Self-adaptive multi-scene exoskeleton robot control method and system

Abstract

The invention belongs to the technical field of robots, and discloses a self-adaptive multi-scene exoskeleton robot control method and system, and the method comprises the steps: carrying out the denoising of collected multi-modal sensor data through employing a scene feature entropy-based Kalman filtering algorithm, carrying out the fusion of the denoised multi-modal sensor data, and carrying out the fusion of the denoised multi-modal sensor data, extracting spatial features and time sequence features in combination with a hybrid deep learning model to identify a current scene type; and according to a current scene type identification result, constructing an emotional state and motion intention mapping model, generating a corresponding power control instruction, and synchronously outputting pulse parameters of active stimulation muscles and passive stimulation muscles based on the power control instruction so as to perform man-machine cooperative control under multiple scenes. The method can adapt to different types of exoskeleton robots, and a more intelligent and accurate solution is provided for multiple application scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of robotics, in particular to a self-adaptive multi-scenario exoskeleton robot control method and system. BACKGROUND

[0002] As a device capable of enhancing human movement ability or assisting human movement, exoskeleton robots are widely used in medical rehabilitation, industrial production, military and other fields. However, the control algorithm of existing exoskeleton robots is mostly designed for specific scenarios, and manual intervention is usually required to adjust parameters when the scene is switched, which leads to poor adaptability and insufficient control accuracy. Especially in complex or changeable environments, the robot cannot quickly adjust according to the needs of different scenes, thereby affecting its overall performance and application effect.

[0003] In addition, most existing exoskeleton robots only provide mechanical assistance, lacking active and passive stimulation functions for muscles, especially in the field of rehabilitation assistance, failing to fully realize the potential of promoting muscle recovery and functional reconstruction. Therefore, the existing technology still has obvious deficiencies in improving the adaptability of exoskeleton robots to different use scenarios, improving control accuracy, and enhancing muscle stimulation function.

[0004] Currently, simple filtering algorithms are mostly used in data processing, which cannot effectively remove noise interference in complex environments. Scene recognition usually relies on single sensor data, resulting in low recognition accuracy and inability to dynamically adjust according to different environments and human states. In addition, the existing motion control strategy lacks flexible algorithm support and cannot adjust in real time according to the motion state and needs of the human body in different scenarios, thereby limiting the universality and utility of exoskeleton robots in different application scenarios.

[0005] The control method for muscle stimulation is also relatively simple, lacking precise pulse stimulation parameter calculation and dynamic adjustment mechanism, making it difficult to achieve precise regulation of muscles and optimize rehabilitation effect. These technical problems in practical application lead to poor universality of exoskeleton robots, limiting their further development and promotion.

[0006] Therefore, how to provide a self-adaptive multi-scenario exoskeleton robot control method and system is a problem to be solved at present. SUMMARY

[0007] The embodiments of the present application provide a self-adaptive multi-scenario exoskeleton robot control method and system to solve the above technical problems existing in the prior art.

[0008] The following presents a simplified summary of some aspects of the disclosed embodiments in order to provide a basic understanding of such embodiments. This summary is not an extensive overview of the embodiments and is intended neither to identify key / critical elements of the embodiments nor to delineate the scope of the embodiments. Its sole purpose is to present some concepts of the embodiments in a simplified form as a prelude to the more detailed description that is presented later.

[0009] According to a first aspect of embodiments of the present application, there is provided a self-adaptive multi-scene exoskeleton robot control method.

[0010] In an embodiment, a self-adaptive multi-scene exoskeleton robot control method comprises: Denoising the collected multi-modal sensor data by using a Kalman filtering algorithm based on scene feature entropy, and fusing the denoised multi-modal sensor data, combining a hybrid deep learning model to extract spatial features and time series features, to identify the current scene type; According to the current scene type identification result, a sentiment state and motion intention mapping model is constructed, corresponding power control instructions are generated, and the pulse parameters of the active stimulating muscles and the passive stimulating muscles are synchronously output based on the power control instructions, to perform human-machine collaborative control in multiple scenes.

[0011] In an embodiment, the denoising the collected multi-modal sensor data by using a Kalman filtering algorithm based on scene feature entropy, and fusing the denoised multi-modal sensor data, combining a hybrid deep learning model to extract spatial features and time series features, to identify the current scene type comprises: Collecting multi-modal sensor data, denoising the multi-modal sensor data by using a Kalman filtering algorithm based on scene feature entropy, and fusing the denoised multi-modal sensor data by combining a biomechanics-constrained reinforcement learning decision mechanism and a space-time fusion algorithm; Inputting the fused multi-modal sensor data into a hybrid deep learning model composed of a convolutional neural network and a recurrent neural network to extract spatial features and time series features respectively; Fusing the extracted spatial features and time series features through a full connection layer, and calculating the probability distribution of each type of scene through a normalized exponential function to identify the current scene type.

[0012] In an embodiment, the collecting multi-modal sensor data, denoising the multi-modal sensor data by using a Kalman filtering algorithm based on scene feature entropy, and fusing the denoised multi-modal sensor data by combining a biomechanics-constrained reinforcement learning decision mechanism and a space-time fusion algorithm comprises: Collecting multi-modal sensor data containing environmental information and human motion data, and setting an initial noise covariance matrix by using a Kalman filtering algorithm to preliminarily denoise the multi-modal 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, a reinforcement learning algorithm is used to encode biomechanical constraints into a penalty function. This is combined with a double-delayed deep deterministic policy gradient algorithm to construct a reinforcement learning framework. Based on the scene feature entropy, a meta-learning algorithm is used to train a reinforcement learning framework, and combined with a spatiotemporal attention mechanism, the weights of multimodal sensor data are dynamically adjusted; The denoised multimodal sensor data is fused according to the adjusted multimodal sensor data weights.

[0013] In one embodiment, the probability distribution is expressed as: ; Where, y i Indicates the i The probability of a scenario; z i represents the output of the fully connected layer i The score of each scene; C Indicates the total number of scene categories; i Represents the index value; z j represents the output of the fully connected layer j The score of the scene.

[0014] In one embodiment, the method of constructing an emotion state and motion intention mapping model based on the current scene type recognition result, generating corresponding power control instructions, and synchronously outputting pulse parameters for active muscle stimulation and passive muscle stimulation based on the power control instructions to perform human-machine collaborative control in multiple scenarios includes: Based on the recognition results of the current scene type and the motion control strategy of the emotional state recognition, a mapping model of the emotional state and motion intention is constructed, and corresponding power control instructions are generated based on the mapping model of the emotional state and motion intention; Based on the generated power control instructions, combined with the collaborative control algorithm of synchronous muscle group stimulation and the target trajectory and torque mapping function, the pulse width is adjusted, and the pulse parameters of the active and passive muscle stimulation are output synchronously to achieve human-machine collaborative control in multiple scenarios.

[0015] In one embodiment, the method of constructing an emotional state and motion intention mapping model based on the recognition result of the current scene type and combining it with the motion control strategy of the emotional state recognition, and generating corresponding power control instructions based on the emotional state and motion intention mapping model includes: According to the recognition result of the current scene type, physiological signals of the user are collected, and emotional features are extracted from the physiological signals; The extracted emotional features are input into a hybrid deep learning model, the weight of the emotional features is optimized through an attention mechanism, and the emotional state of the user and the corresponding quantitative value are output; Based on the quantitative value of the emotional state, an emotional influence factor is defined, and an emotional state and motion intention mapping model is constructed; Combined with the emotional state and the emotional state and motion intention mapping model, the control parameters of the controller are dynamically adjusted, and corresponding power control instructions are generated, including joint angle control instructions, motion speed control instructions, and assistance coefficient calculation.

