Adaptive human-robot collaborative control method with safety constraints

Abstract

This invention relates to the field of human-machine collaborative control technology and discloses an adaptive human-machine collaborative control method that integrates safety constraints. The method includes: acquiring real-time environmental data and human motion state information; determining safety constraint parameters, performing priority evaluation and conflict resolution; performing path planning and iterative adjustment to form a feasible assist path sequence; combining human-machine interaction signals and human motion intention feedback to output an assist behavior prediction result; generating control commands to execute assist control; acquiring feedback data from the actuator and comparing it with the prediction result; if a deviation exists, triggering a re-evaluation and correction of the safety constraint set, and feeding the corrected constraint parameters back to the main control loop to dynamically adjust subsequent assist behaviors. This invention achieves an adaptive balance between safety constraints and dynamic operational needs through real-time scene perception, constraint conflict resolution, human intention fusion, and a closed-loop correction mechanism, improving the fluency, safety, and individual adaptability of human-machine collaboration.

Description

Technical Field

[0001] This invention relates to the field of human-machine collaborative control technology, and in particular to an adaptive human-machine collaborative control method that integrates safety constraints. Background Technology

[0002] During the construction of power transmission lines in mountainous areas, workers face extremely complex working environments, including steep slope climbing, cliff crossings, and heavy-duty walking on soft ground. Processes such as tower material handling, conductor laying, and hardware installation not only place extremely high demands on workers' physical fitness but also present significant challenges to the coordination between humans and equipment due to the complex and varied terrain. Flexible exoskeletons, wearable devices that provide assistance without restricting human freedom of movement, have been gradually introduced into the field of power transmission line construction in recent years, becoming an important technological direction for improving work efficiency and reducing the risk of occupational injuries. However, the complexity of the mountainous construction environment and the dynamic changes in task requirements present a core challenge for human-machine collaborative control: how to achieve a high degree of adaptation between exoskeleton assistance and human movement intentions while ensuring the safety of construction workers. This ensures that under different working conditions such as steep slope climbing, obstacle crossing, and heavy-duty walking, it can provide sufficient assistance to reduce worker fatigue without causing imbalance, collisions with surrounding obstacles, or compensatory muscle injuries due to excessive assistance caused by rigid control strategies.

[0003] To address the aforementioned issues, existing human-machine collaborative control methods primarily employ rule-based control strategies or offline-trained motion pattern recognition schemes. Rule-based methods constrain the exoskeleton's motion output by pre-setting safety boundaries, such as maximum assist torque, joint angle limits, and center-of-gravity shift thresholds. While these methods offer some protection in relatively structured construction environments, their core drawback lies in the static nature of the pre-set rules, which cannot adapt to real-time terrain changes and dynamic adjustments in human movement. In mountainous construction scenarios, with frequent terrain undulations, varying ground support conditions, and complex and diverse human movement patterns, static rules often fail to cover all working conditions. This frequently leads to a dilemma: either being too conservative, resulting in insufficient assistance and requiring workers to expend significant physical effort, or being too aggressive, causing safety risks and even worker falls. Motion pattern recognition-based methods attempt to classify patterns by collecting human motion data, such as distinguishing between climbing, traversing, and walking on flat ground, and then matching corresponding control parameters. However, such methods usually rely on offline trained motion models, and their recognition accuracy is limited by the coverage of training data. When there are sudden changes in terrain or workers temporarily adjust their work paths, recognition lag and pattern misjudgment problems become prominent, making it difficult to achieve both smoothness and safety in human-machine collaboration. The assist output of the exoskeleton often lags behind the actual movement needs of the human body, forming a response lag dilemma of "people waiting for terrain recognition, terrain recognition waiting for control commands". Summary of the Invention

[0004] Therefore, the technical problem to be solved by the present invention is to overcome the difficulties in coordinating and adapting safety constraints and dynamic task requirements in complex dynamic environments such as power transmission line construction in mountainous areas, the inability of static rules to adapt to terrain changes, and the lag in motion pattern recognition in the existing technology, which makes it difficult to achieve both smoothness and safety of human-machine collaboration. The present invention provides an adaptive human-machine collaborative control method that integrates safety constraints, so as to achieve an adaptive balance between safety constraints and dynamic operation requirements, and improve the smoothness, safety and individual adaptability of human-machine collaboration.

[0005] To address the aforementioned technical problems, this invention provides an adaptive human-machine collaborative control method incorporating safety constraints, applied to flexible exoskeleton devices, comprising the following steps: Acquire real-time environmental data and human motion status information, and generate current scene description information through sensor fusion processing; Based on the current scenario description information, determine at least one security constraint parameter, and perform priority evaluation and conflict resolution on each security constraint parameter to generate an optimized security constraint set; Based on the optimized set of safety constraints, an initial assistance path is generated through path planning, and the initial assistance path is iteratively adjusted according to the real-time terrain conditions to form a sequence of feasible assistance paths. Based on the feasible assistance path sequence, combined with human-computer interaction signals and human movement intention feedback, an enhancement strategy model is constructed through fusion processing to output the prediction results of assistance behavior; When the prediction results of the assistive behavior meet the dynamic operation requirements, a sequence of control instructions is generated and output to the actuator of the flexible exoskeleton for assistive control. The system acquires feedback data from the actuators and compares it with the predicted results of the assist behavior. If there is a discrepancy between the two, it triggers a reassessment and correction of the safety constraint set and feeds the corrected constraint parameters back to the main control loop for dynamic adjustment of subsequent assist behaviors.

[0006] In one embodiment of the present invention, real-time environmental data and human motion state information are acquired, and current scene description information is generated through sensor fusion processing, including: The inertial measurement unit collects the attitude angle data and angular velocity data of the human torso and limbs, the electromyography sensor collects the muscle activation signals of the main muscle groups of the human body, and the depth camera or lidar collects the three-dimensional point cloud data of the terrain in front and the ground slope information. The attitude angle data, angular velocity data, muscle activation signals, 3D point cloud data and ground slope information are input into the Kalman filter for time synchronization and data fusion, and output a multi-dimensional state vector representing the current human pose, motion trend and terrain features as the current scene description information.

[0007] In one embodiment of the present invention, the safety constraint parameters include: joint angle limit parameters, joint torque limit parameters, human body center of gravity stability boundary parameters, foot landing point safety area parameters, and assist output response time parameters. Among them, the joint angle limiting parameter is used to limit the maximum flexion and extension angle of each joint of the exoskeleton, the joint torque limiting parameter is used to limit the maximum assist torque output by each joint of the exoskeleton, the human body center of gravity stability boundary parameter is used to characterize the safe offset range of the human body center of gravity within the supporting polygon, the foot landing point safe area parameter is used to characterize the safe distance of the foot landing point relative to the terrain edge or obstacle, and the assist output response time parameter is used to characterize the maximum allowable delay time from intent recognition to assist output.

