A human-machine shared control decision-making method based on biological signals

Abstract

The application discloses a kind of human-machine shared control decision-making methods based on biological signal.The biological signal of operator is collected, and then sequentially amplified, filtered, analog-digital conversion processing is carried out to obtain digital signal;Digital signal is sequentially de-noised and feature extraction to obtain feature vector, and the high-level operation intention of operator is obtained by identifying feature vector;The state parameters of robot and environment are collected to construct state vector, and the state vector is input into the pre-trained reinforcement learning model, and the low-level control action is output;High-level operation intention and low-level control action are fused to generate fusion control quantity, and the robot is driven to work according to fusion control quantity, to realize human-machine shared control to robot movement.The present application realizes the deep fusion of human cognitive decision and machine accurate control ability, so that the remote operation system can respond to the immediate intention of operator, and can also cope with complex environment with the help of machine's autonomous intelligence, which significantly improves the efficiency and safety of human-machine cooperation.

Description

Technical Field

[0001] This invention relates to the field of robot control, and in particular to a human-machine shared control decision-making method based on biosignals. Background Technology

[0002] With increasing demands for power system reliability, live-line working has become a crucial means of ensuring continuous power supply. Traditional live-line working primarily relies on manual operation by personnel wearing heavy protective suits, which presents challenges such as high labor intensity, high safety risks, and limited operational precision. Existing live-line working robot systems mostly employ remote control, but still require operator intervention via handles or consoles, which suffers from operational delays and insufficient precision in complex electromagnetic environments.

[0003] Especially in high-voltage electric field environments, strong electromagnetic interference severely impacts traditional biosignal acquisition systems, leading to decreased signal quality and reduced recognition accuracy. Simultaneously, the extreme safety requirements of live-line work necessitate that control systems respond to safety hazards within milliseconds, a requirement that traditional control methods struggle to meet.

[0004] There is an urgent need for a method that can accurately obtain the operator's intention through anti-interference biosignal recognition technology and integrate it with the machine's autonomous safety decision-making intelligence to achieve reliable coordination between human high-level intentions and machine low-level control in a strong electromagnetic environment, thereby improving the safety, accuracy and efficiency of live-line work. Summary of the Invention

[0005] To address the problems in the background technology, this invention proposes a human-machine shared control decision-making method based on biosignals, which realizes the intelligent integration of human intentions and machine autonomous decision-making, and improves the system's adaptability in complex environments.

[0006] The technical solution adopted in this invention is: This invention includes the following steps: S1. Collect the operator's biological signals, and then amplify, filter, and convert the operator's biological signals into digital signals in sequence. S2. The digital signal is denoised and feature extracted sequentially to obtain a feature vector. The feature vector is then used to identify the high-level operational intent of the operator. S3. Collect the state parameters of the robot and the environment to construct a state vector, and input the state vector into a pre-trained reinforcement learning model to output low-level control actions; S4. Obtain the environmental safety state vector and safety state flag. Combine the environmental safety state vector and safety state flag to fuse the high-level operation intention and low-level control action to generate a fused control quantity. Then, drive the robot to perform operations based on the fused control quantity to achieve human-machine shared control of robot movement.

[0007] The operator's biosignals are specifically acquired through electroencephalography (EEG) or electromyography (EMG) sensors.

[0008] In step S2, the high-level operational intent of the operator is obtained by identifying the feature vector, specifically using a classification or regression model based on deep learning. The high-level operational intent of the operator includes approaching, grasping, installing, avoiding obstacles, and withdrawing the corresponding target state vector. The low-level control actions are specifically the adjustment amounts of impedance control parameters, including virtual mass adjustment, virtual damping adjustment, virtual stiffness adjustment, and attitude adjustment.

[0009] In step S3, the robot's state parameters include joint angle vector, joint angular velocity vector, end-effector contact force vector, end-effector position error, and end-effector posture error. The environmental state parameters include environmental force vector. The state vector includes robot joint angle, joint angular velocity, end-effector contact force, position error, posture error, and environmental force.

[0010] The environmental safety state vector specifically includes electric field strength, insulation distance, and leakage current; The safety status flag is specifically obtained by comparing the parameters in the environmental safety status vector with preset parameter thresholds. If any parameter exceeds the corresponding parameter threshold, the safety status flag is true; otherwise, it is false.

[0011] In step S3, the reinforcement learning model specifically employs a proximal policy optimization algorithm.