[0016] In one embodiment, the expression of the joint angle control instruction is: ; In the formula, θ cmd represents the joint angle control instruction; θ 0 represents the basic angle value; k a represents the acceleration coefficient; a represents the acceleration data; k f represents the force coefficient; f represents the human force data; The expression of the motion speed control instruction is: ; In the formula, v cmd represents the motion speed control instruction; v 0 represents the basic speed; α represents the deceleration coefficient; d represents the distance to the target; The expression of the assistance coefficient is: ; In the formula, λ represents the assistance coefficient; λ 0 represents the basic assistance coefficient; β represents the adjustment coefficient; s represents the human muscle electrical signal intensity.

[0017] In one embodiment, based on the generated power control instructions, the synergistic control algorithm of muscle group synchronous stimulation, the target trajectory, and the torque mapping function are combined to adjust the pulse width and synchronously output the pulse parameters of the active stimulation muscles and the passive stimulation muscles to perform human-machine collaborative control in multiple scenes. According to the generated power control instruction and human body movement physiology, a multi-muscle group function synergy matrix is established, and a correlation coefficient between muscle groups is calculated through muscle electromyogram coherence analysis to determine a time offset of synchronous stimulation; According to the determined time offset, a distributed pulse stimulation array is designed, and a pre-defined synergy stimulation coefficient is combined to dynamically adjust a stimulation intensity ratio of an auxiliary muscle group and a core muscle group; According to the adjusted stimulation intensity ratio, a target trajectory and torque mapping function is established to convert a movement trajectory into a target torque, and a man-machine torque distribution ratio is calculated to cooperatively distribute pulse intensity and exoskeleton mechanical assistance; In combination with a preset torque change rate threshold, a torque change trend predicted by a long short-term memory network is used to adjust a pulse width, and pulse parameters of an active stimulation muscle and a passive stimulation muscle are synchronously output to perform man-machine cooperative control in multiple scenes.

[0018] In one embodiment, the expression of the pulse width is: ; In the expression, W a represents the pulse width; W a0 represents a basic pulse width; k t represents a torque error coefficient; represents a deviation between an actual torque and a target torque; The expression of the pulse parameter of the active stimulation muscle is: ; In the expression, I a represents an active stimulation current intensity; I a0 represents an active stimulation basic current intensity; k s represents a muscle electromyogram intensity coefficient; s represents a human muscle electromyogram intensity; T a represents an active stimulation pulse period; T a0 represents an active stimulation basic pulse period; k d represents a distance coefficient; d represents a distance to a target; The expression of the pulse parameter of the passive stimulation muscle is: ; In the expression, I prepresents passive stimulation current intensity; I p0 represents passive stimulation base current intensity; k θ represents joint angle deviation coefficient; θ represents the deviation of the actual joint angle from the target joint angle; T p represents passive stimulation pulse period; T p0 represents passive stimulation base pulse period; k v represents motion speed coefficient; v represents exoskeleton robot motion speed.

[0019] According to a second aspect of the embodiments of the present application, there is provided an adaptive multi-scene exoskeleton robot control system.

[0020] In one embodiment, the adaptive multi-scene exoskeleton robot control system comprises: a scene recognition module configured to denoise the collected multi-modal sensor data using a Kalman filtering algorithm based on scene feature entropy, fuse the denoised multi-modal sensor data, and extract spatial features and time sequence features in combination with a hybrid deep learning model to identify a current scene type; a cooperative control module configured to construct an emotion state and motion intention mapping model according to the current scene type identification result, generate corresponding power control instructions, and synchronously output pulse parameters of the active stimulation muscle and the passive stimulation muscle based on the power control instructions to perform human-machine cooperative control in multiple scenes.

[0021] According to a third aspect of the embodiments of the present application, there is provided a computer device.

[0022] In some embodiments, the computer device comprises a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0023] According to a fourth aspect of the embodiments of the present application, there is provided a computer readable storage medium.

[0024] In one embodiment, the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the above method.

[0025] The technical solutions provided by the embodiments of the present application can include the following beneficial effects: 1、The application can remove noise and improve data quality by introducing Kalman filtering algorithm, thereby providing stable and reliable data basis for subsequent processing; at the same time, the application of the hybrid deep learning model improves the accuracy and robustness of scene recognition, so that the system can better adapt to various changes in dynamic environment.

[0026] 2、The application can adjust control parameters in real time according to different scenes and human states, so that the exoskeleton robot has self-adaptive ability, thereby improving the safety and comfort of movement; in addition, by accurately controlling the active stimulation and passive stimulation of muscles, the active stimulation helps the human body to actively exert force, and the passive stimulation assists the human body to complete passive movement, greatly enhancing the effect of the exoskeleton robot in rehabilitation assistance and other scenes, and being suitable for different types of exoskeleton robots, providing a more intelligent and accurate solution for various application scenarios.

[0027] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the application. BRIEF DESCRIPTION OF DRAWINGS

[0028] The drawings incorporated into the specification and forming part of the specification, show embodiments consistent with the application, and together with the specification, serve to explain the principles of the application.

[0029] Figure 1 is a flow chart of an adaptive multi-scene exoskeleton robot control method according to an exemplary embodiment; Figure 2 is a principle block diagram of an adaptive multi-scene exoskeleton robot control system according to an exemplary embodiment; Figure 3 is a structural schematic diagram of a computer device according to an exemplary embodiment; Figure 4 is a whole framework diagram of an adaptive multi-scene exoskeleton robot control method according to an exemplary embodiment; Figure 5 is one of the scene recognition hybrid model structure schematic diagrams of an adaptive multi-scene exoskeleton robot control method according to an exemplary embodiment; Figure 6 is the second scene recognition hybrid model structure schematic diagram of an adaptive multi-scene exoskeleton robot control method according to an exemplary embodiment. DETAILED DESCRIPTION

[0030] It should be understood that although the steps in the flowchart are shown in sequential order, such that each step depends on completion of the previous step before execution of the next step, the steps are not necessarily performed in the order shown by the arrows. Unless specifically stated in this document, the steps are not necessarily performed in a strict sequence, and the steps can be performed in other sequences. Also, at least some of the steps in the figure can include multiple sub-steps or multiple stages, which are not necessarily performed at the same time, but can be performed at different times, and which are not necessarily performed sequentially, but can be performed in rotation or alternation with at least some of the other steps or sub-steps or stages of other steps.

[0031] Each module in the device or system of the present application can be implemented wholly or partially by software, hardware, and combinations thereof. The modules can be embedded in or independent of a processor in a computer device in hardware form, or stored in a memory in the computer device in software form, so as to be invoked and executed by a processor to perform operations corresponding to the modules.

[0032] The embodiments in the present application and the features in the embodiments can be combined with each other in the case of no conflict.

[0033] Figure 1 An embodiment of an adaptive multi-scene exoskeleton robot control method of the present application is shown.

