An Adaptive Multi-Scene Exoskeleton Robot Control Method and System

Abstract

This invention belongs to the field of robotics technology and discloses an adaptive multi-scene exoskeleton robot control method and system. The method includes: denoising acquired multimodal sensor data using a Kalman filter algorithm based on scene feature entropy; fusing the denoised multimodal sensor data; and extracting spatial and temporal features using a hybrid deep learning model to identify the current scene type. Based on the scene type identification result, an emotional state and motion intention mapping model is constructed to generate corresponding dynamic control commands. Based on the dynamic control commands, pulse parameters for actively and passively stimulating muscles are simultaneously output for human-machine collaborative control in multiple scenes. This invention can adapt to different types of exoskeleton robots, providing a more intelligent and precise solution for various application scenarios.

Description

Technical Field

[0001] This invention relates to the field of robotics, and in particular to an adaptive multi-scenario exoskeleton robot control method and system. Background Technology

[0002] Exoskeleton robots, as devices that enhance or assist human movement, are widely used in medical rehabilitation, industrial production, and military fields. However, existing control algorithms for exoskeleton robots are mostly designed for specific scenarios, often requiring manual intervention to adjust parameters when switching scenarios. This results in poor adaptability and insufficient control precision. Especially in complex or ever-changing environments, the robot cannot quickly adjust to the needs of different scenarios, thus affecting its overall performance and application effectiveness.

[0003] Furthermore, most existing exoskeleton robots only provide mechanical assistance and lack the ability to actively and passively stimulate muscles, especially in fields such as rehabilitation assistance, failing to fully realize their potential to promote muscle recovery and functional reconstruction. Therefore, existing technologies still have significant shortcomings in improving the adaptability of exoskeleton robots to different usage scenarios, enhancing control precision, and strengthening muscle stimulation functions.

[0004] Currently, data processing often employs simple filtering algorithms, which fail to effectively remove noise interference in complex environments. Scene recognition typically relies on data from a single sensor, resulting in low accuracy and an inability to dynamically adjust based on different environments and human body states. Furthermore, existing motion control strategies lack flexible algorithmic support, failing to adjust in real-time according to the human body's movement state and needs in different scenarios, thus limiting the versatility and effectiveness of exoskeleton robots in various application scenarios.

[0005] The methods for controlling muscle stimulation are also relatively simple, lacking precise pulse stimulation parameter calculation and dynamic adjustment mechanisms, making it difficult to achieve accurate muscle regulation and optimized rehabilitation effects. These technical problems result in poor versatility of exoskeleton robots in practical applications, limiting their further development and promotion.

[0006] Therefore, how to provide an adaptive multi-scenario exoskeleton robot control method and system is an urgent problem to be solved. Summary of the Invention

[0007] This invention provides an adaptive multi-scenario exoskeleton robot control method and system to solve the aforementioned technical problems in the prior art.

[0008] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or to describe the scope of protection of these embodiments. Its sole purpose is to present some concepts in a simple form as a prelude to the detailed description that follows.

[0009] According to a first aspect of the present invention, an adaptive multi-scenario exoskeleton robot control method is provided.

[0010] In one embodiment, an adaptive multi-scenario exoskeleton robot control method includes:

[0011] The Kalman filter algorithm based on scene feature entropy is used to denoise the collected multimodal sensor data, and the denoised multimodal sensor data is fused. Spatial features and time series features are extracted by combining a hybrid deep learning model to identify the current scene type.

[0012] Based on the current scene type recognition results, an emotional state and motor intention mapping model is constructed to generate corresponding motor control commands. Based on the motor control commands, pulse parameters for active and passive muscle stimulation are output synchronously to enable human-machine collaborative control in multiple scenarios.

[0013] In one embodiment, the step of using a Kalman filter algorithm based on scene feature entropy to denoise the acquired multimodal sensor data, fusing the denoised multimodal sensor data, and combining a hybrid deep learning model to extract spatial and temporal features to identify the current scene type includes:

[0014] Multimodal sensor data is collected, and the Kalman filter algorithm based on scene feature entropy is used to denoise the multimodal sensor data. Then, the denoised multimodal sensor data is fused by combining the reinforcement learning decision mechanism based on biomechanical constraints and the spatiotemporal fusion algorithm.

[0015] The fused multimodal sensor data is input into a hybrid deep learning model composed of convolutional neural networks and recurrent neural networks to extract spatial features and time series features respectively.

[0016] Spatial and temporal features extracted by a fully connected layer are fused together, and the probability distribution of various scenarios is calculated by a normalized exponential function to identify the current scenario type.

[0017] In one embodiment, the process of acquiring multimodal sensor data, denoising the multimodal sensor data using a Kalman filter algorithm based on scene feature entropy, and fusing the denoised multimodal sensor data by combining a reinforcement learning decision-making mechanism based on biomechanical constraints with a spatiotemporal fusion algorithm includes:

[0018] Multimodal sensor data containing environmental information and human motion data are collected, and the initial noise covariance matrix is ​​set using the Kalman filter algorithm to perform preliminary denoising on the multimodal sensor data;

[0019] Based on the collected multimodal sensor data, the scene feature entropy is calculated in real time, and the initial noise covariance matrix is ​​dynamically adjusted according to the calculation results to denoise the multimodal sensor data.

[0020] Based on the dynamically adjusted initial noise covariance matrix, the biomechanical constraints are encoded into a penalty function using a reinforcement learning algorithm, and a reinforcement learning framework is constructed by combining it with a dual-delay deep deterministic policy gradient algorithm.

[0021] Based on scene feature entropy, a reinforcement learning framework is trained using a meta-learning algorithm, and combined with a spatiotemporal attention mechanism, the weights of multimodal sensor data are dynamically adjusted.

[0022] Based on the adjusted weights of the multimodal sensor data, the denoised multimodal sensor data is fused.

[0023] In one embodiment, the probability distribution is expressed as:

[0024] ;

[0025] In the formula, y i Indicates the first i The probability of each scenario; z i The output of the fully connected layer represents the first... i Scores for each scenario; C Indicates the total number of scene categories; i Indicates the index value; z j The output of the fully connected layer represents the first... j Scores for each scenario.

[0026] In one embodiment, the step of constructing an emotional state-motor intention mapping model based on the current scene type recognition result, generating corresponding motor control commands, and synchronously outputting pulse parameters for actively and passively stimulating muscles based on the motor control commands to perform human-machine collaborative control in multiple scenarios includes:

[0027] Based on the recognition results of the current scene type, and combined with the motion control strategy of emotion state recognition, an emotion state and motion intention mapping model is constructed, and corresponding power control commands are generated based on the emotion state and motion intention mapping model.

[0028] Based on the generated power control commands, combined with the collaborative control algorithm of synchronous muscle stimulation and the target trajectory and torque mapping function, the pulse width is adjusted, and the pulse parameters of active and passive muscle stimulation are output synchronously to perform human-machine collaborative control in multiple scenarios.

[0029] In one embodiment, the step of constructing an emotion state-motor intention mapping model based on the recognition result of the current scene type, combined with the motion control strategy of emotion state recognition, and generating corresponding power control commands based on the emotion state-motor intention mapping model includes:

[0030] Based on the recognition results of the current scene type, collect the user's physiological signals and extract emotional features from the physiological signals;

[0031] The extracted emotional features are input into a hybrid deep learning model, and the weights of the emotional features are optimized through an attention mechanism to output the user's emotional state and the corresponding quantitative value.

[0032] Based on the quantitative value of emotional state, we define the emotional influence factor and construct a mapping model between emotional state and movement intention.