[0012] In step S3, the reward function of the reinforcement learning model is set according to the following formula: in, This represents the total reward function value of the reinforcement learning model. , , , , These respectively represent task tracking rewards, contact force control rewards, motion smoothness rewards, safety obstacle avoidance rewards, and intention rewards. Figure One Consistent rewards , , , , These respectively represent task tracking rewards, contact force control rewards, motion smoothness rewards, safety obstacle avoidance rewards, and intention rewards. Figure One The weighting parameters corresponding to consistency rewards Indicates positional error. Indicates attitude error. The L2 norm of a vector. The weighting parameters represent the attitude error. This represents the contact force between the end effector of the robotic arm and the environment. Indicates the desired contact force. Indicates the joint control torque. This represents the rate of change of the joint control torque. Indicates the position of the robotic arm's end effector. Indicates the first in the environment The location of the obstacle This indicates the preset safe distance threshold. Indicates the number of obstacles. Indicating the operational intentions of higher management With reinforcement learning output actions A consistency metric function between them.

[0013] In step S4, the high-level operation intention and low-level control action are combined with the environmental safety state vector and safety state flag to generate a fusion control quantity through the safety fusion control module. The safety fusion control module includes a first feature extraction branch, a second feature extraction branch, an attention weight generator, a weighted fusion unit and a safety mandatory intervention unit. The high-level operation intent is input to the first feature extraction branch for feature extraction, outputting a first feature vector. The low-level control action and environmental safety state vector are concatenated and input to the second feature extraction branch for feature extraction, outputting a second feature vector. Subsequently, the first and second feature vectors are element-wise added to obtain a combined feature vector. The combined feature vector is input to the attention weight generator for processing to obtain the initial high-level intent weight. The safety state flag is input to the safety mandatory intervention unit for processing to obtain the final high-level intent weight. The first and second feature vectors are linearly mapped to obtain the first mapped feature vector and the second mapped feature vector, respectively. Then, the final high-level intent weight, the first mapped feature vector, and the second mapped feature vector are input to the weighted fusion unit for weighted summation to obtain the fused control quantity.

[0014] The security mandatory intervention unit processes the situation according to the following steps: it judges the security status flag. If the security status flag is true, it indicates that the current environment is in a dangerous state, and the initial high-level intention weight is set to zero to obtain the final high-level intention weight; otherwise, the initial high-level intention weight remains unchanged and is used as the final high-level intention weight.

[0015] The method employs a human-machine shared control system, which includes: The biosignal acquisition module is used to acquire the operator's electroencephalogram (EEG) or electromyogram (EMG) signals, and then amplify, filter, and convert them from analog to digital signals to obtain digital signals. The intent recognition module is used to extract features and identify high-level operational intents based on the digital signals output by the biosignal acquisition module. The charged environment perception module is used to acquire the state parameters of the robot and the environment, the environmental safety state vector, and the safety state flags. The reinforcement learning decision module is used to construct state vectors based on the state parameters of the robot and the environment, and to process and output low-level control actions based on the state vectors. The safety fusion control module is used to combine environmental safety state vectors and safety state flags to fuse high-level operational intentions with low-level control actions to generate fusion control quantities. The execution feedback module is used to drive the robot to perform tasks based on the fused control variables and interact with the environment to provide feedback to the reinforcement learning decision module, thereby achieving closed-loop control.

[0016] The beneficial effects of this invention are: This invention identifies the operator's high-level operational intentions by decoding the operator's electroencephalogram (EEG) or electromyogram (EMG) signals, while simultaneously utilizing a reinforcement learning agent to generate low-level control actions based on the environmental state. A fusion module combines human intentions with machine autonomous decision-making to generate the final control command. This invention achieves a deep integration of human cognitive decision-making and precise machine control capabilities, enabling the teleoperation system to respond to the operator's immediate intentions while also leveraging the machine's autonomous intelligence to cope with complex environments, significantly improving the efficiency and safety of human-machine collaboration. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system architecture in this embodiment.

[0018] Figure 2 This is a control decision flowchart for this embodiment.

[0019] Figure 3 This is a flowchart of the biosignal acquisition, transmission and processing in this embodiment. Detailed Implementation

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0022] like Figure 2 and Figure 3 As shown, the method in this embodiment includes the following steps: S1. Collect the operator's biological signals, and then amplify, filter, and convert the operator's biological signals into digital signals in sequence. Specifically, biosignal acquisition and processing involves real-time acquisition of the operator's biosignals using electroencephalography (EEG) or electromyography (EMG) sensors. The data signals originate from the bioelectrical signals generated by the operator's brain or muscle activity. The raw biosignals undergo amplification and filtering preprocessing by signal conditioning circuitry to remove power frequency interference and baseline drift. They are then converted into digital signals by an analog-to-digital converter (ADC). The digital signals are transmitted to the main processing unit (such as a computer or embedded processor) via wireless (e.g., Bluetooth) or wired (e.g., USB) communication.

