensing load changes through an electromyography-IMU fusion model, adaptive assist control with "sensing and responding" is realized, i.e., on-demand assist control.

[0083] 2. Fully utilize the early response characteristics of EMG, and combine Bi-LSTM and Transformer architecture in under-assisted states to construct a feedforward + feedback collaborative control strategy to achieve fast and accurate joint torque prediction and adjustment.

[0084] 3. Establish a multi-task learning framework in under-assisted states to unify the coupling relationship between modeling quality, torque and motion state, and significantly improve the model's generalization ability and adaptability to complex dynamic behaviors.

[0085] 4. Based on electromyographic activity recognition of the auxiliary state, the control output is dynamically adjusted to effectively suppress human-machine antagonistic behavior and improve system stability and user comfort.

[0086] Example 2: Based on the control strategy of Example 1, specifically, the deviation between the estimated mass *m* and the actual load mass is reflected in the electromyographic (EMG) signal, which is then used to guide subsequent corrections to the mass estimation. Compared to the actual initiation of movement, the EMG signal typically has a 30–150 millisecond advance response characteristic, enabling the system to quickly and accurately estimate the mass error while the joint motor issues an equivalent torque command, thereby effectively offsetting the load effect.

[0087] Based on the above, this embodiment provides three types of mapping relationships and specific schemes:

[0088] 1. Predict the target joint torque required at the next moment using load estimation mass m and IMU data;

[0089] 2. Estimate the auxiliary torque required by the motor based on electromyography and IMU data;

[0090] 3. Combining the auxiliary torque and IMU information, the equivalent mass error (i.e. the residual mass deviation that needs to be dynamically eliminated) is derived.

[0091] To achieve these mappings, this embodiment designs and implements two types of models: the first mapping uses a bidirectional long short-term memory network (Bi-LSTM) to fully extract the temporal features of electromyography and IMU signals, enabling rapid and accurate prediction of joint torque; the second and third mappings are implemented through a multi-task learning (MTL) model based on the Transformer architecture, modeling the complex coupling relationship between mass, torque, and IMU data. By introducing task identifiers, the two tasks are trained collaboratively within a unified Transformer framework, thereby achieving efficient and synchronous auxiliary control and mass correction functions.

[0092] Regarding model architecture design

[0093] This section describes the architectural design of the two models required for the control strategy.

[0094] First, this approach trains a bidirectional long short-term memory (Bi-LSTM) network model to predict joint torque in real time based on electromyography (EMG) and inertial measurement unit (IMU) signals. Figure 3 The overall structure of the proposed model is shown. This model is trained using user-specific data and its ability to model time dependencies is enhanced through a Bi-LSTM structure.

[0095] The input signals include two electromyography (EMG1, EMG2) channels and two IMU channels (joint angle and angular acceleration), for a total of four sensor channels. To effectively utilize time-series information, the data is processed using a sliding window method: a window length of 60 ms (corresponding to 60 sampling points) and a sliding step size of 2 ms are used to divide the data into an input matrix of size 60×4, where 60 is the time step and 4 is the channel dimension.

[0096] It is worth noting that electromyographic (EMG) signals have a significant time lead characteristic. After device calibration, EMG responses can typically be detected approximately 55 ms before the start of movement. Based on this advantage, the model incorporates time-series inputs from the past 60 ms to predict upcoming joint torque demands, thereby achieving a faster and more precise control response.

[0097] Bidirectional Long Short-Term Memory (Bi-LSTM) is an enhanced recurrent neural network (RNN) that improves the modeling ability of time series regression tasks by simultaneously fusing forward and backward temporal information. Compared to traditional unidirectional LSTM, Bi-LSTM can comprehensively capture bidirectional time dependencies, thereby significantly improving prediction accuracy.

[0098] In one approach, the input layer receives time-series data of size 60×4, where 60 represents the time step and 4 represents the number of sensor channels. The Bi-LSTM layer consists of LSTM units in two directions, each containing 16 cells, processing the forward and reverse sequences respectively. The bidirectional outputs are then passed to a fully connected layer containing 50 neurons to extract deep features, and the output layer generates the final prediction of joint torque. Model hyperparameters (including the number of hidden units, learning rate, optimizer, etc.) are dynamically optimized based on engineering experience and hardware performance constraints.

[0099] Building upon this, this patent further designs a multi-task learning (MTL) model based on the Transformer architecture to model the nonlinear relationship between mass, joint torque, and IMU observations (angle and angular acceleration). Figure 4 As shown, this model uses a self-attention mechanism to learn the implicit physical mapping rules from the input data.