[0034] In this optional embodiment, the adaptive multi-scene exoskeleton robot control method comprises: In step S101, a Kalman filtering algorithm based on scene feature entropy is used to denoise the collected multi-modal sensor data, and the denoised multi-modal sensor data is fused, combined with a hybrid deep learning model to extract spatial features and time sequence features, to identify the current scene type.

[0035] In this optional embodiment, the use of a Kalman filtering algorithm based on scene feature entropy to denoise the collected multi-modal sensor data, and the fusion of the denoised multi-modal sensor data, combined with a hybrid deep learning model to extract spatial features and time sequence features, to identify the current scene type comprises: The multi-modal sensor data is collected, and a Kalman filtering algorithm based on scene feature entropy is used to denoise the multi-modal sensor data, and a reinforcement learning decision mechanism with biomechanical constraints and a spatio-temporal fusion algorithm are combined to fuse the denoised multi-modal sensor data.

[0036] In this optional embodiment, the collecting of 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 in combination with a biomechanically constrained reinforcement learning decision mechanism and a spatiotemporal fusion algorithm include: Collect multimodal sensor data containing environmental information and human motion data, and use the Kalman filter algorithm to set the initial noise covariance matrix to perform preliminary denoising on the multimodal sensor data; It should be noted that different sensor types include: IMU, electromyography sensor, and plantar pressure sensor. For example, the accelerometer of IMU is initially Can be set to , gyroscope initialization Set as , calibrated through multiple no-load tests.

[0037] 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; It should be noted that the expression of scene feature entropy is: ; Where, H represents the scene feature entropy; n Indicates the number of feature categories; i Represents the index value; p i Indicates the i Scenario-like features are calculated through historical data statistics or real-time sliding windows.

[0038] Based on the dynamically adjusted initial noise covariance matrix, a reinforcement learning algorithm is used to encode biomechanical constraints into a penalty function. This is combined with a double-delayed deep deterministic policy gradient algorithm to construct a reinforcement learning framework. Based on the scene feature entropy, a meta-learning algorithm is used to train a reinforcement learning framework, and combined with a spatiotemporal attention mechanism, the weights of multimodal sensor data are dynamically adjusted; It should be noted that the temporal attention weight in the spatiotemporal attention mechanism is allocated based on the gait cycle phase, and the spatial attention weight is allocated based on the sensor signal-to-noise ratio.

[0039] The denoised multimodal sensor data is fused according to the adjusted multimodal sensor data weights.

[0040] It should be noted that the formula of the Kalman filter algorithm is: ; Where, expressk optimal estimation value of the time instant; denotes k is based on k predicted value of the time instant -1; K k denotes the Kalman gain; z k denotes k measurement value of the time instant; H denotes the measurement matrix.

[0041] wherein the prediction equation of the Kalman filtering algorithm is: ; wherein, denotes the predicted value of the current time instant based on the optimal estimation of the previous time instant; A denotes the state transition matrix; denotes the optimal estimation value of the previous time instant, and B denotes the control input matrix; denotes the control input of the previous time instant; denotes the prediction error covariance matrix; Q denotes the process noise covariance matrix.

[0042] the update equation of the Kalman filtering algorithm is: ; wherein, K k denotes the Kalman gain; H denotes the measurement matrix; T denotes the transpose; R denotes the measurement noise covariance matrix; z k denotes the measurement value of the current time instant, I denotes the unit matrix; denotes the optimal estimation value of the current time instant; denotes the estimation error covariance matrix of the current time instant. Through the Kalman filtering algorithm, the noise in the data can be effectively removed, the accuracy and stability of the data can be improved, and a good foundation is laid for subsequent scene recognition, control strategy generation and pulse stimulation parameter calculation.

[0043] In addition, multi-modal sensor data is collected, the multi-modal sensor data is denoised by using the Kalman filtering algorithm based on scene feature entropy, and the denoised multi-modal sensor data is fused by combining the reinforcement learning decision mechanism based on biomechanical constraints and the spatio-temporal fusion algorithm. Specifically, the fusion includes: I. Kalman filtering algorithm: 1. Self-adjusting mechanism of dynamic noise covariance matrix: the noise covariance matrix of the conventional Kalman filtering is a fixed value, which is prone to filtering lag in complex scenes (uphill, uneven ground), and is adjusted in real time by introducing scene feature entropy: ; wherein is the feature entropy of the current scene (calculating the information entropy of gait cycle, joint angle rate of change), is the adaptive coefficient (0.1-0.3 dynamic adjustment), is the initial noise matrix. When a ground mutation (from flat ground to stairs) is detected, sudden increase makes expand 3-5 times, improve the response speed of filter to sudden signal; in smooth scene tends to 0, shrink to reduce noise interference.

[0044] 2, multi-modal observation weighted fusion: traditional algorithm only relies on single sensor (myoelectricity), through fusion of IMU attitude data and plantar pressure sensor signal, through credibility weight dynamic distribution of observation weight: , the weight satisfies , and is updated based on the real-time signal-to-noise ratio of each sensor: .

[0045] For example, when going up and down stairs, the plantar pressure signal credibility is high (0.5 up to 0.5), and when walking on flat ground, the myoelectricity signal weight is dominant (0.6 remains 0.6); when the scene switches, 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. II. Reinforcement learning decision mechanism based on biomechanical constraints: Encode the physical constraints of exoskeleton (joint range of motion, maximum driving torque) as a penalty function of reinforcement learning:

[0046] ; wherein is the actual angle of the jth joint, is the safety limit value, is the penalty coefficient; using double-delay deep deterministic policy gradient algorithm, add biomechanical constraint network in Actor-Critic framework, so that the decision process naturally avoids damage risk. For different scenes (loaded / unloaded, flat / slope), train the basic model through meta-learning, and only 50 interactions are needed in new scenes to achieve convergence; input the scene features (slope, load value) as meta-parameters into the policy network, so that the model quickly adapts to parameterized scene changes; when walking on a 30° slope, the energy consumption of the exoskeleton is reduced by 28%, and the gait synchronization error with the human body is reduced from ± 15° to ± 5°, realizing "adaptive efficiency optimization under safety constraints".

[0047] III. Multi-sensor spatio-temporal fusion algorithm based on attention mechanism:

[0048] ​​Designing spatio-temporal attention module: temporal attention, assigning weights to sensor data at different times (support phase weight of gait cycle is higher than swing phase); spatial attention, dynamically highlighting key sensors (prioritize hip joint sensors when going uphill, focus on knee joints when on flat ground); fusion features: wherein, is an attention weight matrix.

[0049] As a pre-processing module of Kalman filter, the key features selected by attention mechanism can improve the convergence speed of filter by 30%; when the signal is interrupted (EMG sensor falls off), the weight of redundant sensors is automatically enhanced, and the fault tolerance time of the system is extended from 200ms to 500ms; the effective information utilization rate of multi-sensor data is improved from 65% to 92%, and the control accuracy can still be maintained above 85% when the sensor is partially disabled. By introducing biomechanical constraints into reinforcement learning and giving the Kalman filter the ability to adapt to scene entropy, the limitations of traditional fields such as "filtering only for noise reduction" and "reinforcement learning only for action optimization" are broken, and the three technical improvements can form a closed loop to fully leverage their respective advantages; by introducing attention mechanism, the data quality is improved, and the improved Kalman filter can better suppress noise, while reinforcement learning can make efficient decisions based on clean data, and the overall performance including scene adaptation speed and control accuracy is improved, ensuring that the exoskeleton robot can adapt to complex and changing use scenarios more efficiently and accurately.