[0023] S2. The digital signal is denoised and feature extracted sequentially to obtain a feature vector. The feature vector is then used to identify the high-level operational intent of the operator. Specifically, in the main processing unit, the digital signal is denoised and its features are extracted. Based on the processed signal, the high-level operational intent of the operator is identified through a classifier or deep learning model.

[0024] Preferably, in the main processing unit, the signal is further processed, such as by using methods like independent component analysis (ICA) to remove artifacts and extract time-domain, frequency-domain, or time-frequency-domain features.

[0025] The intent recognition module employs a CNN-LSTM hybrid neural network structure. The input is a preprocessed biosignal feature vector, and the output is a probability distribution of the intent to perform live-line work. The network is specifically trained on a biosignal dataset containing electromagnetic interference, demonstrating robust recognition capabilities under strong interference environments.

[0026] High-level intent recognition: Feature vectors extracted from the preprocessed signal are input into an intent recognition model (such as a classification or regression model like a Support Vector Machine (SVM) or Convolutional Neural Network (CNN)) to identify the operator's high-level operational intent (such as "grab," "move," "stop," etc.). The identified high-level intent is then output to the human-machine decision fusion module.

[0027] S3. Collect the state parameters of the robot and the environment to construct a state vector, and input the state vector into a pre-trained reinforcement learning model to output low-level control actions; The robot and its environment are perceived in real time, and based on this state information, a reinforcement learning agent outputs low-level control actions. Specifically, environmental state perception and RL decision-making: At the same time, the robot's six-dimensional force sensor, electric field strength sensor, safe distance sensor and insulation status monitor acquire the state information of the live working environment in real time, and construct a state space containing safety parameters; the reinforcement learning agent based on the proximal policy optimization algorithm outputs low-level control actions that meet safety specifications according to this state.

[0028] The reinforcement learning decision-making module is based on the PPO algorithm. Its state space includes the robot's state, environmental information, safety parameters, and identified operational intentions. The action space comprises the adjustment amounts of admittance control parameters to meet safety regulations. The agent is pre-trained in a live-line working digital twin environment and then fine-tuned online in real live-line working tasks.

[0029] Environmental state perception and RL decision-making: Simultaneously, the system acquires environmental state data through the robot's force sensors, position sensors, and vision sensors. This data is constructed into a state space. The reinforcement learning agent (e.g., using PPO or DDPG algorithms) outputs low-level control actions based on this state. (such as admittance parameter adjustment amount) The RL agent's strategy has been pre-trained offline in a digital twin environment.

[0030] S4. Collect environmental safety state vectors and safety state flags, combine environmental safety state vectors and safety state flags, fuse high-level operation intentions and low-level control actions to generate fused control quantities, and then drive the robot to perform operations based on the fused control quantities to achieve human-machine shared control of robot movement.

[0031] like Figure 3 As shown, the fusion module integrates high-level intentions with low-level actions to generate final control commands. These final control commands drive the teleoperated robot to perform tasks, and the execution feedback is used to update the strategy of the reinforcement learning agent.

[0032] The safety fusion control module adopts a neural network based on an attention mechanism under safety constraints. It dynamically calculates the fusion weight of human and machine decisions based on intent confidence, environmental safety status, and operational complexity, and generates optimal control instructions that comply with safety procedures.

[0033] Human-Machine Decision Fusion: The fusion module receives high-level intentions from the intention recognition model and low-level actions from the RL agent. Through a pre-defined fusion strategy (such as rule-based weighting, fuzzy logic, or another lightweight neural network), the two are collaboratively fused to generate the final control command. This process ensures an effective combination of high-level human planning and low-level autonomous control capabilities of the machine.

[0034] Control Execution and Online Learning: The final control commands are sent to the robot's underlying controller, driving the robotic arm to perform the task. Execution generates new states and rewards. The system uses collected experience data (including biosignal intentions, environmental states, and execution effects) to fine-tune the RL strategy online, enabling it to better adapt to the real environment while adhering to human intentions, thus forming a closed-loop optimization.

[0035] The operator's biosignals are specifically acquired through electroencephalography (EEG) or electromyography (EMG) sensors.

[0036] In this embodiment, the EEG signal acquisition uses a 32-lead electromagnetically shielded dry electrode EEG cap with a sampling rate of not less than 512Hz and a common-mode rejection ratio of not less than 120dB for the signal conditioning circuit, which has anti-interference capability under strong electric field environment.