[0100] This model handles two regression tasks simultaneously:

[0101] Task 1: Map mass and IMU data to joint auxiliary torques;

[0102] Task 2: Map the auxiliary torque and IMU data to the equivalent mass.

[0103] To enable collaborative training of the two tasks within a unified model framework, a task identifier (task id) is embedded in the input vector to distinguish the mapping paths of different tasks. Since both the input and output are fixed-length scalars (4-dimensional input vector, 1-dimensional scalar output), the model adopts a Transformer structure that only contains the encoder.

[0104] The input vector x = (task id, θ, α, q) is first processed by a learned projection matrix. Mapped to a 128-dimensional feature space:

[0105] , , that is:

[0106] Positional codes are added and concatenated to form an embedded input. These are passed through a four-layer transformer encoder, where each layer includes multiple heads with self-attention mechanisms for the input / output dimensions. The attention mechanism calculates the weighted feature interactions as follows:

[0107]

[0108] in These are the query matrix, key matrix, and value matrix. It is the feature dimension of each head.

[0109] The layer outputs are processed through residual connections and layer normalization to preserve dimensionality. The feedforward network (FFN) extends the functionality to 512D using ReLU activation, followed by residual connections and normalization. A fourth location feature is extracted from the final encoder layer. Two separate fully connected layers map this to... Outputs corresponding to the predictions for Task 1 (mass) and Task 2 (torque).

[0110] Regarding model training

[0111] The data acquisition system used in this solution mainly consists of the following modules:

[0112] 1. Dual-channel electromyography (EMG) signal acquisition module: EMG electrodes are placed on the long head of the biceps brachii (BIC-LH) and the lateral head of the triceps brachii (TRI-LAT), respectively, following the SENIAM standard. The sampling frequency is 1000 Hz, the bandwidth range is 20–450 Hz, and a 50 Hz notch filter is configured to suppress power frequency interference.

[0113] 2. Forearm IMU Module: The IMU is installed at the distal third of the forearm and is used to acquire 9-axis motion data at a frequency of 100 Hz.

[0114] Dynamic data. The elbow joint angle θ, angular velocity ω, and angular acceleration α in the sagittal plane are obtained through algorithmic analysis.

[0115] 3. Combined drive and torque measurement system: The system uses an AK80-9 motor as the combined drive unit and integrates a high-precision torque sensor (Futek LSB200, range ±20 Nm, accuracy ±0.5% FS) at its joint interface. The sampling frequency is 1000 Hz, which is used to record the elbow joint output torque in real time.

[0116] The dataset used for training the Bi-LSTM model was collected under loads ranging from 1 to 5 kg. During the experiment, the subject's arm was rigidly coupled to the exoskeleton system (covering from the upper arm to the hand), and the exoskeleton effectively transmitted the force generated by the hand-held load to the elbow joint torque sensor. The motor operated in zero-force follow mode to ensure that no active torque intervention was applied. In this mode, the elbow joint torque measured by the sensor reflects the real muscle output torque required by the human body to overcome external loads, the weight and inertia of the forearm and hand during flexion and extension movements, approximating the biomechanical requirements under natural conditions.

[0117] The dataset covers 10 different load levels, with 20 trials collected under each load condition, covering different elbow flexion and extension speeds. The final data is divided into training, validation, and test sets in a ratio of 70:15:15.

[0118] Preprocessing applied a second-order Butterworth bandpass filter (20-450 Hz) to the raw electromyography data, defined by the transfer function:

[0119]

[0120] in The cutoff frequency, This represents the complex frequency variable in the Laplace transform. Furthermore, a 50 Hz notch filter eliminates power line interference. The filtered signal is then normalized using Z-score normalization.

[0121]

[0122] in: These are the original electromyography values. This is the average value. The standard deviation is denoted as .

[0123] The regression objective is a 1×1 torque matrix. Model training employed iterative parameter tuning using the GPU-based Adam optimizer (final learning rate: The prediction accuracy reached 91.2%.

[0124] In the training of the Transformer-based multi-task learning model, the upper arm structure of the exoskeleton was fixed, and loads of different masses were applied to its distal end. The pose control of the elbow joint was driven by IMU data from the Bi-LSTM dataset, while the motor output torque and corresponding load mass were recorded simultaneously at each moment.