[0050] The fused multi-modal sensor data is input into a hybrid deep learning model composed of convolutional neural network and recurrent neural network to extract spatial features and time series features respectively.

[0051] It should be noted that the hybrid deep learning model training process includes the size of the training data set (500 groups of users, 10 scenes, each group containing 1000 time step data), data enhancement method (time series flip, additive Gaussian noise), loss function (such as cross-entropy loss) and optimizer (Adam, learning rate 0.001, decay coefficient 0.9) to improve model credibility; the parameters of the hybrid deep learning model are shown in Table 1.

[0052] Table 1: Hybrid deep learning model parameter table ; As Figure 5 , Figure 6 shown, the fused multi-modal sensor data is input into a hybrid deep learning model composed of convolutional neural network and recurrent neural network to extract spatial features and time series features, which specifically includes: Convolutional neural network feature extraction: convolutional layer extracts spatial features of data through convolution operation, formula is ;in Output the feature value of the (i, j) position on the feature map of the convolution layer, The ReLU activation function is used, that is , which can solve the gradient vanishing problem and accelerate the convergence of the model. After multiple layers of convolution and pooling operations, the spatial feature vector of the input data is obtained.

[0053] Recurrent neural network time series processing: The spatial feature vector output by the convolutional neural network is input into the recurrent neural network to capture the time series characteristics of the data; the hidden layer state update formula of the RNN is: ;in, is the hidden layer state at time t, is the weight matrix from the input layer to the hidden layer, is the input feature vector at time t, is the weight matrix from hidden layer to hidden layer, is the hidden layer state at time t-1, is the bias term of the hidden layer.

[0054] The spatial features and time series features extracted are fused through the fully connected layer, and the probability distribution of various scenes is calculated through the normalized exponential function to identify the current scene type.

[0055] In this optional embodiment, the probability distribution is expressed as: ; Where, y i Indicates the i The probability of a scenario; z i represents the output of the fully connected layer i The score of each scene; C Indicates the total number of scene categories; i Represents the index value; z j represents the output of the fully connected layer j The score of each scene is calculated; the scene with the highest probability is selected as the recognition result.

[0056] 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.

[0057] Step S102, according to the current scene type recognition result, construct the emotion state and motion intention mapping model, generate the corresponding dynamic control instruction, and based on the dynamic control instruction, synchronously output the pulse parameters of the active stimulation muscle and the passive stimulation muscle, to carry out the human-machine collaborative control in multiple scenes.

[0058] In this optional embodiment, the according to the current scene type recognition result, combined with the emotion state recognition motion control strategy, the emotion state and motion intention mapping model is constructed, and the corresponding dynamic control instruction is generated based on the emotion state and motion intention mapping model. According to the recognition result of the current scene type, combined with the emotion state recognition motion control strategy, the emotion state and motion intention mapping model is constructed, and the corresponding dynamic control instruction is generated based on the emotion state and motion intention mapping model.

[0059] In this optional embodiment, the according to the current scene type recognition result, combined with the emotion state recognition motion control strategy, the emotion state and motion intention mapping model is constructed, and the corresponding dynamic control instruction is generated based on the emotion state and motion intention mapping model. According to the recognition result of the current scene type, the physiological signal of the user is collected, and the emotion features are extracted from the physiological signal. It needs to be supplemented that the physiological signal includes PPG, GSR, voice, etc., and the specific implementation of extracting emotion features is as follows: the standard deviation calculation method of the whole sinus rhythm RR interval of heart rate variability is: , wherein, is the RR interval of the th, is the average RR interval, is the total number of RR intervals; the fundamental frequency (F0) in the voice emotion feature is extracted by the autocorrelation method, the frame length is 20ms, and the frame shift is 10ms.

[0060] The extracted emotion features are input into the mixed deep learning model, the weight of the emotion features is optimized through the attention mechanism, and the emotion state of the user and the corresponding quantization value are output; It needs to be supplemented that, in order to establish the objective calibration standard of the emotion state quantization value, 200 subjects covering different ages, genders and health conditions are recruited to ensure the representativeness of the sample; in the pre-set controllable experimental scene (calm sitting, exercise fatigue, simulated dangerous environment), multi-modal physiological signals (heart rate variability, skin galvanic response, electroencephalogram, etc.) are synchronously collected, and the data are labeled combined with the subjective emotion score of the subjects (using 1-5 Likert scale); based on the above data, the mapping model or quantization control table of the emotion state and physiological signal features is constructed, to improve the explainability and objectivity of the emotion recognition result.

[0061] Based on the quantitative value of the emotional state, an emotional influence factor is defined, and an emotional state and movement intention mapping model is constructed; In combination with the emotional state and the emotional state and movement intention mapping model, the control parameters of the controller are dynamically adjusted to generate corresponding power control instructions, including joint angle control instructions, movement speed control instructions, and assistance coefficient calculation.

[0062] In this optional embodiment, the expression of the joint angle control instruction is: ; In the formula, θ cmd represents the joint angle control instruction; θ 0 represents the basic angle value, which is obtained through experimental calibration; k a represents the acceleration coefficient, which is adjusted according to different scenarios; a represents the acceleration data; k f represents the force coefficient, which is adjusted according to different scenarios; f represents the human force data; The expression of the movement speed control instruction is: ; In the formula, v cmd represents the movement speed control instruction; v 0 represents the basic speed; α represents the deceleration coefficient; d represents the distance to the target, which is calculated through laser radar data, when the distance to the obstacle is close, the speed is automatically reduced to ensure the safety of movement; The expression of the assistance coefficient is: ; In the formula, λ represents the assistance coefficient; λ 0 represents the basic assistance coefficient; β represents the adjustment coefficient; s represents the human muscle electrical signal intensity, the stronger the muscle electrical signal, the smaller the assistance coefficient, realizing human-machine collaborative force.

[0063] It should be noted that, according to the recognition result of the current scene type, in combination with the movement control strategy of emotional state recognition, the emotional state and movement intention mapping model is constructed, and the corresponding power control instruction is generated based on the emotional state and movement intention mapping model, which specifically includes: I. Introduction of emotional state recognition module: 1. Multi-modal emotional feature collection: Based on existing sensor systems, add physiological signal collection units: including wrist PPG heart rate sensor (sampling rate 200Hz), galvanic skin response (GSR) sensor (sampling rate 100Hz), and voice emotion collection microphone (16kHz sampling).

[0064] Extract three types of core emotional features: physiological features, time domain indicators of heart rate variability (HRV) (such as SDNN, RMSSD), peak frequency and amplitude of GSR; voice features, fundamental frequency (F0) fluctuation range, voice energy entropy; motion accompanying features, gait cycle fluctuation rate (≤5% in calm state, rising to 15%-20% in tense state).

[0065] 2. Emotional state classification model: Use CNN-LSTM hybrid model, input layer fusion multi-modal features (dimension 64), through attention mechanism to strengthen key feature weight (HRV weight increases to 0.35 in anxiety state); output is divided into 5 emotional states and corresponding quantitative values: calm (0.1-0.3), focused (0.4-0.6), fatigue (0.7-0.8), anxiety (0.8-0.9), panic (≥0.9), classification accuracy needs to reach more than 89% (through 500 user data training and verification).