[0033] By combining emotional state and the emotional state-motor intention mapping model, the controller's control parameters are dynamically adjusted to generate corresponding power control commands, which include joint angle control commands, motion speed control commands, and assist coefficient calculations.

[0034] In one embodiment, the expression for the joint angle control command is:

[0035] ;

[0036] In the formula, θ cmd Indicates joint angle control commands; θ 0 Indicates the base angle value; k a Indicates the acceleration coefficient; a Represents acceleration data; k f Indicates the force coefficient; f This represents data on human force exertion;

[0037] The expression for the motion speed control command is:

[0038] ;

[0039] In the formula, v cmd This indicates a motion speed control command; v 0 Indicates the base speed;α Indicates the deceleration coefficient; d Indicates the distance to the target;

[0040] The expression for the assist coefficient is:

[0041] ;

[0042] In the formula, λ Indicates the assist coefficient; λ 0 Indicates the basic assist coefficient; β Indicates the adjustment factor; s This indicates the intensity of electrical signals in human muscles.

[0043] In one embodiment, the step of adjusting the pulse width and simultaneously outputting the pulse parameters of actively stimulated and passively stimulated muscles based on the generated power control commands, combined with a collaborative control algorithm for synchronous muscle stimulation and a target trajectory and torque mapping function, to perform human-machine collaborative control in multiple scenarios includes:

[0044] Based on the generated dynamic control commands and human movement physiology, a multi-muscle group functional coordination matrix is ​​established, and the correlation coefficient between muscle groups is calculated through electromyographic signal coherence analysis to determine the time offset of synchronous stimulation.

[0045] Based on the determined time offset, a distributed pulse stimulation array is designed, and combined with a predefined synergistic stimulation coefficient, the stimulation intensity ratio between auxiliary muscle groups and core muscle groups is dynamically adjusted.

[0046] Based on the adjusted stimulation intensity ratio, a target trajectory and torque mapping function is established to convert the motion trajectory into target torque and calculate the human-machine torque distribution ratio to coordinate the distribution of pulse intensity and exoskeleton mechanical assistance.

[0047] By combining a preset torque change rate threshold and using the torque change trend predicted by a long short-term memory network, the pulse width is adjusted, and the pulse parameters for active and passive muscle stimulation are output simultaneously to enable human-machine collaborative control in multiple scenarios.

[0048] In one embodiment, the expression for the pulse width is:

[0049] ;

[0050] In the formula, W a Indicates the pulse width; W a0 Indicates the basic pulse width; k t Indicates the torque error coefficient; This indicates the deviation between the actual torque and the target torque.

[0051] The expression for the pulse parameters of the actively stimulated muscle is:

[0052] ;

[0053] In the formula, I a Indicates the intensity of the active stimulation current; I a0 Indicates the intensity of the active stimulation baseline current; k s Indicates the muscle electrical signal intensity coefficient; s Indicates the intensity of electrical signals in human muscles; T a Indicates the active stimulation pulse period; T a0 This indicates the basic pulse period of active stimulation; k d Indicates the distance coefficient; d Indicates the distance to the target;

[0054] The expression for the pulse parameters of the passively stimulated muscle is:

[0055] ;

[0056] In the formula, I p Indicates the intensity of the passive stimulation current; I p0 Indicates the baseline current intensity of passive stimulation; k θ Indicates the joint angle deviation coefficient; θ This indicates the deviation between the actual joint angle and the target joint angle; T p Indicates the passive stimulation pulse period; T p0 This indicates the basic pulse period of passive stimulation; k v Indicates the velocity coefficient; v This indicates the speed of movement of the exoskeleton robot.

[0057] According to a second aspect of the present invention, an adaptive multi-scenario exoskeleton robot control system is provided.

[0058] In one embodiment, the adaptive multi-scenario exoskeleton robot control system includes:

[0059] The scene recognition module is used to denoise the acquired multimodal sensor data using a Kalman filter algorithm based on scene feature entropy, and to fuse the denoised multimodal sensor data. It then combines a hybrid deep learning model to extract spatial and temporal features in order to identify the current scene type.

[0060] The collaborative control module is used to construct a mapping model between emotional state and motor intention based on the current scene type recognition results, generate corresponding power control commands, and synchronously output pulse parameters for active and passive muscle stimulation based on the power control commands, so as to carry out human-machine collaborative control in multiple scenarios.

[0061] According to a third aspect of the present invention, a computer device is provided.

[0062] In some embodiments, the computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0063] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.

[0064] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the above method.

[0065] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0066] 1. This invention introduces the Kalman filter algorithm to remove noise and improve data quality, thereby providing a stable and reliable data foundation for subsequent processing. At the same time, the application of the hybrid deep learning model improves the accuracy and robustness of scene recognition, enabling the system to better adapt to various changes in dynamic environments.

[0067] 2. This invention enables the exoskeleton robot to adapt to different scenarios and human conditions by adjusting control parameters in real time, thereby improving the safety and comfort of movement. In addition, by precisely controlling active and passive muscle stimulation, active stimulation helps the human body to exert force actively, while passive stimulation assists the human body to complete passive movement, which greatly enhances the effect of the exoskeleton robot in rehabilitation and other scenarios. It can adapt to different types of exoskeleton robots and provides a more intelligent and precise solution for a variety of application scenarios.

[0068] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0069] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0070] Figure 1 This is a flowchart illustrating an adaptive multi-scenario exoskeleton robot control method according to an exemplary embodiment;

[0071] Figure 2 This is a schematic block diagram illustrating an adaptive multi-scenario exoskeleton robot control system according to an exemplary embodiment;

[0072] Figure 3 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment;

[0073] Figure 4 This is an overall framework diagram of an adaptive multi-scenario exoskeleton robot control method according to an exemplary embodiment;

[0074] Figure 5 This is one of the schematic diagrams of a scene recognition hybrid model structure for an adaptive multi-scene exoskeleton robot control method according to an exemplary embodiment;

[0075] Figure 6 This is the second schematic diagram of a scene recognition hybrid model structure for an adaptive multi-scene exoskeleton robot control method, according to an exemplary embodiment. Detailed Implementation

[0076] It should be understood that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order constraint on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the diagram may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0077] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0078] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0079] Figure 1 An embodiment of an adaptive multi-scenario exoskeleton robot control method of the present invention is shown.

[0080] In this optional embodiment, the adaptive multi-scenario exoskeleton robot control method includes:

[0081] Step S101: The Kalman filter algorithm based on scene feature entropy is used to denoise the collected multimodal sensor data, and the denoised multimodal sensor data is fused. Spatial features and time series features are extracted by combining a hybrid deep learning model to identify the current scene type.

[0082] In this optional embodiment, the step of using a Kalman filter algorithm based on scene feature entropy to denoise the acquired multimodal sensor data, fusing the denoised multimodal sensor data, and combining a hybrid deep learning model to extract spatial and temporal features to identify the current scene type includes:

[0083] Multimodal sensor data is collected, and the Kalman filter algorithm based on scene feature entropy is used to denoise the multimodal sensor data. Then, the denoised multimodal sensor data is fused by combining a reinforcement learning decision mechanism based on biomechanical constraints and a spatiotemporal fusion algorithm.

[0084] In this optional embodiment, the process of acquiring multimodal sensor data, denoising the multimodal sensor data using a Kalman filter algorithm based on scene feature entropy, and fusing the denoised multimodal sensor data by combining a reinforcement learning decision-making mechanism based on biomechanical constraints with a spatiotemporal fusion algorithm includes:

[0085] Multimodal sensor data containing environmental information and human motion data are collected, and the initial noise covariance matrix is ​​set using the Kalman filter algorithm to perform preliminary denoising on the multimodal sensor data;

[0086] It should be noted that different sensor types include: IMU, electromyography (EMG) sensors, and plantar pressure sensors. For example, the accelerometer of an IMU initially... Can be set to gyroscope initial Set as It was calibrated through multiple no-load tests.