[0008] In one embodiment of the present invention, priority evaluation and conflict resolution are performed on each security constraint parameter to generate an optimized security constraint set, including: Based on the terrain slope, human movement speed, and muscle activation level in the current scene description information, calculate the real-time risk coefficients of each safety constraint parameter; The safety constraint parameters are sorted from high to low according to the real-time risk coefficient, and the safety constraint parameter with the highest risk coefficient is set as the highest priority constraint. When there is a contradiction between the control instructions of any two safety constraint parameters, the control instruction of the highest priority constraint shall be used as the basis for execution, and the control instructions of the lower priority constraints shall be reduced or delayed until the control instructions of all constraints are compatible with each other, thus forming an optimized set of safety constraints.

[0009] In one embodiment of the present invention, an initial assist path is generated through path planning based on the optimized safety constraint set, including: Based on the current position of the human foot and the target work point in the current scene description information, and combined with the joint angle limit parameters and foot landing safety area parameters in the optimized safety constraint set, a dynamic programming method is used to search for the optimal trajectory from the current position to the target work point on the terrain grid map. The optimal trajectory aims to minimize path length, terrain slope variation, and foot landing safety. It generates a sequence containing a series of foot landing coordinates and corresponding joint drive timings as the initial assist path.

[0010] In one embodiment of the present invention, the initial assistance path is iteratively adjusted according to real-time terrain conditions to form a feasible assistance path sequence, including: During the execution of assisted control, real-time terrain point cloud data and the actual foot placement position of the human body are continuously monitored; The actual foot landing position is compared with the expected landing position of the corresponding gait in the initial assist path, and the landing deviation is calculated. If the landing point deviation exceeds the preset safety deviation threshold, the path planning will be re-executed based on the current measured terrain conditions, starting from the current actual foot landing point position, to generate a corrected local assistance path. The corrected local assistance path will then be combined with the remaining unexecuted initial assistance path to form a feasible assistance path sequence.

[0011] In one embodiment of the present invention, an enhanced strategy model is constructed by combining human-computer interaction signals and human motion intention feedback through fusion processing, which is used to output the prediction result of assisted behavior, including: Collect human movement intention feedback signals, including: muscle activation timing signals collected by electromyography sensors, plantar pressure distribution signals collected by plantar pressure sensors, and human center of mass offset direction signals collected by inertial measurement units. Human-computer interaction signals and human motion intention feedback signals are input into a multilayer perceptron network. The multilayer perceptron network uses historical motion data as training samples to establish a mapping relationship between the current human motion intention and the expected assistance output. The multilayer perceptron network calculates the expected assist torque, expected assist timing, and expected assist duration in real time based on the current input signal, and uses these as the result of assist behavior prediction.

[0012] In one embodiment of the present invention, before inputting the human-computer interaction signal and the human motion intention feedback signal into the multilayer perceptron network, it is necessary to construct an input feature vector, including: The muscle activation time-series signal acquired by the electromyography sensor is segmented according to the acquisition time window. Within each time window, the peak value, mean value, rising slope and duration of muscle activation of each muscle group are extracted to form a muscle activation feature vector. The plantar pressure distribution signal collected by the plantar pressure sensor is divided into the heel area, arch area and forefoot area according to the foot region. The pressure peak, pressure center trajectory and ground contact time ratio of each area are calculated to form a plantar pressure feature vector. The human body center of mass offset direction signal collected by the inertial measurement unit is decomposed into forward and backward offset and left and right offset, and combined with the rate of change of the human body center of mass offset direction signal, a center of mass motion feature vector is formed. The muscle activation feature vector, plantar pressure feature vector, and center of mass motion feature vector are concatenated to form a fused feature vector, which serves as the input layer node data for the multilayer perceptron network.

[0013] In one embodiment of the present invention, the fused feature vector is input into a multilayer perceptron network to establish a mapping relationship between the current human movement intention and the desired assistance output, including: The multilayer perceptron network includes an input layer, at least two hidden layers, and an output layer. The number of nodes in the input layer is the same as the dimension of the fused feature vector. The number of nodes in the hidden layer decreases layer by layer. The number of nodes in the output layer is three, corresponding to the expected assist torque, the expected assist timing, and the expected assist duration, respectively. The multilayer perceptron network is pre-trained by collecting human motion data and corresponding optimal assist parameters under various terrain conditions. During the training process, the backpropagation algorithm is used to adjust the network weights to minimize the mean square error between the expected assist parameters output by the network and the optimal assist parameters in the training samples.

[0014] In one embodiment of the present invention, the dynamic operation requirements include: minimum assist efficiency requirement in flat walking mode, maximum assist torque requirement in uphill climbing mode, braking damping requirement in downhill descent mode, foot lift height requirement in obstacle crossing mode, and waist support torque requirement in load-bearing mode. When the prediction results of the assistive behavior meet the dynamic operation requirements, a sequence of control instructions is generated, including: Identify the current work mode based on the current scenario description information, and retrieve the requirement parameters corresponding to the current work mode from the dynamic work requirements; The expected assist torque, expected assist timing, and expected assist duration in the assist behavior prediction results are compared with the demand parameters. When the expected assist torque, expected assist timing, and expected assist duration all fall within the allowable range defined by the demand parameters, it is determined that the assist behavior prediction results meet the dynamic operation requirements.

[0015] The technical solution of the present invention has the following advantages compared with the prior art: The adaptive human-machine collaborative control method integrating safety constraints described in this invention dynamically generates scene description information by acquiring and fusing environmental and human data in real time. Based on this, it adaptively determines, evaluates, and resolves safety constraints to form an optimized set of safety constraints. Therefore, in complex and variable environments such as power transmission line construction in mountainous areas, it can effectively avoid the insufficient assistance or safety risks caused by traditional static rules.

[0016] Furthermore, this method performs path planning and iterative adjustments in real time based on the optimized constraint set, and constructs an enhanced strategy model by combining human-computer interaction and motion intention feedback to ensure that the prediction results of the assistive behavior are highly matched with the dynamic operation requirements. This overcomes the problems of recognition lag and pattern misjudgment in the existing technology, and realizes smooth, safe and adaptive coordination between the flexible exoskeleton assistive output and human motion intention. It significantly improves the work efficiency and safety of construction workers in conditions such as climbing steep slopes and walking with heavy loads. Attached Figure Description

[0017] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart of the steps of the adaptive human-machine collaborative control method integrating safety constraints of the present invention; Figure 2 This is a flowchart illustrating the steps of generating current scene description information through sensor fusion processing in this invention. Figure 3 This is a flowchart of the steps in this invention to prioritize and resolve conflicts among various security constraint parameters to generate an optimized set of security constraints. Figure 4 This is a flowchart illustrating the steps of path planning and iterative adjustment based on a set of safety constraints in this invention. Figure 5 This is a flowchart illustrating the steps involved in constructing the enhanced strategy model and assisting in behavior prediction according to the present invention. Figure 6 This is a flowchart illustrating the steps involved in generating dynamic job demand matching and control instructions according to the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0019] Reference Figure 1 As shown, this invention proposes an adaptive human-machine collaborative control method that integrates safety constraints. This method is applied to flexible exoskeleton devices and aims to solve the problem of difficulty in coordinating and adapting safety constraints with dynamic task requirements in existing technologies. The method first acquires real-time environmental data and human motion state information through sensor fusion to generate current scene description information. The core function of this step is to provide a precise situational awareness foundation for subsequent control decisions. Unlike existing technologies that only collect single-dimensional data, this method, through multi-sensor fusion, can simultaneously perceive multi-dimensional information such as terrain slope, ground flatness, human joint angles, movement speed, and muscle exertion state, forming a comprehensive understanding of the current working conditions and laying the data foundation for dynamic adjustment of safety constraints.