[0037] The anti-interference biosignal acquisition module employs an electromagnetically shielded 32-lead dry electrode EEG cap and an 8-channel surface electromyography sensor, equipped with active shielding technology and differential signal acquisition circuitry. The signal conditioning circuit boasts a common-mode rejection ratio exceeding 120dB and is specifically optimized for 50Hz power frequency interference suppression. Analog signals are converted to digital signals via a 16-bit ADC and transmitted to the main processing unit through fiber optic communication, ensuring signal integrity even in strong electromagnetic environments.

[0038] In step S2, the high-level operational intent of the operator is obtained by identifying the feature vectors, specifically using a classification or regression model based on deep learning.

[0039] Specifically, high-level intent recognition: In the main processing unit, the digital signal is subjected to adaptive filtering and blind source separation to eliminate residual electromagnetic interference and extract time-frequency domain features; the high-level operation intent of the operator in the live working environment is identified through a deep learning-based classification model, including operation intents such as approaching, grabbing, installing, obstacle avoidance, and withdrawal.

[0040] In step S3, the robot's state parameters include joint angle vectors. Joint angular velocity vector End contact force vector End position error End attitude error The state parameters of the environment include the environmental force vector. The state vector includes robot joint angles, joint angular velocities, end-effector contact force, position error, attitude error, and environmental forces.

[0041] The operator's high-level operational intent includes target state vectors corresponding to approaches, grasping, installing, obstacle avoidance, and withdrawal. The low-level control actions specifically involve adjustments to impedance control parameters, including virtual mass adjustments. Virtual damping adjustment amount Virtual stiffness adjustment amount and target position attitude adjustment amount .

[0042] The environmental safety state vector specifically includes electric field strength, insulation distance, and leakage current; The safety status flag is specifically obtained by comparing the parameters in the environmental safety status vector with preset parameter thresholds. If any parameter exceeds the corresponding parameter threshold, the safety status flag is true; otherwise, it is false.

[0043] The reinforcement learning agent employs a hybrid offline-online training model, which includes offline pre-training in a digital twin environment and online fine-tuning of the policy in a real-world operational environment using data containing biosignal intentions.

[0044] The state space of a reinforcement learning agent is defined as follows: ; Action space is defined as in: Adjustment amount for virtual quality; : Adjustment amount of virtual damping; Adjustment amount for virtual stiffness; : The amount of adjustment to the desired end position or attitude (used for fine-tuning the trajectory).

[0045] The state space is the "global state space," which contains six key variables: : Joint angle (robot's own state); Joint angular velocity (robot's own dynamics); End contact force (interactive information); Position error (task accuracy); : Attitude error (mission attitude accuracy); Environmental forces (such as electric field forces, wind forces, and other environmental disturbances).

[0046] In step S3, the reinforcement learning model specifically employs a proximal policy optimization algorithm.

[0047] In this embodiment, the policy network is a three-layer fully connected neural network with 256 neurons in each layer. During training, the focus is on optimizing the control policy under safety constraints.

[0048] In step S3, the reward function of the reinforcement learning model is set according to the following formula: in, This represents the total reward function value of the reinforcement learning model. , , , , These respectively represent task tracking rewards, contact force control rewards, motion smoothness rewards, safety obstacle avoidance rewards, and intention rewards. Figure One Consistent rewards , , , , These respectively represent task tracking rewards, contact force control rewards, motion smoothness rewards, safety obstacle avoidance rewards, and intention rewards. Figure One The weighting parameters corresponding to consistency rewards Indicates positional error. Indicates attitude error. The L2 norm of a vector. The weighting parameters represent the attitude error. This refers to the contact force between the end effector of the robotic arm (in this patent, "robot" and "robotic arm" are the same concept, both referring to a robotic arm that performs teleoperation tasks) and the environment. Indicates the desired contact force. Indicates the joint control torque. This represents the rate of change of the joint control torque. Indicates the position of the robotic arm's end effector. Indicates the first in the environment The location of the obstacle This indicates the preset safe distance threshold. Indicates the number of obstacles. Indicating the operational intentions of higher management With reinforcement learning output actions Consistency measures between them (such as cosine similarity).