[0087] Based on the collected multimodal sensor data, the scene feature entropy is calculated in real time, and the initial noise covariance matrix is ​​dynamically adjusted according to the calculation results to denoise the multimodal sensor data.

[0088] It should be noted that the expression for scene feature entropy is:

[0089] ;

[0090] In the formula, H Represents the scene feature entropy; n Indicates the number of feature categories; i Indicates the index value; p i Indicates the first i Scene characteristics are calculated through historical data statistics or real-time sliding window calculations.

[0091] Based on the dynamically adjusted initial noise covariance matrix, the biomechanical constraints are encoded into a penalty function using a reinforcement learning algorithm, and a reinforcement learning framework is constructed by combining it with a dual-delay deep deterministic policy gradient algorithm.

[0092] Based on scene feature entropy, a reinforcement learning framework is trained using a meta-learning algorithm, and combined with a spatiotemporal attention mechanism, the weights of multimodal sensor data are dynamically adjusted.

[0093] It should be noted that in the spatiotemporal attention mechanism, the temporal attention weight is based on the gait cycle phase allocation, while the spatial attention weight is based on the sensor signal-to-noise ratio allocation.

[0094] Based on the adjusted weights of the multimodal sensor data, the denoised multimodal sensor data is fused.

[0095] It should be noted that the formula for the Kalman filter algorithm is:

[0096] ;

[0097] In the formula, express k The optimal estimate at time; express k Time based k The predicted value at time -1; K k Indicates Kalman gain; z k express k The measured value at time; H This represents the measurement matrix.

[0098] The prediction equation for the Kalman filter algorithm is as follows:

[0099] ;

[0100] In the formula, A represents the predicted value at the current time based on the optimal estimate at the previous time step; A represents the state transition matrix. Let B represent the optimal estimate at the previous time step, and let B represent the control input matrix. This indicates the control input from the previous moment; represents the prediction error covariance matrix; Q represents the process noise covariance matrix.

[0101] The update equation for the Kalman filter algorithm is:

[0102] ;

[0103] In the formula, K k Indicates Kalman gain; H Represents the measurement matrix; T Indicates transpose; R Represents the measurement noise covariance matrix; z k This represents the measurement value at the current moment. I Represents the identity matrix; This represents the optimal estimate at the current moment; This represents the estimation error covariance matrix at the current moment. The Kalman filter algorithm effectively removes noise from the data, improving its accuracy and stability, thus laying a solid foundation for subsequent scene recognition, control strategy generation, and pulse stimulation parameter calculation.

[0104] In addition, multimodal sensor data is collected, and the Kalman filter algorithm based on scene feature entropy is used to denoise the multimodal sensor data. Then, a reinforcement learning decision-making mechanism based on biomechanical constraints and a spatiotemporal fusion algorithm are combined to fuse the denoised multimodal sensor data. Specifically, this includes:

[0105] I. Kalman Filter Algorithm:

[0106] 1. Self-adjusting mechanism of dynamic noise covariance matrix: The noise covariance matrix of conventional Kalman filtering is a fixed value, which is prone to filtering lag in complex scenes (slopes, uneven ground). This is addressed by introducing scene feature entropy for real-time adjustment. ;in The feature entropy of the current scene (information entropy for calculating gait period and joint angle change rate). The adaptive coefficient (dynamically adjusted from 0.1 to 0.3) This is the initial noise matrix. When a sudden change in ground level (from flat ground to stairs) is detected... sudden increase Increase the filter's response speed to sudden changes by 3-5 times; in stable scenarios... Approaching 0, Shrink to reduce noise interference.

[0107] 2. Multimodal observation weighted fusion: Traditional algorithms rely on only a single sensor (electromyography). By fusing IMU posture data and plantar pressure sensor signals, the weights of the observations are dynamically allocated through confidence weights. The weights satisfy Furthermore, based on the real-time signal-to-noise ratio updates of each sensor: .

[0108] For example, the plantar pressure signal during climbing stairs has high reliability. Increased to 0.5), electromyographic signal weighting was dominant during walking on flat ground ( (Keep 0.6); During scene switching, the root mean square error of the filtered signal is reduced by 42%, and the joint angle prediction lag time is shortened from 80ms to 35ms.

[0109] II. Reinforcement Learning Decision-Making Mechanism Based on Biomechanical Constraints:

[0110] The physical constraints of the exoskeleton (joint range of motion, maximum driving torque) are encoded into a penalty function for reinforcement learning: ;in Let j be the actual angle of the j-th joint. For safety limits, The penalty coefficient is used; a double-delay deep deterministic policy gradient algorithm is adopted, and a biomechanical constraint network is added to the Actor-Critic framework to enable the decision-making process to naturally avoid damage risks.

[0111] For different scenarios (loaded / unloaded, flat / slope), the basic model is trained through meta-learning, and convergence is achieved in new scenarios with only 50 interactions. Scene features (slope, load value) are used as meta-parameters input to the policy network, enabling the model to quickly adapt to parameterized scene changes. When walking on a 30° slope, the exoskeleton's energy consumption is reduced by 28%, and the gait synchronization error with the human body is reduced from ±15° to ±5°, achieving "adaptive efficiency optimization under safety constraints".

[0112] III. Multi-sensor spatiotemporal fusion algorithm based on attention mechanism:

[0113] The design incorporates a spatiotemporal attention module: temporal attention, which assigns weights to sensor data at different times (the support phase of the gait cycle has a higher weight than the swing phase); and spatial attention, which dynamically highlights key sensors (prioritizing hip joint sensors when going uphill or downhill, and emphasizing knee joint sensors on flat ground); and feature fusion. ,in, This is the attention weight matrix.

[0114] As a pre-processing module for Kalman filtering, the key features selected through the attention mechanism can improve the filtering convergence speed by 30%. When the signal is interrupted (EMG sensor detachment), the weight of redundant sensors is automatically enhanced, extending the system fault tolerance time from 200ms to 500ms. The effective information utilization rate of multi-sensor data is increased from 65% to 92%, and control accuracy can still be maintained at over 85% even when some sensors fail. Introducing biomechanical constraints into reinforcement learning and endowing the Kalman filter with scene entropy adaptive capability breaks through the limitations of traditional fields where "filtering is only used for noise reduction" and "reinforcement learning only focuses on action optimization." This allows the three technical improvements to form a closed loop, giving full play to their respective advantages. By introducing the attention mechanism, data quality is improved, and the improved Kalman filter can better suppress noise. Reinforcement learning can achieve efficient decision-making based on clean data. Overall performance, including scene adaptation speed and control accuracy, is improved, ensuring that the exoskeleton robot can adapt to complex and changing usage scenarios more efficiently and accurately.

[0115] The fused multimodal sensor data is input into a hybrid deep learning model composed of convolutional neural networks and recurrent neural networks to extract spatial features and time series features, respectively.

[0116] It should be noted that the training process of the hybrid deep learning model includes the scale of the training dataset (500 groups of users, 10 scenarios, each group containing 1000 time steps of data), data augmentation methods (time series flipping, additive Gaussian noise), loss function (such as cross-entropy loss), and optimizer (Adam, learning rate 0.001, decay coefficient 0.9) to improve the model's credibility; the parameters of the hybrid deep learning model are shown in Table 1.