[0020] Based on this, the method determines at least one safety constraint parameter according to the current scene description information, and performs priority evaluation and conflict resolution on each safety constraint parameter to generate an optimized safety constraint set. The essence of this technical feature lies in changing the traditional method's "hard-stacked" processing logic of safety constraints. In existing technologies, different safety constraints are often treated equally or executed according to a fixed priority. When conflicts arise between constraints, the system either cannot make a decision or mechanically executes according to preset fixed rules, leading to a rigid control strategy. This method uses a priority evaluation mechanism to dynamically judge the urgency of each safety constraint based on the real-time scene. For example, in a steep slope climbing scenario, the priority of the center of gravity stability constraint is higher than the joint assist amplitude constraint; in an obstacle crossing scenario, the priority of the foot landing accuracy constraint is higher than the movement speed constraint. When there are contradictions between different constraints, the system adaptively coordinates through a conflict resolution mechanism, avoiding control rigidity caused by constraint contradictions at the source. This makes the safety constraint set no longer a fixed list of rules, but an adaptive constraint system that dynamically evolves with the scene.

[0021] Furthermore, this method generates an initial assist path based on an optimized set of safety constraints through path planning, and iteratively adjusts the initial assist path according to real-time terrain conditions to form a sequence of feasible assist paths. This feature enables the exoskeleton's assist path planning to no longer rely on preset fixed trajectories or offline trained motion templates, but can be corrected in real time according to terrain changes. When the worker's actual movement trajectory deviates from the initial planned path, the system quickly generates a corrected feasible path through iterative adjustments, ensuring that the assist output always matches the human's actual movement path, fundamentally solving the problem of inconsistent human-machine pacing caused by the lag in motion pattern recognition in existing technologies.

[0022] Building upon this foundation, this method constructs an enhanced strategy model based on feasible assistance path sequences, combined with human-machine interaction signals and human motion intention feedback, through fusion processing. This model is then used to output predicted assistance behavior results. The technical depth of this feature lies in its full utilization of the construction worker's own motion intentions and experiential judgments, integrating proactive human decision-making into the control loop. Unlike existing technologies that rely solely on sensor data for pattern matching, this method collects human motion intention feedback, such as electromyographic signals, joint torque changes, and trends in human posture adjustment. This allows the exoskeleton's assistance behavior to not only respond to the physical environment data collected by sensors but also understand human intentions and real-time decisions. This deeply integrated human-machine strategy model significantly improves the naturalness of human-machine collaboration in complex terrain, transforming the exoskeleton from a passively responding assistance tool into a proactive partner that cooperates with human motion intentions.

[0023] This method further generates a sequence of control commands based on the predicted assistive behavior, which is then output to the actuators of the flexible exoskeleton for assistive control. Simultaneously, the method acquires feedback data from the actuators and compares it with the predicted assistive behavior. If a deviation exists, a reassessment and correction of the safety constraint set is triggered, and the corrected constraint parameters are fed back to the main control loop for dynamic adjustment of subsequent assistive behaviors. This closed-loop feedback mechanism constitutes the core advantage of this method: by comparing the actual execution results with the expected behavior in real time, the system can autonomously detect execution deviations caused by sudden environmental changes, adjustments in human posture, changes in ground support conditions, or differences in equipment dynamic characteristics, and immediately initiate a reassessment and correction of safety constraints. This closed-loop structure of "execution-feedback-reassessment-correction" allows the exoskeleton's control strategy to continuously optimize with changing operating conditions, rather than remaining unchanged after a one-time initial setting, significantly improving the system's adaptability in dynamic and complex environments.

[0024] By organically combining the aforementioned technical features, the method proposed in this application achieves an adaptive balance between safety constraints and dynamic operational requirements. In mountainous power transmission line construction scenarios, this method can dynamically adjust the exoskeleton's assistance strategy based on multi-dimensional information such as the real-time terrain, movement status, and movement intentions of the construction workers. When workers climb steep slopes, the system automatically increases the weight of center-of-gravity stability constraints and foot landing accuracy constraints through priority evaluation, appropriately limiting the range of joint assistance to prioritize body balance. When workers walk with loads on relatively flat sections, the system automatically increases the weight of assistance output constraints, providing sufficient assistance to reduce physical exertion. When workers need to cross obstacles, the system anticipates the crossing action in advance based on feedback of human movement intentions, providing precise assistance torque at critical moments to avoid movement deformation due to inappropriate assistance timing. This scenario-adaptive constraint coordination mechanism effectively solves the inherent problem in existing technologies where safety and task efficiency are difficult to balance.

[0025] Meanwhile, by introducing human motion intention feedback and execution closed-loop correction, this method enables the exoskeleton's control strategy to continuously adapt to individual differences and changes in working conditions. Different construction workers exhibit significant differences in physical condition, movement habits, and work preferences, making the one-size-fits-all control mode in existing technologies insufficient to meet individualized needs. This method, through real-time acquisition of human motion intention feedback, allows the exoskeleton to gradually learn and adapt to the individual user's movement characteristics; through an execution deviation-triggered constraint correction mechanism, the system can identify and compensate for control errors caused by individual differences. This continuous optimization capability ensures that the exoskeleton maintains good human-machine collaboration performance across different users, eliminating the need for offline-trained general motion models and significantly improving the system's robustness and practicality in dynamic and complex environments. Ultimately, this method provides a safe, efficient, and comfortable human-machine collaboration solution for power transmission line construction in mountainous areas, effectively reducing worker fatigue and the risk of work-related injuries while ensuring safety and efficiency in complex terrain operations.

[0026] In practical applications, the efficient, accurate, and comprehensive collection and effective fusion of multi-source heterogeneous data are fundamental to achieving adaptive human-machine collaborative control. This ensures that the generated scene description information accurately reflects the real-time state, movement intentions, and complex environmental characteristics of the human body. Insufficient data collection or improper fusion processing can distort the scene description information, thereby affecting the accuracy of subsequent safety constraints and the effectiveness of assisted control.