[0049] In step S4, the high-level operation intention and low-level control action are combined with the environmental safety state vector and safety state flag to generate a fused control quantity through the safety fusion control module. The safety fusion control module includes a first feature extraction branch, a second feature extraction branch, an attention weight generator, a weighted fusion unit, and a safety mandatory intervention unit. The high-level operational intent is input to the first feature extraction branch for feature extraction, outputting a first feature vector. The low-level control action and environmental safety state vector are concatenated and input to the second feature extraction branch for feature extraction, outputting a second feature vector. Subsequently, the first and second feature vectors are element-wise added to obtain a combined feature vector. The combined feature vector is input to the attention weight generator for processing to obtain the initial high-level intent weight. The safety state flag is input to the safety mandatory intervention unit for processing to obtain the final high-level intent weight. The first and second feature vectors are linearly mapped to obtain the first mapped feature vector and the second mapped feature vector, respectively. Then, the final high-level intent weight, the first mapped feature vector, and the second mapped feature vector are input to the weighted fusion unit for weighted summation to obtain the fused control quantity.

[0050] The safety mandatory intervention unit processes the situation in the following steps: it judges the safety status flag. If the safety status flag is true, it indicates that the current environment is in a dangerous state, and the initial high-level intention weight is set to zero to obtain the final high-level intention weight; otherwise, the initial high-level intention weight remains unchanged and is used as the final high-level intention weight.

[0051] S4 specifically refers to: the security fusion control module receiving high-level operation intentions from the intent recognition module. and low-level control actions from the reinforcement learning decision module This module incorporates an attention-based neural network. Its inputs include intent confidence (provided by the intent recognition module), real-time environmental safety status (such as distance to the nearest obstacle and electric field strength), and task complexity metrics. The network dynamically outputs a fusion weight. When the live-line working safety monitoring submodule detects any safety parameter (such as insulation distance or leakage current) exceeding the preset threshold, the fusion module will... Force it to a lower value (e.g., 0.2), or even switch directly to a pure machine security policy. To ensure operational safety, the final fused control quantity is sent to the execution feedback module, where it is converted into joint control torques to drive the robot.

[0052] That is, a fusion module implemented through a neural network with an attention mechanism based on security constraints, which integrates high-level operational intentions. With low-level control actions The module performs fusion. It takes intent confidence, environmental safety status (such as distance to obstacles and electric field strength), and task complexity as inputs to dynamically calculate the fusion weights of human intent and machine decision-making. and This generates the final fused control input. When the safety monitoring module detects that environmental parameters (such as distance from a charged object) exceed a safety threshold, the fusion module will automatically increase the weight of the machine's autonomous decision-making (i.e., decrease the weight of the machine's autonomous decision-making). Security policies are prioritized. (Security constraints are applied not only to the fusion module but also through the reward function.) The training process of an embedded reinforcement learning model enables soft constraints on the agent's policy. Specifically, this is achieved through a secure fusion control module, which is based on a dual-branch attention neural network structure, including: First feature extraction branch: Receives the target state vector H obtained by mapping from the higher-level operation instructions, processes it sequentially through the first fully connected layer and the ReLU activation function, and outputs the first feature vector f. h ; The second feature extraction branch receives the low-level control action vector a from the reinforcement learning decision module. RL And the environmental safety state vector S from the charged environment sensing module safe After concatenating the two, the result is processed sequentially through a second fully connected layer and a ReLU activation function to output the second feature vector f. a ; Attention weight generator: This generator takes the first feature vector f and assigns it to the weights of the first feature vector f. h With the second eigenvector f a By adding elements together, we obtain the combined feature vector f. comb ; will f comb The input is sequentially fed into a fully connected layer and a Sigmoid activation function, and the output is a dynamically fused weight α, the value of which is in the range of [0,1]. Weighted fusion unit: The first feature vector f is... h Second eigenvector f a The mapped feature vector f is obtained by passing each feature vector through an independent linear mapping layer. h' and f a' Subsequently, the fusion control quantity u is calculated using a weighted summation method. Safety mandatory intervention unit: Receives safety status flags (Flags) from the energized environment sensing module in real time. danger ; When Flagdanger When true, the dynamic fusion weight α is forcibly set to a predefined minimum value ε (ε≥0) or zero, so that the fusion control quantity u is completely or mainly composed of the second feature vector f. a' The decision was made to prioritize the implementation of machine autonomy and safety strategies.

[0053] Preferably, the dimension d of the target state vector H is... h The lower-level control action vector a is 10 to 16. RL dimensional d a The value is 24, corresponding to the admittance control parameter adjustment of a 6-DOF robot; the environmental safety state vector S safe dimensional d s The output dimension d of the first and second fully connected layers is 10. f The output dimension of the linear mapping layer is 32, with a value of 64.

[0054] Security constraints are manifested in two ways: 1) Hard constraints / real-time intervention: In the fusion module, real-time security intervention is achieved by triggering weight adjustments or policy switching through security monitoring. 2) Soft constraints / policy learning: In the reward function of the reinforcement learning model, through... The project trains the agent to learn to proactively avoid obstacles and comply with safety rules at the policy level. Therefore, safety constraints are integrated into both decision fusion and policy learning.