[0117] Table 1: Parameter Table of Hybrid Deep Learning Model

[0118] ;

[0119] like Figure 5 , Figure 6 As shown, the fused multimodal sensor data is input into a hybrid deep learning model composed of convolutional neural networks and recurrent neural networks to extract spatial features and time series features, specifically including:

[0120] Convolutional Neural Network Feature Extraction: Convolutional layers extract spatial features of data through convolution operations, as shown in the formula. ;in The convolutional layer outputs the feature value at position (i,j) on the feature map. The ReLU activation function is used, i.e. It can solve the gradient vanishing problem and accelerate model convergence. After multiple convolution and pooling operations, the spatial feature vector of the input data is obtained.

[0121] Temporal processing of recurrent neural networks: The spatial feature vector output by the convolutional neural network is input into the recurrent neural network to capture the time series features of the data; the hidden layer state update formula of the RNN is: ;in, Let be the hidden layer state at time t. This is the weight matrix from the input layer to the hidden layer. Let be the input feature vector at time t. This is the weight matrix from hidden layer to hidden layer. The hidden layer state at time t-1 This is the bias term for the hidden layer.

[0122] Spatial and temporal features extracted by a fully connected layer are fused together, and the probability distribution of various scenarios is calculated by a normalized exponential function to identify the current scenario type.

[0123] In this optional embodiment, the expression for the probability distribution is:

[0124] ;

[0125] In the formula, y i Indicates the first i The probability of each scenario; z i The output of the fully connected layer represents the first... i Scores for each scenario; C Indicates the total number of scene categories; i Indicates the index value; z j The output of the fully connected layer represents the first... j The scores for each scene are calculated; the scene with the highest probability is selected as the recognition result.

[0126] It should be noted that the total number of scene categories C It includes at least five scenarios: walking on flat ground, going up and down stairs, climbing slopes, rehabilitation training, and static support, and is determined through training with more than 500 sets of multi-scenario sample data.

[0127] Step S102: Based on the current scene type recognition result, construct an emotional state and motion intention mapping model, generate corresponding power control commands, and synchronously output pulse parameters for active and passive muscle stimulation based on the power control commands to perform human-machine collaborative control in multiple scenarios.

[0128] In this optional embodiment, the step of constructing an emotional state and motor intention mapping model based on the current scene type recognition result, generating corresponding motor control commands, and synchronously outputting pulse parameters for active and passive muscle stimulation based on the motor control commands to perform human-machine collaborative control in multiple scenarios includes:

[0129] Based on the recognition results of the current scene type, and combined with the motion control strategy of emotion state recognition, an emotion state and motion intention mapping model is constructed, and corresponding power control commands are generated based on the emotion state and motion intention mapping model.

[0130] In this optional embodiment, the step of constructing an emotion state-motor intention mapping model based on the recognition result of the current scene type, combined with the motion control strategy of emotion state recognition, and generating corresponding power control commands based on the emotion state-motor intention mapping model includes:

[0131] Based on the recognition results of the current scene type, collect the user's physiological signals and extract emotional features from the physiological signals;

[0132] It should be noted that physiological signals include PPG, GSR, and speech, and a specific example for extracting emotional features is as follows: The standard deviation of the total sinus RR intervals in heart rate variability is calculated as follows: ,in, For the first RR intervals, The mean RR interval, The total number of RR intervals; the fundamental frequency (F0) in the speech emotion features is extracted by autocorrelation method, with a frame length of 20ms and a frame shift of 10ms.

[0133] The extracted emotional features are input into a hybrid deep learning model, and the weights of the emotional features are optimized through an attention mechanism to output the user's emotional state and the corresponding quantitative value.

[0134] It should be further noted that, in order to establish an objective standard for quantifying emotional states, 200 subjects will be recruited, covering different ages, genders, and health conditions to ensure the representativeness of the sample. Under a pre-set, controllable experimental scenario (calm sitting, exercise fatigue, simulated dangerous environment), multimodal physiological signals (heart rate variability, skin conductance response, EEG, etc.) will be collected simultaneously, and the data will be labeled in conjunction with the subjects' subjective emotional scores (using a 1–5 Likert scale). Based on the above data, a mapping model or quantitative comparison table between emotional states and physiological signal characteristics will be constructed to improve the interpretability and objectivity of the emotion recognition results.

[0135] Based on the quantitative value of emotional state, we define the emotional influence factor and construct a mapping model between emotional state and movement intention.

[0136] By combining emotional state and the emotional state-motor intention mapping model, the controller's control parameters are dynamically adjusted to generate corresponding power control commands, which include joint angle control commands, motion speed control commands, and assist coefficient calculations.

[0137] In this optional embodiment, the expression for the joint angle control command is:

[0138] ;

[0139] In the formula, θ cmd Indicates joint angle control commands; θ 0 This represents the basic angle value, obtained through experimental calibration. k a This represents the acceleration coefficient, which is adjusted according to different scenarios. a Represents acceleration data; k f This indicates the force coefficient, which can be adjusted according to different scenarios; f This represents data on human force exertion;

[0140] The expression for the motion speed control command is:

[0141] ;

[0142] In the formula, v cmd This indicates a motion speed control command; v 0 Indicates the base speed; α Indicates the deceleration coefficient; d It indicates the distance to the target, calculated using LiDAR data. When the distance to an obstacle is close, the speed automatically decreases to ensure safety during movement.

[0143] The expression for the assist coefficient is:

[0144] ;

[0145] In the formula, λ Indicates the assist coefficient; λ 0 Indicates the basic assist coefficient; β Indicates the adjustment factor; s It represents the intensity of the electrical signal in human muscles. The stronger the electrical signal, the smaller the assist coefficient, thus enabling human-machine coordinated force exertion.

[0146] It should be further explained that, based on the recognition results of the current scene type and combined with the motion control strategy of emotion state recognition, an emotion state-motor intention mapping model is constructed, and corresponding dynamic control commands are generated based on the emotion state-motor intention mapping model, specifically including:

[0147] I. Introduction of the Emotional State Recognition Module:

[0148] 1. Multimodal emotional feature acquisition: Based on the existing sensor system, a new physiological signal acquisition unit is added, including a wrist PPG heart rate sensor (sampling rate 200Hz), a skin conductance response (GSR) sensor (sampling rate 100Hz), and a voice emotion acquisition microphone (16kHz sampling).

[0149] Three types of core emotional features were extracted: physiological features, such as time-domain indicators of heart rate variability (HRV) (e.g., SDNN, RMSSD), peak frequency and amplitude of GSR; speech features, such as fundamental frequency (F0) fluctuation range and speech energy entropy; and motion-related features, such as gait cycle fluctuation rate (≤5% in calm state, rising to 15%-20% in tense state).

[0150] 2. Emotional State Classification Model: A CNN-LSTM hybrid model is adopted, with the input layer fusing multimodal features (64 dimensions). The weight of key features is enhanced through an attention mechanism (HRV weight is increased to 0.35 in the anxious state). The output is divided into 5 emotional states and corresponding quantitative values: calm (0.1-0.3), focused (0.4-0.6), fatigued (0.7-0.8), anxious (0.8-0.9), and panicked (≥0.9). The classification accuracy should reach more than 89% (validated by training with 500 sets of user data).

[0151] 3. Real-time optimization: An incremental learning mechanism is introduced, and the model updates the classification boundary every 10 sets of new data to avoid drift caused by individual emotional differences; model quantization compression technology is adopted to control the inference latency within 20ms to meet the real-time control requirements of the exoskeleton (control cycle is 50ms).

[0152] II. Construction of the Emotional State-Motor Intent Mapping Model:

[0153] 1. Mathematical expression of the mapping relationship: Define the emotional influence factor λ (range 0.8-1.5), and establish a mapping function between emotional state and motor assistance force: in As a basic supportive force without emotional intervention, The current emotional quantification value, This is the baseline calm value.