[0027] Reference Figure 2 As shown, in this embodiment, sensor fusion processing achieves the fusion of multi-source heterogeneous data through a Kalman filter. The Kalman filter is a recursive optimal state estimation algorithm. Its core principle is to predict the state at the next moment using the system state equation, while simultaneously using the observation equation to introduce sensor measurements to correct the prediction result. By dynamically weighting the confidence levels of the predicted and observed values, the optimal state estimate is output.

[0028] Specifically, the sensors involved in this embodiment include an inertial measurement unit (IMU), an electromyography (EMG) sensor, and a depth camera or lidar. The IMU is fixed to the human torso and various joints of the exoskeleton, used to collect the posture angles, angular velocities, and linear accelerations of the human torso and limbs, with a sampling frequency typically between 100Hz and 200Hz. The EMG sensor is attached to major muscle groups, including the rectus femoris, gastrocnemius, erector spinae, and deltoid muscles, used to collect muscle activation signals, with a sampling frequency up to 1000Hz, capable of sensing human movement intentions approximately 50 to 200 milliseconds in advance. The depth camera or lidar is fixed to the waist or back of the exoskeleton, used to collect three-dimensional point cloud data of the terrain in front, with a sampling frequency between 30Hz and 60Hz. Through algorithms such as point cloud segmentation, plane fitting, and obstacle detection, terrain feature parameters such as ground slope angle, step height, and obstacle distance can be extracted.

[0029] Before fusion processing, time synchronization and spatial alignment of multi-source data are required. Since the sampling frequencies of each sensor differ, a hardware-triggered method is used, with the unified clock of the exoskeleton control system as the reference. Each frame of sensor data is timestamped, and linear interpolation is used to interpolate the low-frequency sensor data to the sampling time of the high-frequency sensor, achieving time alignment. For spatial alignment, the relative pose transformation matrix between the coordinate systems of each sensor is obtained through pre-calibration. Data from the inertial measurement unit, electromyography sensor, and depth camera are then uniformly transformed into a world coordinate system with the human body's center of mass as the origin, ensuring that all data are fused within the same reference frame.

[0030] After time synchronization and spatial alignment are completed, multi-source data is input into a Kalman filter for state estimation. The state vector of the Kalman filter is used to characterize the three-dimensional position and velocity of the human body's center of mass, the posture angle and angular velocity of the human torso, and parameters such as the slope angle of the terrain in front, step height, and distance to obstacles. The state equation describes the evolution of the system state over time. According to the human kinematics model, the position and velocity of the center of mass satisfy an integral relationship, as do the posture angle and angular velocity. The terrain feature parameters change slowly over a short period of time, so a uniform change model is adopted. The state equation also considers the assist torque output by the exoskeleton actuator as a control input. The assist torque is mapped to the human joint angular acceleration through the exoskeleton dynamics model, which in turn affects the motion state of the center of mass and the posture angle. Process noise is used to characterize the uncertainty of the model prediction and follows a Gaussian distribution.

[0031] The observation equations describe the mapping relationship between the sensor measurements and the system state. The inertial measurement unit directly measures the attitude angles and angular velocities in the state vector, and its measurements are linearly related to the corresponding state components. The linear acceleration measurements have a kinematic differential relationship with the velocity and position in the state vector, and the mapping is established through pre-calibrated differential coefficients. The electromyography (EMG) sensor measurements reflect the human's movement intention and have a nonlinear relationship with joint angles and angular velocities. In this embodiment, a pre-trained linear regression model is used to approximate muscle activation as a linear combination of joint angles, thus incorporating the EMG observations into the linear observation equation framework. The terrain feature parameters extracted by the depth camera or lidar have a direct linear correspondence with the corresponding terrain parameters in the state vector. The observation noise of each sensor is modeled according to its physical characteristics, follows a Gaussian distribution, and characterizes the measurement uncertainty.

[0032] The Kalman filter achieves optimal state estimation through two alternating steps: prediction and updating. In the prediction step, the filter uses the state equation, based on the optimal state estimate from the previous time step, and combines the exoskeleton assist torque as the control input to predict the current state according to the human kinematics model, thus obtaining a prior estimate. Simultaneously, the filter calculates the uncertainty covariance matrix of the prior estimate, which transmits the uncertainty from the previous time step to the current time step through the state transition relationship, and adds process noise.

[0033] In the update step, the filter incorporates the actual measurements from each sensor at the current moment. First, the Kalman gain is calculated; this gain is a dynamic weighting matrix whose magnitude depends on the relative relationship between prediction uncertainty and observation noise. When prediction uncertainty is high and observation noise is low, the gain increases, and the system trusts the current observation more; conversely, when prediction uncertainty is low and observation noise is high, the gain decreases, and the system trusts the prediction more. Through this adaptive weighting mechanism, the filter achieves intelligent fusion of multi-sensor information. Subsequently, the filter compares the actual and predicted observations at the current moment, calculates the observation residuals, and uses the Kalman gain to map the observation residuals to the state space, performing weighted correction on the prior estimate to obtain the posterior optimal state estimate. Finally, the filter updates the uncertainty covariance matrix of the posterior estimate for recursive calculation in the next moment. As observation data is continuously introduced, the uncertainty of the state estimate gradually decreases, and the estimation accuracy continuously improves.

[0034] The Kalman filter recursively executes the above prediction and update steps in each control cycle, outputting the optimal state estimate for the current moment, i.e., the current scene description information. This state vector includes the three-dimensional position and velocity of the human body's center of mass, used to determine the human body's balance state and motion trend; it includes the posture angle and angular velocity of the human body's torso, used to characterize the human body's current pose and rotation state; and it includes the slope angle of the terrain ahead, the height of the steps, and the distance to obstacles, used to provide an environmental perception basis for subsequent path planning and safety constraint adjustment.

[0035] In adaptive human-machine collaborative control methods that integrate safety constraints, safety constraint parameters need to be determined based on the current scene description information. However, if the specific and comprehensive types and definitions of safety constraints are not clearly defined, the system may be unable to effectively identify potential risks in complex and changing environments, thereby affecting the safety and assistive effect of the flexible exoskeleton device, and may even cause harm to the user. To address this, this application further proposes safety constraint parameters including joint angle limit parameters, joint torque limit parameters, human body center of gravity stability boundary parameters, foot landing safety area parameters, and assistive output response time parameters.