[0125] The model input is a four-dimensional vector [task_id, θ, α, q], where:

[0126] θ: Elbow joint angle;

[0127] α: Elbow joint angular acceleration;

[0128] q: Task-related quantity, representing known torque in Task 1 and known mass in Task 2;

[0129] task_id: Task identifier, used to distinguish the regression target (0 indicates task 1: predicting quality, 1 indicates task 2: predicting torque).

[0130] The output is a one-dimensional scalar, corresponding to the target variable of the selected task (task 1 outputs mass, task 2 outputs torque).

[0131] To enhance the model's generalization ability, the input data was standardized:

[0132] Angle θ is mapped to [−π,π].

[0133] Angular acceleration, torque, and mass are standardized to zero mean and unit variance (μ=0, σ=1);

[0134] Task identifiers are used to distinguish tasks using one-hot encoding (0: Task 1, 1: Task 2).

[0135] Learnable position embedding vectors are used to encode type differences between input physical quantities, improving the model's ability to express heterogeneous features.

[0136] The dataset consists of 10 different load levels, with 20 replicates for each load. All data are divided into 70% training, 15% validation, and 15% testing.

[0137] The model is trained for a maximum of 300 training epochs, and an early stopping strategy is introduced to prevent overfitting.

[0138] Model performance is measured using the following evaluation metrics:

[0139] Mean absolute error (MAE);

[0140] Root mean square error (RMSE);

[0141] Coefficient of determination (R) 2 ).

[0142] The formulas for calculating these indicators are as follows:

[0143] Mean Absolute Error

[0144] Root mean square error

[0145] Coefficient of determination

[0146] in: , These are the actual value and the predicted value, respectively. The average of the actual values, the result is as follows Figure 5 As shown.

[0147] Example 3: To realize the exoskeleton control method that integrates load mass and model prediction, an exoskeleton control system that integrates load mass and model prediction is also provided. The system includes:

[0148] A dual-channel electromyography (EMG) signal acquisition module is used to acquire EMG signals.

[0149] The IMU module is used to collect and calculate elbow joint angle, angular velocity, and angular acceleration.

[0150] The combined drive and torque measurement module is used to drive forearm movement and collect elbow joint output torque.

[0151] An auxiliary state processing module is used for auxiliary state identification, outputting auxiliary torque, and updating load quality estimates. The system includes:

[0152] A dual-channel electromyography (EMG) signal acquisition module is used to acquire EMG signals.

[0153] The IMU module is used to collect and calculate elbow joint angle, angular velocity, and angular acceleration.

[0154] The combined drive and torque measurement module is used to drive forearm movement and collect elbow joint output torque.

[0155] The auxiliary status processing module is used for auxiliary status identification, outputting auxiliary torque, and updating load quality estimates.

[0156] Example 4: The assistive efficacy of the elbow exoskeleton system was evaluated by analyzing the amplitude fluctuations of electromyographic (EMG) signals during flexion movements under unknown load conditions, reflecting the degree of muscle exertion and observing the assistive effect. Simultaneously, the actual load mass was compared with the iterative mass prediction curve generated by the model to verify the accuracy and stability of the load estimation. Three different load masses were prepared for the wearer, and EMG signals and mass estimation curves were recorded during the loaded curling movements. Figure 6 The experimental setup and workflow are explained.

[0157] The specific experimental protocol was as follows: Five healthy subjects were recruited (age: 23.0 ± 2.0 years; height: 1.77 ± 0.03 meters; weight: 70.0 ± 10.0 kilograms). To evaluate the on-demand assist control strategy, four load configurations were set up: three 5-liter water buckets filled to approximately 1.5 kilograms, 2.5 kilograms, and 3.5 kilograms respectively, and an adhesive load of approximately 0.5 kilograms. Electromyography (EMG) sensors embedded in the biceps brachii (BIC) and triceps brachii (TRI) muscles were used to acquire EMG signals in real time, and the system output a quality estimation curve. Pre-experimental preparation included: calibration of the exoskeleton in zero-force following mode, adjustment of body size, and 3 minutes of seated acclimatization.

[0158] The static load experiment consisted of three fixed-mass bicep curls, with simultaneous recording of electromyographic signals and load estimation data. The dynamic load experiment introduced an adhesive load during the bicep curl, with relevant data collected simultaneously. The quality of the static and dynamic loads was validated after the experiments. To reduce physiological fatigue and ensure data independence, the same test was repeated three times over three consecutive days.