[0154] Values ​​of λ under different emotional states: Calm / Focused: λ=0.1 (assistive force fluctuation ≤5%); Fatigue: λ=-0.2 (assistive force increased by 20%-30%, with a focus on increasing hip joint driving torque); Anxiety / Panic: λ=0.3 (assistive force decreased by 15%-20%, while increasing the damping coefficient to inhibit excessive movement).

[0155] 2. Dynamic Mapping Adjustment Mechanism: Constructing an emotion-scenario dual-factor adjustment matrix; for example, anxiety state + stair climbing scenario: λ is increased to 0.4, assist force is reduced by 25% while joint stiffness is increased by 10% to enhance stability; fatigue state + walking on flat ground: λ is decreased to -0.3, assist force is increased by 35%, and step length is automatically shortened (from 70cm to 55cm) to reduce energy consumption; the mapping relationship is generated through offline training using reinforcement learning, and the reward function is designed as follows: .

[0156] 3. Individual Adaptability Optimization: An emotion-motor preference database is established for each user, recording subjective comfort scores (1-5 points) under different emotional states, and the mapping parameters are iteratively updated through a Bayesian optimization algorithm; upon first use, a 3-minute guided training is conducted, including simulated scenarios such as calm, fatigue, and focus, to complete the calibration of basic parameters.

[0157] III. Achieving dynamic adjustment of motion assistance force:

[0158] 1. In the original control strategy generation algorithm, a new emotion adjustment layer is added: receiving E(t) output from the emotion state recognition module; calling the mapping model to calculate the λ value and the auxiliary force correction amount; dynamically adjusting the PID parameters of the joint controller (reducing the proportional coefficient Kp by 20% when anxious, and shortening the integral time Ti by 30% when fatigued); adjustment granularity: the hip joint, knee joint, and ankle joint are adjusted independently, for example, the knee joint auxiliary force is increased by 10% higher than that of the hip joint when fatigued.

[0159] In a fatigued state: user walking energy consumption is reduced by 32% (verified by metabolic rate monitoring), and gait cycle stability is improved by 40% (fluctuation rate reduced from 18% to 11%); in an anxious state: the probability of accidental imbalance is reduced by 65% ​​(through pressure platform testing, the imbalance judgment standard is that the center of pressure on the sole of the foot shifts by more than 2cm); in mixed scenarios (fatigue + uphill and downhill): the assistive force adjustment response speed is 2 times faster than the traditional solution, avoiding the feeling of sluggishness caused by adjustment lag.

[0160] 2. When a panic state is detected (E(t)≥0.9), the safety protection mode is triggered: the assist force is instantly reduced to 30%, the maximum flexion angle of the knee joint is locked (≤90°), and a voice prompt "Please stop walking" is given; when the emotion recognition module fails, it automatically switches to the basic control mode to ensure the robustness of the system.

[0161] In addition, the basic angle values ​​in the joint angle control commands are adjusted for different scenario types (walking on flat ground, going up and down stairs, rehabilitation training). acceleration coefficient Force coefficient The specific values ​​are shown in Table 2.

[0162] Table 2: Numerical Table of Joint Angle Control Commands

[0163] ;

[0164] Among these, to ensure the safety and adaptability of human-machine collaboration, the assist coefficient is... Adjusting and setting dynamic boundary conditions, when the intensity of human muscle electrical signals... When the threshold is exceeded (50mV, corresponding to excessive force), the assist coefficient... The lower limit is set to 0.2 to avoid the exoskeleton hindering movement in the opposite direction; when When the voltage drops below the threshold (5mV, corresponding to a lack of power), The upper limit is set to 0.9 to prevent excessive assistance from leading to dependency.

[0165] Based on the generated power control commands, combined with the collaborative control algorithm of synchronous muscle stimulation and the target trajectory and torque mapping function, the pulse width is adjusted, and the pulse parameters of active and passive muscle stimulation are output synchronously to perform human-machine collaborative control in multiple scenarios.

[0166] In this optional embodiment, the step of adjusting the pulse width and synchronously outputting the pulse parameters of actively stimulated and passively stimulated muscles based on the generated power control commands, combined with the collaborative control algorithm of synchronous muscle stimulation and the target trajectory and torque mapping function, to perform human-machine collaborative control in multiple scenarios includes:

[0167] Based on the generated dynamic control commands and human movement physiology, a multi-muscle group functional coordination matrix is ​​established, and the correlation coefficient between muscle groups is calculated through electromyographic signal coherence analysis to determine the time offset of synchronous stimulation.

[0168] Based on the determined time offset, a distributed pulse stimulation array is designed, and combined with a predefined synergistic stimulation coefficient, the stimulation intensity ratio between auxiliary muscle groups and core muscle groups is dynamically adjusted.

[0169] Based on the adjusted stimulation intensity ratio, a target trajectory and torque mapping function is established to convert the motion trajectory into target torque and calculate the human-machine torque distribution ratio to coordinate the distribution of pulse intensity and exoskeleton mechanical assistance.

[0170] It should be further noted that, in the specific embodiment, for knee flexion and extension movements, the target trajectory and torque mapping function is set as follows: ,in (proportion coefficient) (differential coefficients) (Coefficient of viscous friction) is set according to the scenario, such as when walking on flat ground. .

[0171] By combining a preset torque change rate threshold and using the torque change trend predicted by a long short-term memory network, the pulse width is adjusted, and the pulse parameters for active and passive muscle stimulation are output simultaneously to enable human-machine collaborative control in multiple scenarios.

[0172] It should be noted that the Long Short-Term Memory (LSTM) network is used to predict torque change trends. Its input features include the torque sequence and joint angle change rate over the past 100ms. The network structure uses a 2-layer LSTM, with each layer containing 64 neural units. The prediction step size is 50ms. To improve prediction accuracy, an error correction mechanism based on real-time torque feedback is introduced, which dynamically adjusts the prediction results by setting a correction coefficient of 0.1.

[0173] In this optional embodiment, the expression for the pulse width is:

[0174] ;

[0175] In the formula, W a Indicates the pulse width; W a0 This indicates the basic pulse width, which is 50-200μs. k t This represents the torque error coefficient, which is 5 μs / (N·m). This indicates the deviation between the actual torque and the target torque.

[0176] The expression for the pulse parameters of the actively stimulated muscle is:

[0177] ;

[0178] In the formula, I a Indicates the intensity of the active stimulation current; I a0 Indicates the intensity of the active stimulation baseline current; k s Indicates the muscle electrical signal intensity coefficient; s Indicates the intensity of electrical signals in human muscles; T a This indicates the active stimulation pulse period, in milliseconds (ms). T a0 This represents the basic pulse period of active stimulation, with a value ranging from 50 to 100 ms. kd This represents the distance coefficient, with a value ranging from 5 to 10. d It indicates the distance to the target, in meters (m). The closer the distance to the target, the shorter the pulse period and the higher the stimulation frequency, alerting the body to prepare.

[0179] Furthermore, active stimulation is used to assist the human body in actively exerting force. Its pulse parameters are related to the intensity of the human muscle electrical signal and the distance to the target, and vary depending on the scenario. Different values, such as in the scenario of walking on flat ground. Scenes of going up and down stairs ; The intensity coefficient of the muscle electrical signal ranges from 0.1 to 0.3. The intensity of the electrical signal in human muscles is expressed in mV. The greater the intensity of the electrical signal, the stronger the body's willingness to exert force, and the greater the intensity of the active stimulation current, which helps the body exert force better.