[0066] 3. Real-time optimization: Introduce incremental learning mechanism, the model updates the classification boundary once every 10 new data received, avoiding drift caused by individual emotional differences; use model quantization compression technology, control the inference delay within 20ms, meet the real-time control requirements of exoskeleton (control cycle is 50ms).

[0067] II. Construction of emotional state-movement intention mapping model: 1. Mathematical expression of mapping relationship: Define emotional influence factor λ (value range 0.8-1.5), establish mapping function between emotional state and motion assistance force: Where is the basic assistance force without emotional intervention, is the current emotional quantitative value, is the baseline calm value.

[0068] λ values under different emotional states: calm / focused: λ = 0.1 (auxiliary force fluctuation ≤5%); fatigue: λ = -0.2 (auxiliary force increases by 20%-30%, with emphasis on increasing hip joint driving torque); anxiety / panic: λ = 0.3 (auxiliary force decreases by 15%-20%, while increasing damping coefficient to suppress excessive movement).

[0069] 2、Dynamic mapping adjustment mechanism: Construct a sentiment-scene double-factor adjustment matrix; For example, anxiety state + up and down stairs scene: λ up to 0.4, auxiliary force reduced by 25% while joint stiffness increased by 10% to enhance stability; Fatigue state + walking on flat ground: λ down to -0.3, auxiliary force increased by 35%, and step length automatically shortened (from 70 cm to 55 cm) to reduce energy consumption; The mapping relationship is generated by offline training through reinforcement learning, and the reward function is designed as: .

[0070] 3、Individual adaptability optimization: Establish a sentiment-motion preference database for each user, record subjective comfort ratings (1-5) under different emotional states, and update mapping parameters iteratively through Bayesian optimization algorithm; First use through 3-minute guided training, including simulation scenes of calm, fatigue, and concentration, to complete basic parameter calibration.

[0071] Three, dynamic adjustment of motion assistance force: 1、In the original control strategy generation algorithm, a new emotional adjustment layer is added: receive E(t) output by the emotional state recognition module; Call the mapping model to calculate λ value and auxiliary force correction; Dynamically adjust the PID parameters of the joint controller (anxiety state: Kp reduced by 20%, fatigue state: Ti shortened by 30%); Adjustment granularity: hip joint, knee joint, and ankle joint are independently adjusted, for example, knee joint auxiliary force increases by 10% more than hip joint in fatigue state.

[0072] Fatigue state: user walking energy consumption reduced by 32% (verified by metabolic rate monitoring), gait cycle stability improved by 40% (fluctuation rate reduced from 18% to 11%); Anxiety state: unexpected imbalance probability reduced by 65% (tested by pressure platform, imbalance criterion is foot pressure center offset more than 2 cm); Mixed scene (fatigue + uphill / downhill): auxiliary force adjustment response speed is 2 times faster than traditional scheme, avoiding the jerk caused by adjustment lag.

[0073] 2、When panic state (E(t) ≥ 0.9) is detected, trigger safety protection mode: auxiliary force instantaneously reduced to 30%, at the same time, lock knee joint maximum flexion angle (≤90°), and through voice prompt "please pause walking"; When emotional recognition module fails, automatically switch to basic control mode to ensure system robustness.

[0074] In addition, for different scene types (walking on flat ground, up and down stairs, rehabilitation training), the specific values of basic angle value , acceleration coefficient , and force coefficient in joint angle control instruction are shown in Table 2.

[0075] Table 2: Joint angle control instruction value table ; Wherein, in order to ensure the safety and adaptability of human-machine cooperation, the adjustment of the assistance coefficient sets dynamic boundary conditions. When the intensity of the human muscle electrical signal exceeds the threshold value (50 mV, corresponding to excessive force), the lower limit of the assistance coefficient is set to 0.2 to avoid the exoskeleton hindering movement in the opposite direction; when is lower than the threshold value (5 mV, corresponding to no force), the upper limit is set to 0.9 to prevent excessive assistance leading to dependence.

[0076] Based on the generated power control instruction, the cooperative control algorithm of muscle group synchronous stimulation, and the target trajectory and torque mapping function, the pulse width is adjusted, and the pulse parameters of the active stimulation muscle and the passive stimulation muscle are output synchronously to perform human-machine cooperative control in multiple scenarios.

[0077] In this optional embodiment, the cooperative control algorithm of muscle group synchronous stimulation based on the generated power control instruction, the target trajectory and torque mapping function, the pulse width is adjusted, and the pulse parameters of the active stimulation muscle and the passive stimulation muscle are output synchronously to perform human-machine cooperative control in multiple scenarios include: According to the generated power control instruction and human movement physiology, a multi-muscle group function cooperation 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; According to the determined time offset, a distributed pulse stimulation array is designed, and the stimulation intensity ratio of the auxiliary muscle group and the core muscle group is dynamically adjusted in combination with the pre-defined cooperative stimulation coefficient; According to the adjusted stimulation intensity ratio, a target trajectory and torque mapping function is established to convert the movement trajectory into a target torque, and a human-machine torque distribution ratio is calculated to cooperatively distribute the pulse intensity and the mechanical assistance of the exoskeleton; It should be noted that in specific embodiments, for knee flexion and extension movement, the target trajectory and torque mapping function is set as: , wherein (proportionality coefficient), (differential coefficient), (viscous friction coefficient) are respectively set according to the scene, such as walking on flat ground .

[0078] In combination with the preset torque change rate threshold, the torque change trend predicted by the long short-term memory network is used to adjust the pulse width, and the pulse parameters of the active stimulation muscle and the passive stimulation muscle are output synchronously to perform human-machine cooperative control in multiple scenarios.

[0079] ​It should be noted that the long short-term memory network is used to predict the torque change trend, the input features of which include the torque sequence and the joint angle change rate in the past 100 ms; the network structure adopts 2 layers of LSTM, each layer containing 64 neural units; the prediction step is 50 ms; in order to improve the prediction accuracy, an error correction mechanism based on real-time torque feedback is introduced, and the prediction result is dynamically adjusted by setting a correction coefficient of 0.1.

[0080] In this optional embodiment, the expression of the pulse width is: ; In the formula, W a represents the pulse width; W a0 represents the basic pulse width, which is 50-200 μs; k t represents the torque error coefficient, which is 5 μs / (N·m); represents the deviation of the actual torque from the target torque.

[0081] The expression of the pulse parameter of the active stimulation muscle is: ; In the formula, I a represents the active stimulation current intensity; I a0 represents the active stimulation basic current intensity; k s represents the muscle electrical signal intensity coefficient; s represents the human muscle electrical signal intensity; T a represents the active stimulation pulse cycle, with the unit of ms; T a0 represents the active stimulation basic pulse cycle, with the value range of 50-100 ms; k d represents the distance coefficient, with the value range of 5-10; d represents the distance to the target, with the unit of m; the closer the distance to the target, the smaller the pulse cycle, the higher the stimulation frequency, and the human body is reminded to be well prepared.

[0082] In addition, the active stimulation is used to assist the human body in active force, and the pulse parameters thereof are related to the human muscle electrical signal intensity and the distance to the target, which have different values in different scenarios, such as the walking on flat ground scenario , the up and down stairs scenario ; represents the muscle electrical signal intensity coefficient, with the value range of 0.1-0.3; ​The human muscle electrical signal intensity is in mV; the greater the muscle electrical signal intensity, the stronger the human active force intention, the active stimulation current intensity is increased accordingly, and the human body is better assisted to exert force.