[0055] The method also includes an intent confidence assessment module that automatically increases the weight of machine autonomous safety decisions when it detects that the quality of biosignals has deteriorated due to electromagnetic interference or psychological stress on the part of the operator.

[0056] like Figure 1 As shown, the biosignal-based live-line working human-machine shared control system of this invention includes an anti-interference biosignal acquisition module, a live-line environment perception module, an intent recognition module, a reinforcement learning decision-making module, a safety fusion control module, and an execution feedback module. These modules are connected via a shielded data bus to form a complete closed-loop control system.

[0057] The method employs a human-machine shared control system, which includes: The biosignal acquisition module is used to acquire the operator's electroencephalogram (EEG) or electromyogram (EMG) signals, and then amplify, filter, and convert them from analog to digital signals to obtain digital signals. The intent recognition module is used to extract features and identify high-level operational intents based on the digital signals output by the biosignal acquisition module. The charged environment perception module is used to acquire the state parameters of the robot and the environment, the environmental safety state vector, and the safety state flags. The method also includes a live-line work safety monitoring module (which is a functional subset or specialized application of the live-line environment sensing module. The live-line environment sensing module is responsible for collecting a wide range of environmental state parameters (such as force, vision, distance, etc.), while the safety monitoring module focuses on extracting specific safety-related variables (such as electric field strength, insulation distance, leakage current) from these parameters, comparing and judging them with preset safety thresholds, and triggering alarms or protection mechanisms when limits are exceeded. The two are related as inclusion and being included, perception and decision-making), which monitors electric field strength, insulation distance, and leakage current parameters in real time, and automatically triggers safety protection mechanisms when safety thresholds are exceeded.

[0058] The reinforcement learning decision module is used to construct state vectors based on the state parameters of the robot and the environment, and to process and output low-level control actions based on the state vectors. The safety fusion control module is used to fuse high-level operational intentions with low-level control actions to generate fused control quantities. The fusion module employs a neural network structure based on a security constraint-based attention mechanism. It dynamically adjusts the weight ratio between human intent and machine autonomous decision-making, ensuring that security policies are automatically prioritized when security risks are detected. Human-machine decision fusion: The security constraint-based fusion module collaboratively integrates high-level intents with low-level actions, dynamically adjusting the weight ratio based on real-time security status to generate final control commands that comply with security procedures.

[0059] The execution feedback module is used to drive the robot to perform tasks based on the fused control variables and interact with the environment to provide feedback to the reinforcement learning decision module, thereby achieving closed-loop control.

[0060] Control execution and online learning: The final control commands are used to drive the live-line working robot to perform tasks, and the execution feedback, safety status and intention confidence evaluation results are used to update the strategy of the reinforcement learning agent online.

[0061] In summary, the data signal flow can be summarized as follows: Sources: Operator's electrical biosignals (EEG / EMG); sensor data from the robot and environment.

[0062] Transmission path: Biosignals: Sensor → Signal Conditioning → ADC → Transmission Module → Main Processing Unit. Environmental data: Robot Sensor → Main Processing Unit.

[0063] How it is used: Biosignals are used to identify high-level intentions; environmental data is used for RL decision-making; the two are combined in the fusion module to generate control commands, drive the robot, and provide feedback to optimize the RL strategy.

[0064] Example 1: Application of live-line working on 10kV distribution lines In this embodiment, the system is applied to a 10kV distribution network line live-line connection operation robot. The operator wears an anti-interference EEG cap on an insulated bucket truck or at a ground control point and expresses the operation intention through specific motion imagery patterns (e.g., left-hand motion imagery corresponds to conductor positioning, right-hand motion imagery corresponds to clamp installation).

[0065] Biosignal acquisition parameters: Sampling rate: 512Hz Bandpass filter: 1-40Hz (enhanced 50Hz notch filtering) Feature extraction: Robust common space pattern + wavelet packet transform Intent recognition accuracy: 91.2% (under 10kV environmental interference) Reinforcement learning parameters: State space dimensions: 32 dimensions (including safety parameters such as electric field strength and safe distance) Action space dimensions: 6 dimensions (including safe speed limits) Reward function weights: 0.35, 0.20, 0.15, 0.15, 0.10, 0.05 Network structure: 256-256-128 fully connected Security Fusion Parameters: Baseline human weight: 0.7 Maximum machine weight (safe mode): 0.8 Safe distance threshold: 0.7 meters (10kV voltage level) Electric field strength safety threshold: 1.5kV / m Example 2: Application of live-line maintenance of 10kV power distribution equipment This embodiment provides a specific implementation process for live-line maintenance of 10kV power distribution equipment. The system identifies the operator's maintenance intentions through electromyography signals and generates a safe movement trajectory adapted to the compact space of the power distribution equipment by combining reinforcement learning agents.