[0180] The expression for the pulse parameters of the passively stimulated muscle is:

[0181] ;

[0182] In the formula, I p Indicates the intensity of the passive stimulation current; I p0 Indicates the baseline current intensity of passive stimulation; k θ Indicates the joint angle deviation coefficient; θ This indicates the deviation between the actual joint angle and the target joint angle; T p This indicates the passive stimulation pulse period, in milliseconds (ms). T p0 This represents the basic pulse period of passive stimulation, with a value range of 80-150ms; k v This represents the velocity coefficient, with a value ranging from 10 to 20. v This indicates the movement speed of the exoskeleton robot, measured in m / s. The faster the movement speed, the longer the pulse period, to avoid overstimulating the muscles.

[0183] Furthermore, passive stimulation is used to induce passive movement in the human body and is suitable for scenarios such as rehabilitation training. Its pulse parameters are related to joint angle deviation and exoskeleton robot movement speed, and vary depending on the scenario. Different values, for example, in rehabilitation training scenarios. This is the joint angle deviation coefficient, with a value range of 0.2-0.5; The deviation between the actual joint angle and the target joint angle is expressed in degrees. The greater the deviation of the joint angle, the greater the intensity of the passive stimulation current, which in turn causes the joint to move toward the target angle.

[0184] It should be further explained that, based on the generated power control commands, combined with the collaborative control algorithm for synchronous muscle stimulation and the target trajectory and torque mapping function, the pulse width is adjusted, and the pulse parameters for both actively and passively stimulated muscles are output synchronously to enable human-machine collaborative control in multiple scenarios, specifically including:

[0185] I. Synergistic control strategy for synchronous stimulation of muscle groups:

[0186] 1. Construction of Muscle Group Synergistic Mapping Relationship: Based on human exercise physiology, a multi-muscle group functional synergistic matrix is ​​established to clarify the combination of core and auxiliary muscle groups under different movement trajectories; during hip flexion, the core muscle group is the iliopsoas (stimulation weight 0.6), and the auxiliary muscle groups are the rectus femoris (0.3) and tensor fasciae latae (0.1); during knee extension, the core muscle group is the quadriceps femoris (0.7), and the auxiliary muscle groups are the gluteus maximus (0.2) and gastrocnemius (0.1); through electromyographic signal coherence analysis (calculating the cross-correlation coefficient between muscle groups), the time offset of synchronous stimulation (≤10ms) is determined to avoid muscle group contraction timing conflicts.

[0187] 2. Multi-channel pulse synchronous triggering mechanism: A distributed pulse stimulation array is designed, containing 8 independent output channels (each channel supports 0-100mA current adjustment and pulse width adjustable from 50-500μs). Time diversity technology is used to achieve synchronous triggering. The main controller sends a synchronous clock signal (1kHz), and each channel starts pulse output simultaneously on the rising edge of the clock. Inter-channel phase compensation (±2μs) is introduced to offset the synchronization error caused by cable transmission delay. For muscle groups at risk of spasticity (hamstrings in stroke patients), inhibitory pre-stimulation is set: a low-intensity pulse (10-15mA) is output 50ms before the main stimulation to reduce muscle spindle excitability.

[0188] 3. Muscle group balance regulation of stimulus intensity: Define a synergistic stimulation coefficient ξ (value 0-1) to characterize the ratio of stimulus intensity between accessory and core muscle groups: ξ=I 辅助 / I 核心 The dynamic range is 0.2-0.8.

[0189] For example, when walking on flat ground, the ξ of the hip flexor and extensor muscles is 0.5 (balanced force exertion); when going up and down stairs, the ξ of the knee extensor muscles is 0.8 (strengthening the coordination of auxiliary muscle groups).

[0190] II. Dynamic adjustment mechanism based on joint torque data:

[0191] 1. Human-machine torque fusion model: Establish a target trajectory-torque mapping function to map the desired motion trajectory (such as the joint angle curve within the gait cycle). Converted into target torque Where K p K d The proportional-differential coefficients, The actual angle and angular velocity (acquired by the joint encoder).

[0192] Calculate the human-machine torque distribution ratio: Let the mechanical assist torque of the exoskeleton be T. m (t), the electrical stimulation-induced torque is T e (t), satisfying Define the allocation coefficient .

[0193] 2. Linkage adjustment of pulse intensity and mechanical assistance: Design a dual closed-loop adjustment system;

[0194] Outer ring: Compare actual torque T actual (t) and target torque T d (t), calculate torque error Inner ring: according to Adjust the α value and pulse parameters. When ΔT>0 (insufficient torque), if α<0.7, increase the pulse intensity (ΔI=k_p·ΔT), otherwise increase the mechanical assist. When ΔT<0 (excess torque): prioritize reducing the pulse intensity (the rate is twice that of the mechanical assist).

[0195] Pulse intensity adjustment formula: I(t) = I0(t) · [1 + β · ΔT / T] d [(t)], where β is the adjustment gain (0.5-2.0), and I0(t) is the basic intensity.

[0196] 3. Look-ahead regulation based on torque change rate: Introduce a torque change rate threshold γ (e.g., 5 N·m / s). When dT_actual(t) / dt>γ, look-ahead regulation is initiated. The torque change trend in the next 100ms is predicted through an LSTM network (the input is the torque sequence of the past 500ms). The pulse width is adjusted in advance (ΔW=0.1·ΔT_predicted) to achieve a "prediction-compensation" closed loop.

[0197] III. Implementation of Human-Machine Collaborative Trajectory Tracking: In the original pulse stimulation control algorithm, a torque distribution decision layer is added. Execution steps: Receive the target trajectory θ planned by the upper layer. d (t); Collect joint torque sensor data T actual (t) and electromyographic signal (EMG(t)); calculate the target torque T d(t) and allocation coefficient α; output mechanical assistance command (adjust servo motor current) and pulse stimulation parameters (intensity, width, frequency); set differentiated control parameters for different motion phases (support phase / swing phase): support phase α=0.3 (mechanical assistance dominant), swing phase α=0.7 (electric stimulation dominant).

[0198] Tracking accuracy improved, with angle error reduced from ±5° to ±1.5° and torque tracking error reduced from ±8% to ±3%; energy consumption optimized: when the proportion of electrically induced torque increased to 50%, the exoskeleton battery life increased by 40%; human-machine collaboration smoothness: the subject's subjective score (1-10 points) improved from 6.2 points to 8.9 points, with no obvious lag. A torque upper limit protection was set: when T... actual (t)>1.2·T d When (t), immediately cut off the pulse output and reduce mechanical assistance (to 30% within 100ms); abnormal EMG signal detection: if the EMG signal amplitude exceeds the threshold (5 times that of the resting state), it is determined to be muscle spasm, and the reverse stimulation is automatically started (pulse frequency 100Hz, lasting 200ms).

[0199] Furthermore, active and passive stimulation of the muscles can be achieved based on the generated pulse parameters.

[0200] like Figure 4 As shown, in a specific embodiment, taking the scenario of going up and down stairs as an example, the working process of the present invention is as follows:

[0201] 1. Data Acquisition and Preprocessing: The perception module uses LiDAR, a visual camera, and an electromyography (EMG) sensor to collect environmental information and human motion data. The LiDAR acquires the height of the stair steps (approximately 15cm), the distance between steps (approximately 30cm), and the distribution of surrounding obstacles. The visual camera captures images of the stair area and identifies the edge contours of the steps. The EMG sensor, attached to the quadriceps and hamstring muscles, acquires a muscle electrical signal intensity of 25mV. The pressure sensor detects a foot contact pressure of 300N with the step. The inertial measurement unit records the leg acceleration as 1.2m / s². The control module uses a Kalman filter algorithm to preprocess the data and sets an initial state estimate for the acceleration data. Process noise covariance Measure noise covariance State transition matrix Control input matrix Measurement matrix After three iterations of filtering, a smooth acceleration data of 1.18 m / s² was obtained, removing random noise of ±0.3 m / s².