[0159] Figure 7 The above is shown Figure 6 Summary of experimental results; Figure 7 Both static and dynamic loads employ three types of load mass. Reference represents the actual value of the load mass, Load tracing represents the predicted value of the load mass, the horizontal axis represents time, and the vertical axis represents the load mass.

[0160] First, compare Figure 7 As shown in the comparison of predicted and measured loads, after implementing the control strategy proposed in this experiment, the estimated load can rapidly approach the measured value and converge within 0.2 seconds, with no significant fluctuations throughout the process. Furthermore, the root mean square error (RMSE) between the estimated and actual values ​​is 0.18 kg. In the dynamic load experiment, when an additional load is introduced, the control framework can quickly adjust the estimated value and accurately match the new load mass, maintaining stability.

[0161] Figure 8 Changes in two-channel electromyography (EMG) signals during bicep curls with and without exoskeleton assistance were compared. Without assistance, activation of the agonist muscle (BIC) significantly increased with increasing load mass; however, with assistance, activation remained consistently low and independent of load. The antagonist muscle (TRI) remained inhibited (RMS <0.05 mV) with assistance, occasionally even below baseline levels without assistance.

[0162] Figure 9 This study presents the mean electromyographic (EMG) amplitude analysis of the biceps brachii (BIC) and triceps brachii (TRI) muscles in five participants under three load conditions. Under unassisted conditions, the activation level of the agonist muscle (ordinate) increased with increasing load (short axis), while it remained stable under assisted conditions. The decrease in antagonist muscle activation further confirms that the human-machine antagonism phenomenon has been significantly reduced.

[0163] Through experimental analysis, the following conclusions were drawn:

[0164] First, the comparison between the load estimation curve and the actual measurement results shows that this strategy has high efficiency and accuracy in estimating unknown loads. It should be noted that the model estimate is slightly lower than the measured value. This steady-state error may originate from subconscious exertion during movement, causing fluctuations in electromyographic signals to be transmitted to the control system, thus leading to estimation bias.

[0165] Secondly, electromyography analysis in static biceps curl experiments with three loads showed that the assist control strategy effectively counteracts the effects of load, making the extra burden almost imperceptible thanks to accurate load estimation and real-time assist output.

[0166] Finally, statistical analysis of data from multiple rounds of multi-subject experiments showed that the trend of voluntary muscle activity validated the effectiveness of the strategy, while the trend of antagonistic muscle activity indicated that the human-machine antagonistic effect was minimal, further demonstrating the user-friendliness and adaptability of the method.

[0167] This application has been disclosed above with preferred embodiments, but it is not intended to limit this application. Any person skilled in the art can make some changes or modifications to the above-disclosed structure and technical content to create equivalent embodiments without departing from the scope of the technical solution of this application, and all such modifications and modifications are within the scope of the technical solution of this application.

Claims

1. An exoskeleton control method that integrates load mass and model prediction, characterized in that: The method includes the following steps: S1. Data Collection (1) Electromyographic signals of the biceps brachii and triceps brachii; (2) Elbow joint angle, angular velocity and angular acceleration calculated based on IMU; (3) The current load quality estimate and the previous load quality estimate are initially set to zero; All data is aligned and segmented using a timestamped sliding window mechanism to ensure the synchronization and timing of data from different channels. S2, Exoskeleton On-Demand Assisted Control (1) Assisted state recognition: By analyzing the elbow joint angular velocity and electromyographic signals within the sliding window, it is determined whether the current state is "over-assisted" or "under-assisted". (2) Output auxiliary torque: Generate auxiliary torque for the motor based on the identified auxiliary state; (3) Update the load mass estimate: Based on the auxiliary state and auxiliary torque, dynamically correct the load mass estimate and update the load mass estimate. Repeat this iterative cycle to achieve dynamic adjustment of the joint motor auxiliary torque.

2. The exoskeleton control method integrating load mass and model prediction according to claim 1, characterized in that: When the identified state is an over-assisted state, perform the following steps: 1) Determine the current load estimation quality Overrated; 2) Return to the previous load estimate quality That is, set ; 3) Utilize Input the data calculated by the current IMU into the Transformer model, and output the predicted feedforward torque. ; 4) The predicted feedforward torque The target torque is sent to the motor; 5) Simultaneously using correction coefficients Adjust and update the current load estimate quality. ; 6) Update history: Assigned ,Will Assigned .