[0083] The expression of the pulse parameters of the passive stimulation muscle is: ; In the formula, I p The passive stimulation current intensity is represented by Ipassive; I p0 The passive stimulation base current intensity is represented by Ibase; k θ The joint angle deviation coefficient is represented by K; θ The actual joint angle and the target joint angle are represented by the deviation; T p The passive stimulation pulse period is represented by Tp, and the unit is ms; T p0 The passive stimulation base pulse period is represented by Tbase, and the value range is 80-150 ms; k v The movement speed coefficient is represented by K, and the value range is 10-20; v The exoskeleton robot movement speed is represented by V, and the unit is m / s; the faster the movement speed, the greater the pulse period, and the muscle is avoided from being stimulated too much.

[0084] In addition, the passive stimulation is used to drive the passive movement of the human body, and is suitable for rehabilitation training and the like. The pulse parameters thereof are related to the joint angle deviation and the exoskeleton robot movement speed, and the values are different in different scenes, for example, the joint angle deviation coefficient K in the rehabilitation training scene is 0.2-0.5. The joint angle deviation coefficient is represented by K, and the value range is 0.2-0.5; The actual joint angle and the target joint angle are represented by the deviation, and the unit is degree; the greater the joint angle deviation, the greater the passive stimulation current intensity, and the joint is promoted to move to the target angle.

[0085] It should be noted that based on the generated dynamic control instruction, the pulse width is adjusted by combining the synergistic control algorithm of the muscle group synchronous stimulation, the target trajectory and the torque mapping function, and the pulse parameters of the active stimulation muscle and the passive stimulation muscle are output synchronously to perform human-machine synergistic control in multiple scenes, which specifically includes: I. Synergistic control strategy of synchronous stimulation muscle group: ​1. Muscle group synergy mapping relationship construction: Based on human movement physiology, a multi-muscle group functional synergy matrix is established to clearly define the combination of core muscle groups and auxiliary muscle groups under different movement trajectories; when the hip joint is flexed, the core muscle group is the iliopsoas muscle (stimulation weight 0.6), and the auxiliary muscle group is the rectus femoris muscle (0.3) and the tensor fascia lata muscle (0.1); when the knee joint is extended, the core muscle group is the quadriceps femoris muscle (0.7), and the auxiliary muscle group is the gluteus maximus muscle (0.2) and the gastrocnemius muscle (0.1); through coherence analysis of electromyographic signals (calculate the cross-correlation coefficient between muscle groups), determine the time offset (≤10 ms) of synchronous stimulation to avoid muscle contraction timing conflicts.

[0086] 2. Multi-channel pulse synchronous triggering mechanism: design a distributed pulse stimulation array, including 8 independent output channels (each channel supports 0-100 mA current adjustment, pulse width 50-500 μs adjustable), use time diversity technology to achieve synchronous triggering; the main controller sends a synchronous clock signal (1 kHz), and each channel starts pulse output at the same time as the clock rising edge; introduce inter-channel phase compensation (±2 μs) to offset the synchronization error caused by cable transmission delay; for spastic risk muscle groups (hamstring muscles of stroke patients), set an inhibitory pre-stimulation: output a low-intensity pulse (10-15 mA) 50 ms before the main stimulation to reduce muscle spindle excitability.

[0087] 3. Balanced regulation of stimulation intensity between muscle groups: define the synergy stimulation coefficient ξ (value 0-1), which represents the stimulation intensity ratio of auxiliary muscle groups to core muscle groups: ξ = I 辅助 / I 核心 , dynamic range 0.2-0.8.

[0088] For example: when walking on flat ground, the ξ of the hip flexor and extensor muscle groups is 0.5 (balanced effort); when going up and down stairs, the ξ of the knee extensor muscle is 0.8 (strengthening auxiliary muscle synergy).

[0089] II. Dynamic adjustment mechanism combined with joint torque data: 1. Human-machine torque fusion model: establish a target trajectory-torque mapping function to convert the desired movement trajectory (such as the joint angle curve within a gait cycle into the target torque , where K p , K d are proportional-differential coefficients, are the actual angle and angular velocity (collected by joint encoders).

[0090] Calculate the human-machine torque distribution ratio: let the exoskeleton mechanical assistance torque be T m (t), the electric stimulation induced torque be T e (t), satisfy , define the distribution coefficient .

[0091] 2. Linkage adjustment of pulse intensity and mechanical assistance: design a double closed-loop adjustment system; Outer ring: Compare actual torque T actual (t) and target torque T d (t), calculate the 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 assistance. When ΔT<0 (excess torque): give priority to reducing the pulse intensity (the rate is twice that of the mechanical assistance).

[0092] 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.

[0093] 3. Forward-looking adjustment based on torque change rate: A torque change rate threshold γ (e.g., 5 N·m / s) is introduced. When dT_actual(t) / dt>γ, forward-looking adjustment is initiated. The LSTM network predicts the torque change trend over the next 100 ms (the input is the torque sequence over the past 500 ms). The pulse width is adjusted in advance (ΔW = 0.1·ΔT_predicted), implementing a "prediction-compensation" closed loop.

[0094] 3. Implementation of human-machine collaborative trajectory tracking: In the original pulse stimulation control algorithm, a new torque distribution decision layer is added. The execution steps are: receiving 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 distribution coefficient α; output mechanical assistance instructions (to adjust the servo motor current) and pulse stimulation parameters (intensity, width, frequency); set differentiated control parameters for different movement phases (stance phase / swing phase): α=0.3 in the stance phase (mechanical assistance dominates), α=0.7 in the swing phase (electrical stimulation dominates).

[0095] Tracking accuracy is improved, with angle error reduced from ±5° to ±1.5°, and torque tracking error reduced from ±8% to ±3%; energy consumption is optimized: when the proportion of electrical stimulation-induced torque is increased to 50%, the exoskeleton battery life is extended by 40%; human-machine collaboration fluency: the subjective score (1-10 points) of the subjects is improved from 6.2 points to 8.9 points, with no obvious sense of frustration. Set torque limit protection: When T actual (t)>1.2·T dWhen the pulse output is cut off immediately and the mechanical assistance is reduced (to 30% within 100ms); abnormal electromyographic signal detection: if the EMG signal amplitude exceeds the threshold (5 times the resting state), it is judged as muscle spasm and antiphase stimulation is automatically started (pulse frequency 100Hz, duration 200ms).

[0096] In addition, active and passive stimulation of muscles is achieved according to the generated pulse parameters.

[0097] like Figure 4 As shown, in a specific embodiment, taking the scene of going up and down stairs as an example, the working process of the present invention is as follows: 1. Data acquisition and preprocessing: The perception module's lidar, visual camera, and electromyographic sensor collect environmental information and human motion data. The lidar acquires stair step height (approximately 15 cm), step spacing (approximately 30 cm), and the distribution of surrounding obstacles. The visual camera captures images of the stair area and identifies the edge contours of the steps. The electromyographic sensor is attached to the quadriceps and hamstrings, collecting a muscle electrical signal strength of 25 mV. The pressure sensor detects 300 N of contact pressure between the foot and the step. The inertial measurement unit records leg acceleration of 1.2 m / s². The control module uses the Kalman filter algorithm to preprocess the data and set the initial state estimate for the acceleration data. , process noise covariance , measurement noise covariance , state transition matrix , control input matrix , the measurement matrix After three iterations of filtering, the smoothed acceleration data is 1.18 m / s², removing the random noise of ±0.3 m / s².