[0066] The training and implementation process includes the following steps: Step S1: Construct a power distribution equipment model and an electric field distribution model in a 10kV distribution network digital twin environment to simulate the electromagnetic distribution under complex conditions of the distribution network.

[0067] Step S2: Collect biosignal data of operators in the distribution network operation environment, establish a dataset for recognizing the intent of live-line operation in the distribution network, and pay special attention to the operation characteristics in space-constrained environments.

[0068] Step S3: Initialize the parameters of the reinforcement learning agent and the safety constraint reward function, and pre-train them in the power distribution network simulation environment, focusing on optimizing the precise control strategy under compact space.

[0069] Step S4: Deploy the pre-trained strategy to the 10kV distribution network operation robot system, and use real-time sensor data and biosignals for online fine-tuning, paying particular attention to its adaptability in complex distribution network environments.

[0070] Step S5: During the online fine-tuning process, monitor the system's safety status in real time, periodically evaluate the strategy performance, and ensure control reliability under typical power distribution network disturbances.

[0071] In this embodiment, offline training uses a digital twin platform based on real distribution network data, while online fine-tuning employs real-time data streaming. The training time is approximately 36 hours to achieve the required performance for the job. Typical job tasks include common distribution network operations such as replacing fuses, installing grounding rings, and replacing surge arresters.

[0072] Through the above embodiments, the system of the present invention can operate stably in a 10kV distribution network environment. The biosignal recognition system maintains extremely high accuracy under typical interference in the distribution network. The reinforcement learning agent can autonomously adapt to the compact working space of the distribution network and automatically adjust the posture of the robotic arm when a nearby live object is detected, effectively ensuring the safety and efficiency of live-line work in the distribution network.

[0073] The specific manifestations of the beneficial effects in 10kV distribution network scenarios: At the 10kV voltage level, the system can achieve more precise operation and control, adapting to the characteristics of dense equipment and limited space in the distribution network. To address the frequent power outage and restoration needs of the distribution network, the system response speed is greatly improved, meeting the requirements for high efficiency in distribution network operations; In the complex electromagnetic environment of the power distribution network, the anti-interference ability of biosignals is further enhanced, and the recognition stability is improved; The reinforcement learning strategy is optimized for the characteristics of power distribution network operations, significantly improving operational efficiency while ensuring safety. The system can adapt to diverse equipment types and operating scenarios in the power distribution network, and has good versatility and scalability.

[0074] In summary, the method of this invention can effectively adapt to the special needs of live-line work on 10kV distribution networks, and realize the efficient integration of biosignal recognition and machine intelligence in the distribution network environment, which is of great significance for improving the safety, accuracy and efficiency of live-line work on distribution networks.

[0075] The above detailed embodiments illustrate the technical solution and beneficial effects of the present invention. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A human-machine shared control decision-making method based on biosignals, characterized in that, The method includes the following steps: S1. Collect the operator's biological signals, and then amplify, filter, and convert the operator's biological signals into digital signals in sequence. S2. The digital signal is denoised and feature extracted sequentially to obtain a feature vector. The feature vector is then used to identify the high-level operational intent of the operator. S3. Collect the state parameters of the robot and the environment to construct a state vector, and input the state vector into a pre-trained reinforcement learning model to output low-level control actions; S4. Obtain the environmental safety state vector and safety state flag. Combine the environmental safety state vector and safety state flag to fuse the high-level operation intention and low-level control action to generate a fused control quantity. Then, drive the robot to perform operations based on the fused control quantity to achieve human-machine shared control of robot movement.

2. The human-machine shared control decision-making method based on biosignals according to claim 1, characterized in that: The operator's biosignals are specifically acquired through electroencephalography (EEG) or electromyography (EMG) sensors.

3. The human-machine shared control decision-making method based on biosignals according to claim 1, characterized in that: In step S2, the high-level operational intent of the operator is obtained by identifying the feature vector, specifically using a classification or regression model based on deep learning. The high-level operational intent of the operator includes approaching, grasping, installing, avoiding obstacles, and withdrawing the corresponding target state vector. The low-level control actions are specifically the adjustment amounts of impedance control parameters, including virtual mass adjustment, virtual damping adjustment, virtual stiffness adjustment, and attitude adjustment.