[0202] 2. Scene Recognition: The preprocessed multi-source data is input into a CNN-RNN hybrid model. The CNN part uses a two-layer convolutional structure: Convolutional Layer 1 uses 3×3 convolutional kernels (32 kernels) with a stride of 1 to extract features from a 64×64 environmental image, outputting a feature map of size 62×62×32; Pooling Layer 1 uses 2×2 max pooling, outputting 31×31×32; Convolutional Layer 2 uses 3×3 convolutional kernels (64 kernels) with a stride of 1, outputting 29×29×64; Pooling Layer 2 uses 2×2 max pooling, outputting 14×14×64, which is flattened to obtain a 12544-dimensional feature vector; The RNN part uses an LSTM network (128 hidden units) to process 50 consecutive frames of temporal data (sampling frequency 10Hz), and the hidden layer state update formula is as follows: ,in It is a 32×128 weight matrix. It is a 128×128 weight matrix. The bias vector is 128-dimensional; finally, the scene probability is output through a fully connected layer and a softmax function: going up and down stairs (92%), walking on flat ground (5%), climbing a slope (3%), and the current scene is determined to be going up and down stairs.

[0203] 3. Motion control strategy generation:

[0204] Joint angle control: Calls preset parameters for the stair-climbing scene, including basic joint angles. (Knee flexion angle), acceleration coefficient Force coefficient According to the formula The target knee joint angle of 42° is generated.

[0205] Motion speed control: Base speed Deceleration coefficient The lidar measured the distance to the previous step. According to the formula The movement speed is limited to 0.25 m / s.

[0206] Assist coefficient calculation: Basic assist coefficient Adjustment coefficient Substituting the muscle electrical signal intensity of 25mV, we get That is, it outputs 45% of the auxiliary driving force.

[0207] 4. Calculation of pulse stimulation parameters:

[0208] Active stimulation: Targeting the quadriceps (active knee extensor muscles), setting a baseline current. muscle electrical signal coefficient Distance from the previous step Basic cycle Distance coefficient ; calculated .

[0209] Passive stimulation: targeting the hamstrings (passive knee flexor muscles), the deviation between the actual joint angle of 38° and the target angle of 42°. Base current Angular deviation coefficient Speed ​​of movement Basic cycle speed coefficient ; calculated (Frequency approximately 9.6 Hz).

[0210] 5. Command Execution and Feedback: After receiving the control command, the power module drives the knee joint to rotate through the reducer: rotating from the initial angle of 20° to 42° within 0.8 seconds, maintaining an angular velocity of 27.5° / s, and dynamically adjusting the driving torque according to the assist coefficient (peak value 8 N·m); the pulse stimulation module outputs current through the electrode pads: active stimulation continuously outputs a 20mA pulse signal with a period of 75.2ms during the knee extension phase (0-0.4s), and passive stimulation outputs a 9.2mA pulse signal with a period of 103.75ms during the knee flexion phase (0.4-0.8s); the control module receives the status feedback from the power module (actual joint angle error <1°) and the secondary sampling data from the sensing module in real time, updating the control parameters every 100ms to ensure that the step frequency is stable at 1.2 steps / second during the process of going up and down stairs, and the overlap between the knee joint movement trajectory and the preset curve is >95%.

[0211] When the scene switches to walking on flat ground, the above process repeats automatically: the lidar detects that the ground flatness error is <5mm, the scene recognition model outputs a flat ground probability of 97%, and the control module calls the flat ground parameters (joint base angle 15°, base speed 0.6m / s, active stimulation base current 10mA) to achieve seamless switching.

[0212] In addition, to ensure the safety and effectiveness of neuromuscular electrical stimulation, the intensity of the active stimulation current is set. The upper limit is set at 50mA to avoid muscle damage, and the lower limit is set at 5mA to ensure effective stimulation; pulse cycle The range is set to 20-200ms, corresponding to a frequency of 5-50Hz, covering the frequency band of human muscle response. The above parameter boundaries take into account the physiological response characteristics of muscles and safety requirements, and are set with reference to the national standard "Safety Requirements for Muscle Stimulators in Medical Electrical Equipment" (GB 9706.15-2008) to comply with the safety specifications for medical electrical equipment.

[0213] Different muscle groups require different stimulation parameters due to their different physiological characteristics. For example, in upper limb exoskeleton applications, the biceps brachii, which actively flexes the elbow, requires active stimulation with a baseline current. Set at 12mA, and for the quadriceps femoris muscle responsible for active knee extension in the lower limbs, Setting it to 15mA reflects the adaptability of the stimulation parameters to the physiological characteristics of the muscle group.

[0214] Figure 2 An embodiment of an adaptive multi-scenario exoskeleton robot control system of the present invention is shown.

[0215] In this optional embodiment, the adaptive multi-scenario exoskeleton robot control system includes:

[0216] The scene recognition module 201 is used to denoise the acquired multimodal sensor data using a Kalman filter algorithm based on scene feature entropy, and to fuse the denoised multimodal sensor data. It then combines a hybrid deep learning model to extract spatial features and time series features to identify the current scene type.

[0217] The collaborative control module 202 is used to construct an emotional state and motion intention mapping model based on the current scene type recognition result, generate corresponding power control commands, and synchronously output pulse parameters for active and passive muscle stimulation based on the power control commands, so as to carry out human-machine collaborative control in multiple scenarios.

[0218] It should be noted that the scene recognition module 201 includes a data denoising unit (using scene feature entropy adaptive Kalman filtering), a data fusion unit (biomechanical constraint reinforcement learning + spatiotemporal attention), and a feature recognition unit (CNN-RNN hybrid model), used to output scene type recognition results; the collaborative control module 202 includes an emotion-intention mapping unit (physiological signal acquisition + emotion-intention model), a control command generation unit (joint / velocity / assistance coefficient calculation), and a pulse stimulation unit (muscle group coordination + torque mapping). The emotion-intention mapping unit receives the output results of the scene recognition module, and the pulse stimulation unit outputs active and passive stimulation parameters to realize multi-scene human-machine collaborative control.

[0219] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0220] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0221] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0222] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0223] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0224] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. An adaptive multi-scenario exoskeleton robot control method, characterized in that, The method includes: The Kalman filter algorithm based on scene feature entropy is used to denoise the collected multimodal sensor data, and the denoised multimodal sensor data is fused. Spatial features and time series features are extracted by combining a hybrid deep learning model to identify the current scene type. Based on the recognition results of the current scene type, and combined with the motion control strategy of emotion state recognition, an emotion state-motor intention mapping model is constructed. Corresponding motion control commands are then generated based on this model, specifically including: Based on the recognition results of the current scene type, the user's physiological signals are collected, and emotional features are extracted from the physiological signals. The physiological signals include heart rate variability, skin conductance response, and speech fundamental frequency. The emotional features include the standard deviation of heart rate variability, peak amplitude of skin conductance, and fluctuation range of speech fundamental frequency. The extracted emotional features are input into a hybrid deep learning model, and the weights of the emotional features are optimized through an attention mechanism. The user's emotional state and corresponding quantitative value are output. Based on the quantitative value of the emotional state, an emotional influence factor is defined, and a mapping model between emotional state and movement intention is constructed. Combining the emotional state and the emotional state-movement intention mapping model, the control parameters of the controller are dynamically adjusted to generate corresponding power control commands. The power control commands include joint angle control commands, movement speed control commands, and assist coefficient calculation. Based on the generated power control commands, combined with the collaborative control algorithm of synchronous muscle stimulation and the target trajectory and torque mapping function, the pulse width is adjusted, and the pulse parameters of active and passive muscle stimulation are output synchronously to perform human-machine collaborative control in multiple scenarios.