[0036] Specifically, joint angle limiting parameters are used to limit the maximum flexion and extension angles of each joint of the flexible exoskeleton. These settings aim to prevent joint movement from exceeding the physiological limits of the human body or the mechanical limits of the equipment, thereby avoiding joint injury to the user or damage to the equipment structure. For example, the flexion and extension angles of the knee joint are typically limited to between 0 degrees (fully extended) and approximately 140 degrees (maximum flexion). These limits can be adjusted based on individual differences, exoskeleton design, and specific operational requirements. Joint torque limiting parameters are used to limit the maximum assist torque output by each joint of the flexible exoskeleton. The purpose is to ensure that the assistance provided by the exoskeleton is not excessive, avoiding unnecessary stress or damage to the human joints, muscles, and ligaments, while also protecting the exoskeleton's own drive mechanism. The setting of this parameter must comprehensively consider the human body's load-bearing capacity, the exoskeleton's design load, and the intensity of the work. Human center of gravity stability boundary parameters characterize the safe deviation range of the human body's center of gravity within the supporting polygon. Maintaining the center of gravity within the supporting polygon is crucial for balance during human movement. This parameter defines a safe zone where the risk of instability increases significantly when the human body's center of gravity exceeds this zone. The settings for these parameters are typically based on human dynamics models and balance control theory. The foot landing safety zone parameter characterizes the safe distance between the foot's landing point and the edge of the terrain or an obstacle. When walking or crossing obstacles, the foot needs to land in a safe, stable area to avoid missteps, slips, or collisions. This parameter defines a minimum safe distance to ensure sufficient support space around the foot landing point and that it is far from potentially dangerous areas. The assist output response time parameter characterizes the maximum permissible delay time from intent recognition to assist output. In human-machine collaborative control, the exoskeleton's assist response speed is crucial for achieving smooth and natural coordinated movement. Excessive response time can lead to assist lag, causing user discomfort or increasing the risk of falls. This parameter sets a time limit, requiring the system to begin providing assistance within this time window after recognizing the user's movement intention.

[0037] In practical applications, simply listing safety constraint parameters cannot guarantee that the system can intelligently cope with various potential risks in complex and ever-changing human-machine collaborative control scenarios. This may lead to confusion in control commands or a reduction in safety margins. Therefore, referring to... Figure 3As shown, this application further proposes to prioritize and resolve conflicts among various safety constraint parameters to generate an optimized set of safety constraints. Specifically, this includes: calculating the real-time risk coefficient of each safety constraint parameter based on the terrain slope, human movement speed, and muscle activation level in the current scene description information; sorting the safety constraint parameters from high to low according to the real-time risk coefficient, and setting the safety constraint parameter with the highest risk coefficient as the highest priority constraint; when there is a contradiction between the control instructions of any two safety constraint parameters, the control instruction of the highest priority constraint is used as the basis for execution, and the control instructions of the lower priority constraints are reduced or delayed until the control instructions of each constraint are compatible, thus forming an optimized set of safety constraints.

[0038] In this embodiment, the first step in priority assessment and conflict resolution is to calculate the real-time risk coefficient of each safety constraint parameter based on the terrain slope, human movement speed, and muscle activation level in the current scene description information. The real-time risk coefficient is used to quantify the urgency or potential harm of violating each safety constraint under the current working conditions. The higher the risk coefficient, the more priority the constraint needs to be protected at the current moment.

[0039] The current scene description information is output from the sensor fusion processing step, including the three-dimensional position and velocity of the human body's center of mass, the posture angle and angular velocity of the human torso, and parameters such as the slope angle of the terrain in front, the height of steps, and the distance to obstacles. Based on this, the system further extracts key features for risk coefficient calculation: terrain slope angle, representing the degree of inclination of the current or forward ground, directly obtained from the terrain parameters output by the depth camera or LiDAR fusion; human movement velocity, representing the rate of movement of the human body's center of mass in the forward direction, extracted from the state vector output by the Kalman filter; and muscle activation level, representing the force level of the major muscle groups at the current moment, obtained from the muscle activation signal collected by the electromyography sensor and preprocessed, which reflects the intention and intensity of the human body's active force exertion.

[0040] The safety constraint parameters involved in this embodiment include joint angle limit parameters, joint torque limit parameters, human body center of gravity stability boundary parameters, foot landing point safety zone parameters, and assist output response time parameters. The real-time risk coefficients of each constraint parameter are calculated as follows.

[0041] For joint angle limit parameters, the risk coefficient reflects the degree to which the current joint angle approaches its mechanical or physiological limit. The system calculates the difference between the current angle of each joint and the preset safety limit angle based on the human posture angle data in the current scene description information, combined with the current angle values ​​of each joint on the exoskeleton. The risk coefficient of the joint angle limit parameter is calculated using a normalization method: the difference between the current angle and the limit angle is divided by the safety margin range, and then mapped using an exponential function, causing the risk coefficient to rise rapidly when the joint angle approaches the limit. When the terrain slope increases, the human body will naturally adjust the joint angles to maintain balance, making the joints more likely to approach the limit state; therefore, the risk coefficient increases with the slope. When the human body's movement speed increases, the range of change in joint angles increases, the probability of approaching the limit increases, and the risk coefficient also increases accordingly. When muscle activation increases, it indicates that the human body is actively exerting force, and the joints are under greater load. If the joint approaches the limit at this time, the risk of injury is higher; therefore, the risk coefficient is further amplified when muscle activation is high.

[0042] For joint torque limit parameters, the risk coefficient reflects the degree to which the current exoskeleton output torque or the torque borne by the human joint approaches the safe torque limit. The system estimates the required assist torque for the joint and the torque borne by the human joint itself based on the human movement speed and muscle activation level in the current scene description information, combined with the exoskeleton dynamics model. The risk coefficient of the joint torque limit parameter is calculated as the ratio of the currently estimated torque to the preset safe torque limit. When the terrain slope increases, the human body needs to do work against the component of gravity, significantly increasing the torque borne by the joint and raising the risk coefficient accordingly. When the human movement speed increases, the inertial force increases, increasing the joint torque demand and raising the risk coefficient. When the muscle activation level increases, it indicates that the human body is outputting a large muscle torque, increasing the likelihood that the total joint torque is approaching the limit and correspondingly raising the risk coefficient.

[0043] For the stability boundary parameters of the human body's center of gravity, the risk coefficient reflects the proximity of the current position of the human body's center of gravity to the boundary of the supporting polygon. Based on the 3D position of the human body's center of gravity and the foot landing point in the current scene description information, the system calculates the shortest distance between the projection point of the human body's center of gravity on the horizontal plane and each boundary of the foot supporting polygon. The supporting polygon is composed of the landing points of both feet; in the single-foot support phase, it is composed only of the single-foot landing area. The risk coefficient of the stability boundary parameters of the human body's center of gravity is calculated as the ratio of this shortest distance to a preset stability margin; the smaller the distance, the higher the risk coefficient. When the terrain slope increases, the component of gravity along the slope direction will pull the center of gravity downhill, reducing the effective area of ​​the supporting polygon and significantly increasing the risk coefficient. When the human body's movement speed increases, inertial forces disturb the position of the center of gravity, increasing the magnitude of the center of gravity shift and raising the risk coefficient. When muscle activation increases, it indicates that the human body is actively adjusting its posture to maintain balance through muscle exertion. If the center of gravity is already close to the boundary at this time, the adjustment space is limited, and the risk coefficient is appropriately amplified when muscle activation is high.