3. The exoskeleton control method integrating load mass and model prediction according to claim 1, characterized in that: When the identification status is under-assisted, perform the following steps: a) Estimate the current load quality The elbow joint angle and elbow joint angular acceleration calculated by the IMU are input into the Transformer model, and the output is the predicted feedforward torque. ; b) Simultaneously input the electromyographic signals and the elbow joint angle and elbow joint angular acceleration calculated by the IMU into the Bi-LSTM model, and output the feedback torque. ; c) Auxiliary torque The auxiliary torque is then sent to the motor via a fuzzy PID controller. d) Input the feedback torque, elbow joint angle, and elbow joint angular acceleration into the Transformer model, and output the load mass correction value. and update the load quality estimate. The current load quality estimate Record as Then Assign to This iterative cycle enables dynamic adjustment of the joint motor's auxiliary torque by integrating both feedforward and feedback models.

4. The exoskeleton control method integrating load mass and model prediction according to claim 1, characterized in that: The Bi-LSTM model structure is as follows: it has two electromyography signal channels and two IMU channels. The data is processed using a sliding window method: a window length of 60 ms and a sliding step size of 2 ms are used; the data is divided into an input matrix of size 60×4, where 60 is the time step and 4 is the channel dimension.

5. The exoskeleton control method integrating load mass and model prediction according to claim 1, characterized in that: The Transformer model structure is as follows: a Transformer structure containing only encoders is used to implement two regression tasks: the first task is to map load mass and IMU data to joint torque, and the second task is to map joint torque and IMU data to equivalent mass.

6. The exoskeleton control method integrating load mass and model prediction according to claim 5, characterized in that: The mapping process for the two tasks is as follows: The input vector is x = (task id, θ, α, q), where θ represents the elbow joint angle; α represents the elbow joint angular acceleration; q represents the task-related quantity; and task_id represents the task identifier, used to distinguish the regression target. First, a learned projection matrix is ​​used. Mapped to a 128-dimensional feature space: , ; Positional encodings are added and concatenated to form an embedded input, which is then passed through a four-layer encoder, where each layer includes a self-attention mechanism for multiple heads in the input / output dimensions. The attention mechanism calculates the weighted feature interactions as follows: in, These are the query matrix, key matrix, and value matrix. It is the feature dimension of each head; The layer output is processed through residual connections and layer normalization. The feedforward network uses ReLU activation to extend the functionality to 512D, extracting the fourth location feature from the final encoder layer. Two independent fully connected layers map this to... The outputs correspond to the mass prediction and torque prediction.

7. The exoskeleton control method integrating load mass and model prediction according to claim 2, characterized in that: The Transformer model training process is as follows: The model input is a four-dimensional vector [task_id, θ, α, q], where: θ: Elbow joint angle α: Elbow joint angular acceleration q: Task-related quantities task_id: Task identifier, used to distinguish the regression target; The output is a one-dimensional scalar, corresponding to the target variable of the selected task; The input data undergoes the following standardization: angle θ is mapped to [−π,π]; angular acceleration, torque, and mass are standardized with zero mean and unit variance; task identifiers are differentiated using one-hot encoding; and learnable position embedding vectors are used to encode type differences between input physical quantities, improving the model's ability to express heterogeneous features.

8. The exoskeleton control method integrating load mass and model prediction according to claim 2 or 3, characterized in that: The training process for the Bi-LSTM model is as follows: The raw electromyography data were preprocessed using a second-order Butterworth bandpass filter, defined by the transfer function: in The cutoff frequency, Let the complex frequency variable in the Laplace transform be represented, and then the filtered signal be normalized using the Z-score: in: These are the original electromyography values. This is the average value. Standard deviation; The regression objective is a 1×1 torque matrix. Training employs a GPU-based Adam optimizer for iterative parameter tuning.

9. The exoskeleton control method integrating load mass and model prediction according to claim 8, characterized in that: Model performance is measured using mean absolute error, root mean square error, and coefficient of determination metrics. The calculation formula is as follows: Mean Absolute Error Root mean square error Coefficient of determination in , These are the actual value and the predicted value, respectively. The average of the actual values.

10. An exoskeleton control system integrating load mass and model prediction, characterized in that: The control system includes: A dual-channel electromyography (EMG) signal acquisition module is used to acquire EMG signals. The IMU module is used to collect and calculate elbow joint angle, angular velocity, and angular acceleration. The combined drive and torque measurement module is used to drive forearm movement and collect elbow joint output torque. The auxiliary status processing module is used for auxiliary status identification, outputting auxiliary torque, and updating load quality estimates.

Images