[0098] 2. Scene recognition: The preprocessed multi-source data is input into the CNN-RNN hybrid model; the CNN part adopts a two-layer convolution structure: the convolution layer 1 uses 3×3 convolution kernels (32) with a step size of 1 to extract features from the 64×64 environment image, and the output feature map size is 62×62×32; the pooling layer 1 uses 2×2 maximum pooling and outputs 31×31×32; the convolution layer 2 uses 3×3 convolution kernels (64) with a step size of 1 and outputs 29×29×64; the pooling layer 2 uses 2×2 maximum pooling and outputs 14×14×64. After flattening, a 12544-dimensional spatial feature vector is obtained; the RNN part uses an LSTM network (128 hidden units) to process 50 consecutive frames of time series data (sampling frequency 10Hz), and the hidden layer state update formula is ,in is a 32×128 weight matrix, is a 128×128 weight matrix, is a 128-dimensional bias vector; finally, the scene probability is output through a fully connected layer and a softmax function: going up or down stairs (92%), walking on flat ground (5%), and climbing a slope (3%). The current scene is determined to be going up or down stairs.

[0099] 3. Motion control strategy generation: Joint angle control: call the preset parameters of the going up or down stairs scene, the basic joint angle (knee flexion angle), acceleration coefficient , force coefficient ; according to the formula , the target angle of the knee joint is generated as 42°.

[0100] Motion speed control: basic speed , deceleration coefficient , the distance from the laser radar to the previous step ; according to the formula , the motion speed is limited to 0.25 m / s.

[0101] Force assistance coefficient calculation: basic force assistance coefficient , adjustment coefficient , the muscle electrical signal intensity is 25 mV, and , that is, the output auxiliary driving force is 45%.

[0102] 4. Pulse stimulation parameter calculation: Active stimulation: for the quadriceps femoris (active knee extension muscle group), set the basic current , muscle electrical signal coefficient , distance from the previous step , basic period , distance coefficient ; the calculation result is .

[0103] Passive stimulation: for the hamstrings (passive knee flexion muscle group), the deviation of the actual joint angle 38° from the target angle 42° , basic current , angle deviation coefficient ; motion speed , basic period , speed coefficient ; the calculation result is (frequency about 9.6 Hz).

[0104] 5、Instruction execution and feedback: After receiving the control instruction, the power module drives the knee joint to rotate through the servo motor and the reducer: within 0.8 seconds, the initial angle is changed from 20° to 42°, the angular velocity is kept at 27.5° / s, and the driving torque is dynamically adjusted according to the assistance coefficient (peak value 8 N·m); the pulse stimulation module outputs current through the electrode sheet: the active stimulation continuously outputs a pulse signal of 20 mA and a 75.2 ms cycle during the knee extension stage (0-0.4 s), and the passive stimulation outputs a pulse signal of 9.2 mA and a 103.75 ms cycle during the knee flexion stage (0.4-0.8 s); the control module receives the state feedback of the power module (actual angle error of the joint <1°) and the secondary sampling data of the perception module in real time, updates the control parameters every 100 ms, ensures that the step frequency is stable at 1.2 steps / s during the process of ascending and descending stairs, and the coincidence degree of the knee joint motion trajectory and the preset curve is >95%.

[0105] When the scene is switched to walking on flat ground, the above process is automatically repeated: the laser radar detects that the ground flatness error is <5 mm, the scene recognition model outputs a flat ground probability of 97%, and the control module calls the flat ground parameters (joint basic angle 15°, basic speed 0.6 m / s, active stimulation basic current 10 mA) to realize seamless switching.

[0106] In addition, in order to ensure the safety and effectiveness of neuromuscular electrical stimulation, the upper limit of the active stimulation current intensity is set to 50 mA to avoid muscle damage, and the lower limit is set to 5 mA to ensure effective stimulation; the pulse cycle is set to 20-200 ms, corresponding to a frequency of 5-50 Hz, covering the human muscle response frequency band. The above parameter boundaries comprehensively consider the physiological response characteristics and safety requirements of muscles, and are set in reference to the national standard "Medical electrical equipment Muscle stimulator safety requirements" (GB 9706.15-2008) to meet the safety specifications of medical electrical equipment.

[0107] Different muscle groups need to use different stimulation parameters due to physiological differences, for example, in the application of upper limb exoskeletons, the active stimulation basic current for the biceps brachii muscle responsible for active elbow flexion is set to 12 mA, and for the quadriceps femoris muscle responsible for active knee extension, it is set to 15 mA, which reflects the adaptability of stimulation parameters to muscle physiological characteristics.

[0108] Figure 2 An embodiment of the adaptive multi-scene exoskeleton robot control system of the application is shown.

[0109] In this alternative embodiment, the adaptive multi-scene exoskeleton robot control system comprises: The scene recognition module 201 is configured to utilize a Kalman filtering algorithm based on scene feature entropy to denoise the collected multi-modal sensor data, fuse the denoised multi-modal sensor data, combine a hybrid deep learning model to extract spatial features and time sequence features, and identify a current scene type. The cooperative control module 202 is configured to construct an emotion state and motion intention mapping model according to the current scene type identification result, generate corresponding power control instructions, and synchronously output pulse parameters of the active stimulation muscle and the passive stimulation muscle based on the power control instructions, so as to perform human-machine cooperative control in multiple scenes.

[0110] It should be noted that the scene recognition module 201 includes a data denoising unit (adopting a scene feature entropy self-adaptive Kalman filter), a data fusion unit (biomechanics constraint reinforcement learning + space-time attention), and a feature recognition unit (a CNN-RNN hybrid model), which are configured to output a scene type identification result; the cooperative control module 202 includes an emotion intention mapping unit (physiological signal acquisition + emotion-intention model), a control instruction generation unit (joint / speed / assistance coefficient calculation), and a pulse stimulation unit (muscle group cooperation + torque mapping), the emotion intention mapping unit receives the output result of the scene recognition module, and the pulse stimulation unit outputs active and passive stimulation parameters, so as to realize multi-scene human-machine cooperative control.

[0111] In one embodiment, a computer device, which can be a server, has an internal structure diagram as shown in Figure 3 The computer device includes a processor, a memory, and a network interface connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The database of the computer device is configured to store static information and dynamic information data. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement the steps in the above method embodiments.

[0112] Those skilled in the art can understand that Figure 3 The structure shown in the above

[0113] In addition, the present application further provides a computer device, comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0114] In addition, the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps in the above method embodiments.

[0115] A person of ordinary skill in the art can understand that all or part of the processes in the above method embodiments can be completed by a computer program instructing related hardware, and the computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, the computer program can include the processes of the above embodiments. Any reference to the memory, storage, database or other medium in each embodiment of the present application can include at least one of the non-volatile and volatile memories. The non-volatile memory can include a read-only memory (ROM), a magnetic tape, a floppy disk, a flash memory or an optical memory. The volatile memory can include a random access memory (RAM) or an external cache memory. As an illustration but not limitation, the RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0116] The present application is not limited to the structures described above and shown in the drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present application 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. The denoised multimodal sensor data is then fused and combined with a hybrid deep learning model to extract spatial and time series features to identify the current scene type. Based on the current scene type recognition results, an emotional state and motion intention mapping model is constructed to generate corresponding power control instructions. Based on the power control instructions, the pulse parameters of active and passive muscle stimulation are synchronously output to achieve human-machine collaborative control in multiple scenarios.