[0044] The risk coefficient for the foot landing safety zone parameter reflects the degree to which the current or next foot landing position is close to the terrain edge, obstacle, or danger zone. Based on terrain parameters in the current scene description information, including obstacle locations, step edges, steep slope boundaries, etc., combined with the human's current movement speed and direction, the system predicts the next foot landing position and calculates the shortest distance between the predicted landing point and the nearest danger zone boundary. The risk coefficient of the foot landing safety zone parameter is calculated as the ratio of this shortest distance to a preset safe distance threshold; the smaller the distance, the higher the risk coefficient. When the terrain slope increases, the effective flat area of ​​the foot landing zone decreases, increasing the risk of the landing point being close to the edge, and thus the risk coefficient rises. When the human movement speed increases, the uncertainty of the landing point prediction increases, and if the landing point is incorrect, the adjustment time is shorter, and the consequences are more severe; therefore, the risk coefficient increases with speed. When muscle activation increases, it indicates that the human is performing a larger amplitude movement, the foot swing amplitude increases, the accuracy of landing point control decreases, and the risk coefficient increases accordingly.

[0045] For the assist output response time parameter, its risk coefficient reflects how close the delay between the system recognizing the human's movement intention and outputting assistance approaches the safe response time limit. The system estimates the required response time margin under the current operating condition based on the human movement speed and muscle activation level in the current scene description information, combined with the actual measured response delay in historical control cycles. The risk coefficient of the assist output response time parameter is calculated as the ratio of the currently estimated delay to the preset safe response time limit. The larger the ratio, the more delayed the response, and the higher the risk coefficient. When the terrain slope increases, the human's need for assistance becomes more urgent, the delay tolerance decreases, and the risk coefficient increases. When the human movement speed increases, the movement rhythm accelerates, the requirement for response time is higher, and the risk coefficient increases. When the muscle activation level increases, it indicates that the human's movement intention is strong and the movement is about to occur; if the response delay is too long at this time, the optimal assistance opportunity will be missed, and the risk coefficient will increase significantly.

[0046] After calculating the real-time risk coefficients for each safety constraint parameter, the system obtains a set of values ​​reflecting the current urgency of each constraint. These risk coefficients are all normalized to a range of 0 to 1, where 0 indicates that the constraint is currently in an absolutely safe state with no risk of violation; 1 indicates that the constraint has approached or reached the safety boundary and requires immediate action. By incorporating three key working condition characteristics—terrain slope, human movement speed, and muscle activation level—into the calculation logic of each risk coefficient, the system can accurately perceive changes in the urgency of each safety constraint under different working conditions, providing a quantitative basis for subsequent prioritization and conflict resolution. For example, in a steep downhill scenario, the increased slope leads to a significant increase in the risk coefficients of the center of gravity stability boundary parameters and the foot landing safety zone parameters, while the risk coefficient of the joint angle limit parameters is relatively low. Based on this, the system determines that the current constraints should prioritize center of gravity stability and landing safety. In a heavy-load uphill scenario, the increased speed of human movement and the increased degree of muscle activation lead to a significant increase in the risk coefficients of the joint torque limit parameters and the assist output response time parameters. Based on this, the system determines that the current constraints should prioritize torque limitation and response timeliness.

[0047] After calculating the real-time risk coefficients of each security constraint parameter, the system sorts these parameters from highest to lowest risk coefficient. This sorting process aims to establish a dynamic priority system, ensuring that the most critical security requirements are met first when multiple constraints exist simultaneously or may conflict. Setting the security constraint parameter with the highest risk coefficient as the highest priority constraint means that this constraint is crucial to the safe operation of the system at the current moment, and its control command will have the highest execution weight.

[0048] When the system detects a conflict between control commands for any two safety constraint parameters—for example, one constraint requires increasing assist while the other requires decreasing assist, or one requires a rapid response while the other requires a smooth transition—the system will execute the command of the highest-priority constraint. For lower-priority constraints that conflict with the highest-priority constraint, their control commands will be reduced or delayed. Reduction refers to narrowing the strength or scope of the lower-priority constraint's command to within the allowable range of the higher-priority constraint. Delay refers to postponing the execution of a lower-priority constraint's command if its control cannot be immediately satisfied but its importance is relatively low, waiting for the higher-priority constraint to be satisfied. In this way, the system can effectively resolve conflicts between constraints, ensuring that all control commands are compatible, ultimately forming an optimized set of safety constraints. This optimized set of safety constraints not only considers the real-time risks of each constraint but also guarantees the coordination and safety of control commands through priority and conflict resolution mechanisms.

[0049] Reference Figure 4 As shown, based on the optimized safety constraint set, this application proposes to generate an initial assist path through path planning, including: according to the current position of the human foot and the target work point position in the current scene description information, combined with the joint angle limit parameters and foot landing safety area parameters in the optimized safety constraint set, a dynamic programming method is used to search for the optimal trajectory from the current position to the target work point on the terrain grid map; the optimal trajectory takes the shortest path length, the smallest change in terrain slope, and the highest foot landing safety as the optimization objectives, and generates a sequence containing a series of foot landing coordinates and corresponding joint driving timing as the initial assist path.

[0050] Specifically, the current position of the human foot refers to the precise spatial coordinates of the user's foot at the current moment. This position information can be collected in real time by sensors such as foot pressure sensor arrays and inertial measurement units integrated into the sole or foot of the flexible exoskeleton, and estimated with high accuracy using sensor fusion algorithms (such as extended Kalman filtering). The target work point location refers to the specific spatial point that the user expects to reach. It can be pre-set, input through a user interface (such as a touchscreen or voice command), or automatically determined by an environmental perception module (such as a depth camera or LiDAR) after semantic segmentation and target recognition of the work environment. These two locations together constitute the starting and ending points of the path planning.

[0051] During path planning, the pre-assessed and optimized safety constraints must be strictly adhered to. Joint angle limit parameters define the maximum flexion and extension angles that each joint of the flexible exoskeleton can achieve during movement, preventing damage to the mechanical structure or causing discomfort or even injury to the human body. Foot landing safety zone parameters define the permissible safe space range for the foot when it lands, ensuring that the foot does not land on unstable terrain, obstacles, or dangerous edges. These parameters serve as hard constraints on the path planning algorithm, ensuring that the generated path meets safety requirements at all times.

[0052] Dynamic programming is an effective pathfinding algorithm that decomposes the complex path planning problem into a series of subproblems and avoids redundant computation by storing the solutions to these subproblems, thus efficiently finding the global optimum. A terrain raster map is a discretized representation of the current scene environment, abstracting continuous 3D terrain information into a 2D grid. Each grid cell can store its corresponding terrain slope, accessibility, obstacle information, and other attributes. The path planning algorithm searches this raster map, progressively constructing a path by evaluating the cost from one grid cell to an adjacent cell.