4. The human-machine shared control decision-making method based on biosignals according to claim 1, characterized in that: In step S3, the robot's state parameters include joint angle vector, joint angular velocity vector, end-effector contact force vector, end-effector position error, and end-effector posture error. The environmental state parameters include environmental force vector. The state vector includes robot joint angle, joint angular velocity, end-effector contact force, position error, posture error, and environmental force.

5. The human-machine shared control decision-making method based on biosignals according to claim 1, characterized in that: The environmental safety state vector specifically includes electric field strength, insulation distance, and leakage current; The safety status flag is specifically obtained by comparing the parameters in the environmental safety status vector with preset parameter thresholds. If any parameter exceeds the corresponding parameter threshold, the safety status flag is true; otherwise, it is false.

6. The human-machine shared control decision-making method based on biosignals according to claim 1, characterized in that: In step S3, the reinforcement learning model specifically employs a proximal policy optimization algorithm.

7. The human-machine shared control decision-making method based on biosignals according to claim 1, characterized in that: In step S3, the reward function of the reinforcement learning model is set according to the following formula: in, This represents the total reward function value of the reinforcement learning model. , , , , These represent task tracking rewards, contact force control rewards, motion smoothness rewards, safe obstacle avoidance rewards, and intent consistency rewards, respectively. , , , , These represent the weight parameters for task tracking reward, contact force control reward, motion smoothness reward, obstacle avoidance reward, and intent consistency reward, respectively. Indicates positional error. Indicates attitude error. The L2 norm of a vector. The weighting parameters represent the attitude error. This represents the contact force between the end effector of the robotic arm and the environment. Indicates the desired contact force. Indicates the joint control torque. This represents the rate of change of the joint control torque. Indicates the position of the robotic arm's end effector. Indicates the first in the environment The location of the obstacle This indicates the preset safe distance threshold. Indicates the number of obstacles. Indicating the operational intentions of higher management With reinforcement learning output actions A consistency metric function between them.

8. The human-machine shared control decision-making method based on biosignals according to claim 1, characterized in that: In step S4, the high-level operation intention and low-level control action are combined with the environmental safety state vector and safety state flag to generate a fusion control quantity through the safety fusion control module. The safety fusion control module includes a first feature extraction branch, a second feature extraction branch, an attention weight generator, a weighted fusion unit and a safety mandatory intervention unit. The high-level operation intent is input to the first feature extraction branch for feature extraction, outputting a first feature vector. The low-level control action and environmental safety state vector are concatenated and input to the second feature extraction branch for feature extraction, outputting a second feature vector. Subsequently, the first and second feature vectors are element-wise added to obtain a combined feature vector. The combined feature vector is input to the attention weight generator for processing to obtain the initial high-level intent weight. The safety state flag is input to the safety mandatory intervention unit for processing to obtain the final high-level intent weight. The first and second feature vectors are linearly mapped to obtain the first mapped feature vector and the second mapped feature vector, respectively. Then, the final high-level intent weight, the first mapped feature vector, and the second mapped feature vector are input to the weighted fusion unit for weighted summation to obtain the fused control quantity.

9. The human-machine shared control decision-making method based on biosignals according to claim 8, characterized in that: The security mandatory intervention unit processes the situation according to the following steps: it judges the security status flag. If the security status flag is true, it indicates that the current environment is in a dangerous state. The initial high-level intent weight is set to zero to obtain the final high-level intent weight. Otherwise, keep the initial high-level intent weight unchanged and use the initial high-level intent weight as the final high-level intent weight.

10. The human-machine shared control decision-making method based on biosignals according to claim 1, characterized in that: The method employs a human-machine shared control system, which includes: The biosignal acquisition module is used to acquire the operator's electroencephalogram (EEG) or electromyogram (EMG) signals, and then amplify, filter, and convert them from analog to digital signals to obtain digital signals. The intent recognition module is used to extract features and identify high-level operational intents based on the digital signals output by the biosignal acquisition module. The charged environment perception module is used to acquire the state parameters of the robot and the environment, the environmental safety state vector, and the safety state flags. The reinforcement learning decision module is used to construct state vectors based on the state parameters of the robot and the environment, and to process and output low-level control actions based on the state vectors. The safety fusion control module is used to combine environmental safety state vectors and safety state flags to fuse high-level operational intentions with low-level control actions to generate fusion control quantities. The execution feedback module is used to drive the robot to perform tasks based on the fused control variables and interact with the environment to provide feedback to the reinforcement learning decision module, thereby achieving closed-loop control.

Images