2. The adaptive multi-scenario exoskeleton robot control method according to claim 1 is characterized in that: 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. The hybrid deep learning model is combined to extract spatial features and time series features to identify the current scene type, including: Collect multimodal sensor data, denoise it using a Kalman filter algorithm based on scene feature entropy, and fuse the denoised multimodal sensor data using a biomechanically constrained reinforcement learning decision-making mechanism and a spatiotemporal fusion algorithm. The fused multimodal sensor data is input into a hybrid deep learning model consisting of a convolutional neural network and a recurrent neural network to extract spatial features and time series features respectively; The spatial features and time series features extracted are fused through the fully connected layer, and the probability distribution of various scenes is calculated through the normalized exponential function to identify the current scene type.

3. The adaptive multi-scenario exoskeleton robot control method according to claim 2, characterized in that: The multimodal sensor data is collected, denoised using a Kalman filter algorithm based on scene feature entropy, and the denoised multimodal sensor data is fused in combination with a biomechanically constrained reinforcement learning decision-making mechanism and a spatiotemporal fusion algorithm, including: Collect multimodal sensor data containing environmental information and human motion data, and use the Kalman filter algorithm to set the initial noise covariance matrix 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, a reinforcement learning algorithm is used to encode biomechanical constraints into a penalty function. This is combined with a double-delayed deep deterministic policy gradient algorithm to construct a reinforcement learning framework. Based on the scene feature entropy, a meta-learning algorithm is used to train a reinforcement learning framework, and combined with a spatiotemporal attention mechanism, the weights of multimodal sensor data are dynamically adjusted; The denoised multimodal sensor data is fused according to the adjusted multimodal sensor data weights.

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

5. The adaptive multi-scenario exoskeleton robot control method according to claim 1, characterized in that: The method includes: constructing an emotion state and motion intention mapping model based on the current scene type recognition result, generating corresponding power control instructions, and synchronously outputting pulse parameters for active muscle stimulation and passive muscle stimulation based on the power control instructions to perform human-machine collaborative control in multiple scenarios. Based on the recognition results of the current scene type and the motion control strategy of the emotional state recognition, a mapping model of the emotional state and motion intention is constructed, and corresponding power control instructions are generated based on the mapping model of the emotional state and motion intention; Based on the generated power control instructions, combined with the collaborative control algorithm of muscle group synchronous stimulation and the target trajectory and torque mapping function, the pulse width is adjusted, and the pulse parameters of the active and passive muscle stimulation are output synchronously to achieve human-machine collaborative control in multiple scenarios.

6. The adaptive multi-scenario exoskeleton robot control method according to claim 5, characterized in that: The method of constructing an emotional state and motion intention mapping model based on the recognition result of the current scene type and combining it with the motion control strategy of the emotional state recognition, and generating corresponding power control instructions based on the emotional state and motion intention mapping model includes: Based on the recognition result 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 galvanic response, and voice fundamental frequency. The emotional features include the standard deviation of heart rate variability, skin galvanic peak amplitude, and voice fundamental frequency fluctuation range. The extracted emotional features are input into the hybrid deep learning model, the weight of the emotional features is optimized through the attention mechanism, and the user's emotional state and corresponding quantitative value are output; Based on the quantitative value of emotional state, the emotional influencing factors are defined, and a mapping model between emotional state and movement intention is constructed; Combined with the emotional state and the emotional state and motion intention mapping model, the control parameters of the controller are dynamically adjusted to generate corresponding power control instructions, which include joint angle control instructions, motion speed control instructions and power assistance coefficient calculation.

7. The adaptive multi-scenario exoskeleton robot control method according to claim 6, characterized in that: The expression of the joint angle control instruction is: ； Where, Indicates joint angle control instructions; θ 0 Indicates the basic angle value; k a represents the acceleration coefficient; a Represents acceleration data; k f represents the force coefficient; f Indicates human body force data; The expression of the motion speed control instruction is: ； Where, v cmd Indicates motion speed control instruction; v 0 Indicates the base speed; α represents the deceleration coefficient; d Indicates the distance to the target; The expression of the power assist coefficient is: ； Where, λ Indicates the power assist coefficient; λ 0 Indicates the basic power coefficient; β represents the adjustment factor; s Indicates the strength of human muscle electrical signals.

8. The adaptive multi-scenario exoskeleton robot control method according to claim 7, characterized in that: The generated power control instructions are combined with the collaborative control algorithm of muscle group synchronous stimulation and the target trajectory and torque mapping function to adjust the pulse width and synchronously output the pulse parameters of the active and passive muscle stimulation to perform human-machine collaborative control in multiple scenarios, including: Based on the generated power control instructions and human movement physiology, a multi-muscle group functional synergy matrix is ​​established, and the correlation coefficients between muscle groups are 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 pre-defined synergistic stimulation coefficient, the stimulation intensity ratio of the auxiliary muscle groups and the 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 the target torque, and the human-machine torque distribution ratio is calculated to coordinate the distribution of pulse intensity and exoskeleton mechanical assistance. Combined with the preset torque change rate threshold, the torque change trend predicted by the long short-term memory network is used to adjust the pulse width, and the pulse parameters of active and passive muscle stimulation are output synchronously to achieve human-machine collaborative control in multiple scenarios.

9. The adaptive multi-scenario exoskeleton robot control method according to claim 8, characterized in that: The expression of the pulse width is: ； Where, W a Indicates pulse width; W a0 Indicates the basic pulse width; k t represents the torque error coefficient; Indicates the deviation between the actual torque and the target torque; The expression of the pulse parameter of the active muscle stimulation is: ； Where, I a represents the intensity of active stimulation current; I a0 represents the basic current intensity of active stimulation; k s Represents the muscle electrical signal intensity coefficient; s Indicates the strength of human muscle electrical signals; T a represents the active stimulation pulse period; T a0 represents the basic pulse period of active stimulation; k d represents the distance coefficient; d Indicates the distance to the target; The expression of the pulse parameter of the passive muscle stimulation is: ； Where, I p represents the passive stimulation current intensity; I p0 represents the basic current intensity of passive stimulation; k θ represents the joint angle deviation coefficient; θ Indicates the deviation between the actual joint angle and the target joint angle; T p represents the passive stimulation pulse period; T p0 represents the basic pulse period of passive stimulation; k v represents the motion speed coefficient; v Indicates the movement speed of the exoskeleton robot.

10. An adaptive multi-scenario exoskeleton robot control system, characterized in that: The system includes: The scene recognition module is used to denoise the collected multimodal sensor data using a Kalman filter algorithm based on scene feature entropy, fuse the denoised multimodal sensor data, and extract spatial and time series features using a hybrid deep learning model to identify the current scene type. The collaborative control module is used to build an emotional state and motion intention mapping model based on the current scene type recognition results, generate corresponding power control instructions, and synchronously output the pulse parameters of active muscle stimulation and passive muscle stimulation based on the power control instructions to perform human-machine collaborative control in multiple scenarios.

Images