[0053] The optimal trajectory is defined as a multi-objective optimization problem, aiming to balance efficiency, comfort, and safety. Minimizing path length aims to improve movement efficiency and reduce energy consumption. Minimizing terrain slope variation aims to ensure a smooth journey, reducing discomfort and the risk of falls. Maximizing foot placement safety ensures that each step is taken within a safe zone, avoiding potential dangers. In dynamic programming algorithms, these optimization objectives are represented by a comprehensive cost function; for example, the cost of each path segment can be obtained by weighted summing of its length, slope variation, and a penalty term for foot placement safety.

[0054] Finally, a sequence containing a series of foot landing coordinates and corresponding joint drive timings is generated as the initial assist path. The initial assist path is not merely a spatial geometric trajectory; it is a complete sequence containing time and motion commands. The foot landing coordinates define the precise landing position of the foot for the user of the flexible exoskeleton in each step. Based on this, by combining the human gait model and the kinematic and dynamic models of the flexible exoskeleton, the drive timing data of each joint (such as the hip, knee, and ankle joints) at different time points, including angles, angular velocities, and torques, required to achieve these foot landings can be further calculated.

[0055] During the execution of assist control, real-time terrain point cloud data and the actual foot landing position of the human body are continuously monitored; the actual foot landing position is compared with the expected landing position of the corresponding gait in the initial assist path, and the landing deviation is calculated; if the landing deviation exceeds the preset safety deviation threshold, the path planning is re-executed based on the current measured terrain conditions, starting from the current actual foot landing position, to generate a corrected local assist path, and the corrected local assist path is spliced ​​with the remaining unexecuted initial assist path to form a feasible assist path sequence.

[0056] Real-time terrain point cloud data refers to a dataset continuously collected by environmental perception sensors such as depth cameras or LiDAR, reflecting the three-dimensional geometric features of the current work area. This data provides information such as real-time terrain undulations and obstacle distribution, and is crucial for dynamic environmental perception. "The actual foot placement position" refers to the precise spatial coordinates of the human foot's actual contact with the ground during the assistance process of the flexible exoskeleton device. This position can be calculated in real time using a pressure sensor array integrated into the sole of the foot, an inertial measurement unit (IMU) combined with a kinematic model, or by tracking the foot through an external vision system. Continuous monitoring ensures that the system can promptly grasp the real-time state of human-environment interaction.

[0057] The expected landing point of the foot in the initial assist path is determined based on the pre-planned initial assist path, representing the target position that the human foot should reach within a specific gait cycle. Landing point deviation refers to the spatial distance or vector difference between the actual foot landing point and the expected landing point. This deviation can be calculated using Euclidean distance, Manhattan distance, or other suitable distance metrics to quantify the degree of inconsistency between the actual movement and the planned path.

[0058] The preset safety deviation threshold is a pre-defined, allowable upper limit for foot landing deviation. This threshold is typically determined based on a combination of factors, including the stability of the exoskeleton device, human comfort, terrain complexity, and safety margin. When the calculated landing deviation exceeds this threshold, it indicates that the current path is no longer safe or optimal and requires adjustment. The current measured terrain conditions refer to the environmental information reflected in the latest real-time terrain point cloud data obtained through continuous monitoring when path replanning is triggered. Re-executing path planning from the current actual foot landing position means that the system no longer uses the original initial assist path, but instead uses the current actual position of the human foot as the new planning starting point. Re-executing path planning can employ methods similar to those used to generate the initial assist path, such as dynamic programming, but its planning scope is usually a local area, with the goal of generating the optimal trajectory from the current actual foot landing to the next safe landing point or target point. During the planning process, the joint angle limit parameters and foot landing safety area parameters in the optimized safety constraint set will be considered again, with the optimization objectives being the shortest path length, the smallest change in terrain slope, and the highest foot landing safety. The "revised local assistance path" is a new assistance trajectory adapted to the current environment, obtained through replanning and starting from the current actual foot landing point.

[0059] The unexecuted remaining initial assist path refers to the portion of the initial assist path that has not yet been executed when replanning is triggered. Stitching refers to connecting the newly generated, corrected local assist path with the unexecuted remaining initial assist path to form a complete and continuous assist path. The stitching process needs to ensure a smooth transition of the path, avoiding abrupt changes at the connection points. This can be achieved through path smoothing algorithms or spline interpolation techniques. The feasible assist path sequence is the final path, dynamically adjusted, that can guide the flexible exoskeleton device in assist control.

[0060] In some of the embodiments described above in this application, an enhanced strategy model is proposed to output assisted behavior prediction results. However, in its implementation, the key to achieving efficient human-machine collaborative control lies in how to accurately and in real time identify human movement intentions from complex human physiological and motion signals and transform them into precise assisted behavior predictions.

[0061] In this regard, refer to Figure 5As shown, this application further proposes a step of combining human-computer interaction signals and human motion intention feedback to construct an enhanced strategy model through fusion processing, which is used to output the predicted result of assisted behavior. The steps include: collecting human motion intention feedback signals, including: muscle activation timing signals collected by electromyography sensors, plantar pressure distribution signals collected by plantar pressure sensors, and human center of mass offset direction signals collected by inertial measurement units; inputting the human-computer interaction signals and human motion intention feedback signals into a multilayer perceptron network, which uses historical motion data as training samples to establish a mapping relationship between the current human motion intention and the expected assisted output; and calculating the expected assisted torque, expected assisted timing, and expected assisted duration in real time based on the currently input signals, as the predicted result of assisted behavior.

[0062] Human movement intention feedback signals are crucial for understanding a user's current and upcoming movement state. Electromyography (EMG) sensors capture electrophysiological signals generated during muscle activity; their muscle activation timing signals reflect the initiation, intensity, and duration of muscle contractions, directly indicating the user's movement intention. Plantar pressure sensors measure the pressure distribution when the foot contacts the ground; their plantar pressure distribution signals reveal the shift of the body's center of gravity, gait cycle characteristics, and forces exerted on the ground, thus inferring the user's movement patterns and intentions. Furthermore, inertial measurement units (IMUs) measure acceleration and angular velocity to provide signals indicating the direction of the body's center of gravity shift. This signal is essential for assessing the overall movement trend and stability of the body, helping to predict the user's next movement direction.

[0063] In addition to the aforementioned human motion intention feedback signals, human-computer interaction signals can also include commands directly input by the user through other means (such as user interface, voice commands, etc.), as well as the status information of the flexible exoskeleton device itself. These signals, together with the human motion intention feedback signals, constitute the input data for a comprehensive understanding of the user's intention and the device's operating status. As a type of feedforward artificial neural network, the multilayer perceptron network (MLPF) possesses the ability to learn complex nonlinear mapping relationships. This network uses pre-collected historical motion data as training samples, which include the user's physiological signals, human-computer interaction commands, and corresponding ideal assist output parameters in different motion scenarios. Through training, the MPF can learn and establish a complex mapping relationship between these input signals and the desired assist output, enabling it to accurately infer the user's current motion intention and predict the required assistance when receiving real-time input.