2. The adaptive multi-scenario exoskeleton robot control method according to claim 1, characterized in that, The process involves using a Kalman filter algorithm based on scene feature entropy to denoise the acquired multimodal sensor data, fusing the denoised multimodal sensor data, and combining it with a hybrid deep learning model to extract spatial and temporal features to identify the current scene type, including: Multimodal sensor data is collected, and the Kalman filter algorithm based on scene feature entropy is used to denoise the multimodal sensor data. Then, the denoised multimodal sensor data is fused by combining the reinforcement learning decision mechanism based on biomechanical constraints and the spatiotemporal fusion algorithm. The fused multimodal sensor data is input into a hybrid deep learning model composed of convolutional neural networks and recurrent neural networks to extract spatial features and time series features respectively. Spatial and temporal features extracted by a fully connected layer are fused together, and the probability distribution of various scenarios is calculated by a normalized exponential function to identify the current scenario type.

3. The adaptive multi-scenario exoskeleton robot control method according to claim 2, characterized in that, The process of acquiring multimodal sensor data, denoising the multimodal sensor data using a Kalman filter algorithm based on scene feature entropy, and fusing the denoised multimodal sensor data by combining a reinforcement learning decision-making mechanism based on biomechanical constraints with a spatiotemporal fusion algorithm includes: Multimodal sensor data containing environmental information and human motion data are collected, and the initial noise covariance matrix is ​​set using the Kalman filter algorithm to perform preliminary denoising on the multimodal sensor data; Based on the collected multimodal sensor data, the scene feature entropy is calculated in real time, and the initial noise covariance matrix is ​​dynamically adjusted according to the calculation results to denoise the multimodal sensor data. Based on the dynamically adjusted initial noise covariance matrix, the biomechanical constraints are encoded into a penalty function using a reinforcement learning algorithm, and a reinforcement learning framework is constructed by combining it with a dual-delay deep deterministic policy gradient algorithm. Based on scene feature entropy, a reinforcement learning framework is trained using a meta-learning algorithm, and combined with a spatiotemporal attention mechanism, the weights of multimodal sensor data are dynamically adjusted. Based on the adjusted weights of the multimodal sensor data, the denoised multimodal sensor data is fused.

4. The adaptive multi-scenario exoskeleton robot control method according to claim 3, characterized in that, The expression for the probability distribution is: ； In the formula, y i Indicates the first i The probability of each scenario; z i The output of the fully connected layer represents the first... i Scores for each scenario; C Indicates the total number of scene categories; i Indicates the index value; z j The output of the fully connected layer represents the first... j Scores for each scenario.

5. The adaptive multi-scenario exoskeleton robot control method according to claim 1, characterized in that, The expression for the joint angle control command is: ； In the formula, θ cmd Indicates joint angle control commands; θ 0 Indicates the base angle value; k a Indicates the acceleration coefficient; a Represents acceleration data; k f Indicates the force coefficient; f This represents data on human force exertion; The expression for the motion speed control command is: ； In the formula, v cmd This indicates a motion speed control command; v 0 Indicates the base speed; α Indicates the deceleration coefficient; d Indicates the distance to the target; The expression for the assist coefficient is: ； In the formula, λ Indicates the assist coefficient; λ 0 Indicates the basic assist coefficient; β Indicates the adjustment factor; s This indicates the intensity of electrical signals in human muscles.

6. The adaptive multi-scenario exoskeleton robot control method according to claim 5, characterized in that, The generated power control commands, combined with a collaborative control algorithm for synchronized muscle stimulation and a target trajectory-torque mapping function, adjust the pulse width and synchronously output the pulse parameters for both actively and passively stimulated muscles to enable human-machine collaborative control in multiple scenarios, including: Based on the generated dynamic control commands and human movement physiology, a multi-muscle group functional coordination matrix is ​​established, and the correlation coefficient between muscle groups is calculated through electromyographic signal coherence analysis to determine the time offset of synchronous stimulation. Based on the determined time offset, a distributed pulse stimulation array is designed, and combined with a predefined synergistic stimulation coefficient, the stimulation intensity ratio between auxiliary muscle groups and core muscle groups is dynamically adjusted. Based on the adjusted stimulation intensity ratio, a target trajectory and torque mapping function is established to convert the motion trajectory into target torque and calculate the human-machine torque distribution ratio to coordinate the distribution of pulse intensity and exoskeleton mechanical assistance. By combining a preset torque change rate threshold and using the torque change trend predicted by a long short-term memory network, the pulse width is adjusted, and the pulse parameters for active and passive muscle stimulation are output simultaneously to enable human-machine collaborative control in multiple scenarios.

7. The adaptive multi-scenario exoskeleton robot control method according to claim 6, characterized in that, The expression for the pulse width is: ； In the formula, W a Indicates the pulse width; W a0 Indicates the basic pulse width; k t Indicates the torque error coefficient; This indicates the deviation between the actual torque and the target torque. The expression for the pulse parameters of the actively stimulated muscle is: ； In the formula, I a Indicates the intensity of the active stimulation current; I a0 Indicates the intensity of the active stimulation baseline current; k s Indicates the muscle electrical signal intensity coefficient; s Indicates the intensity of electrical signals in human muscles; T a Indicates the active stimulation pulse period; T a0 This indicates the basic pulse period of active stimulation; k d Indicates the distance coefficient; d Indicates the distance to the target; The expression for the pulse parameters of the passively stimulated muscle is: ； In the formula, I p Indicates the intensity of the passive stimulation current; I p0 Indicates the baseline current intensity of passive stimulation; k θ Indicates the joint angle deviation coefficient; θ This indicates the deviation between the actual joint angle and the target joint angle; T p Indicates the passive stimulation pulse period; T p0 This indicates the basic pulse period of passive stimulation; k v Indicates the velocity coefficient; v This indicates the speed of movement of the exoskeleton robot.

8. An adaptive multi-scenario exoskeleton robot control system, characterized in that, The system includes: The scene recognition module is used to denoise the acquired multimodal sensor data using a Kalman filter algorithm based on scene feature entropy, and to fuse the denoised multimodal sensor data. It then combines a hybrid deep learning model to extract spatial and temporal features in order to identify the current scene type. The collaborative control module is used to construct an emotion state-motor intention mapping model based on the recognition results of the current scene type and the motion control strategy of emotion state recognition, and to generate corresponding dynamic control commands based on the emotion state-motor intention mapping model, specifically including: Based on the recognition results of the current scene type, the user's physiological signals are collected, and emotional features are extracted from the physiological signals. The physiological signals include heart rate variability, skin conductance response, and speech fundamental frequency. The emotional features include the standard deviation of heart rate variability, peak amplitude of skin conductance, and fluctuation range of speech fundamental frequency. The extracted emotional features are input into a hybrid deep learning model, and the weights of the emotional features are optimized through an attention mechanism. The user's emotional state and corresponding quantitative value are output. Based on the quantitative value of the emotional state, an emotional influence factor is defined, and a mapping model between emotional state and movement intention is constructed. Combining the emotional state and the emotional state-movement intention mapping model, the control parameters of the controller are dynamically adjusted to generate corresponding power control commands. The power control commands include joint angle control commands, movement speed control commands, and assist coefficient calculation. Based on the generated power control commands, combined with the collaborative control algorithm of synchronous muscle stimulation and the target trajectory and torque mapping function, the pulse width is adjusted, and the pulse parameters of active and passive muscle stimulation are output synchronously to perform human-machine collaborative control in multiple scenarios.

Images