[0064] In this embodiment, the multilayer perceptron network includes an input layer, at least two hidden layers, and an output layer. The number of nodes in the input layer is the same as the dimension of the fused feature vector, ensuring the network can fully receive and process all input feature information. The number of nodes in the hidden layers is designed to decrease layer by layer. This structure helps the network gradually extract and abstract high-level semantic information from the input features, while also avoiding overfitting to some extent and improving the model's generalization ability. The number of nodes in the output layer is set to three, precisely corresponding to the desired assist torque, desired assist timing, and desired assist duration required by the flexible exoskeleton device. These parameters collectively define the assist behavior of the exoskeleton.

[0065] To enable this multilayer perceptron network to accurately predict assistive behaviors, it needs to be pre-trained using motion data of humans under various terrain conditions and corresponding optimal assistive parameters. During training, the backpropagation algorithm is used to adjust the network's weights and biases. The backpropagation algorithm calculates the error between the network output and the true optimal assistive parameters and propagates this error back along the network to update the network parameters, thereby minimizing the mean squared error between the network's expected assistive parameters and the optimal assistive parameters in the training samples. This training method ensures that the network can learn the complex nonlinear mapping relationship between human movement intentions and optimal assistive behaviors from a large amount of historical data.

[0066] After receiving real-time human-computer interaction signals and human motion intention feedback signals, the multilayer sensor network can quickly calculate and output the predicted results of the assisted behavior in real time. These prediction results specifically include the expected assist torque, the expected assist timing, and the expected assist duration. The expected assist torque refers to the magnitude of the assist torque that each joint of the flexible exoskeleton needs to provide to support or enhance the user's movement. The expected assist timing refers to the precise point in time when the exoskeleton begins or ends providing assistance, ensuring synchronization between assistance and human movement. The expected assist duration refers to the effective duration of the assist effect. These three parameters together constitute a complete assist command, guiding the actuators of the flexible exoskeleton to perform precise and coordinated assist control to achieve coordinated movement highly matched with the user's intentions.

[0067] However, during implementation, the original sensor signals may contain a lot of redundant information or noise and have a high dimensionality. Directly inputting them into the multilayer perceptron network may lead to low model training efficiency and difficulty in accurately capturing the complex movement intentions of the human body, thereby affecting the accuracy and real-time performance of assisted behavior prediction.

[0068] To address this, this application further proposes that before inputting the human-computer interaction signals and the human motion intention feedback signals into the multilayer perceptron network, it is necessary to construct an input feature vector. Specifically, this construction process includes: segmenting the muscle activation time-series signals collected by the electromyography sensor according to the acquisition time window, and extracting the peak value, mean value, rising slope, and duration features of muscle activation for each muscle group within each time window to form a muscle activation feature vector; dividing the plantar pressure distribution signals collected by the plantar pressure sensor into the heel area, arch area, and forefoot area according to the foot region, and calculating the pressure peak value, pressure center trajectory, and grounding time ratio for each area to form a plantar pressure feature vector; decomposing the human center of mass offset direction signal collected by the inertial measurement unit into the forward and backward offset and the left and right offset, and combining the change rate of the human center of mass offset direction signal to form a center of mass motion feature vector; finally, concatenating the muscle activation feature vector, the plantar pressure feature vector, and the center of mass motion feature vector to form a fused feature vector, which serves as the input layer node data for the multilayer perceptron network.

[0069] This method involves segmenting the muscle activation time-series signal acquired by an electromyography (EMG) sensor according to the acquisition time window. Within each time window, the peak value, mean value, ramp rate, and duration of muscle activation for each muscle group are extracted to form a muscle activation feature vector. The aim is to extract key information from the raw EMG signal that effectively characterizes the intensity, duration, and dynamic changes of muscle activity. EMG signals, as direct physiological signals reflecting human movement intentions, are typically high-frequency, high-dimensional time-series data. Directly using them as model input increases computational burden and may introduce noise. By setting an appropriate acquisition time window (e.g., 50 ms to 200 ms), segmenting the signal, and calculating the peak value (reflecting maximum activation intensity), mean value (reflecting average activation level), ramp rate (reflecting the initial velocity or explosive force of muscle contraction), and duration (reflecting the persistence of muscle activation) within each window, the complex raw signal can be transformed into more representative and discriminative low-dimensional features, thereby more accurately capturing human movement intentions.

[0070] Plantar pressure distribution signals collected by plantar pressure sensors are divided into heel, arch, and forefoot regions. The peak pressure, pressure center trajectory, and contact time percentage are calculated for each region to form a plantar pressure feature vector. This aims to quantify the biomechanical characteristics of the foot-ground interaction, reflecting postural balance, gait cycle, and weight transfer. Plantar pressure distribution is a crucial indicator for assessing standing and walking stability. By dividing the foot into biomechanically significant regions and extracting the peak pressure (indicating localized force concentration), pressure center trajectory (reflecting the dynamic movement of the center of gravity on the foot), and contact time percentage (indicating the duration of contact between different foot regions and the ground, closely related to the gait cycle) for each region, high-dimensional plantar pressure distribution maps can be effectively transformed into refined feature vectors, providing crucial information for determining the body's movement state and intentions.

[0071] The human center of mass offset direction signal acquired by the inertial measurement unit is decomposed into forward / backward offset and left / right offset. Combined with the rate of change of this offset signal, a center of mass motion feature vector is formed. This vector aims to accurately describe the dynamic position and motion trend of the human center of mass in space, which is the core basis for assessing human balance and predicting the next direction of movement. The human center of mass offset direction signal directly reflects the overall stability of the body. By decomposing it into forward / backward and left / right offsets, the degree of tilting or movement of the body in different planes can be clearly quantified. Further combining the rate of change of these offsets (i.e., center of mass velocity or acceleration) allows for the capture of trends in the dynamic changes of the human center of mass, such as an impending forward step, lateral adjustment of balance, or deceleration and stop, thus providing the exoskeleton with a proactive decision-making basis.

[0072] Muscle activation feature vectors, plantar pressure feature vectors, and center-of-mass motion feature vectors are concatenated to form a fused feature vector, which serves as the input layer node data for a multilayer perceptron network. This aims to integrate information from different physiological and biomechanical dimensions to form a comprehensive, multimodal input representation. This fusion of multi-source information can overcome the limitations of single-sensor data, providing a deeper understanding of human movement intentions. By concatenating these preprocessed and feature-extracted vectors, a moderately dimensional and information-rich fused feature vector can be constructed. This vector, as input to the multilayer perceptron network, provides the network with a structured and highly discriminative data representation, thereby optimizing the network's learning efficiency and prediction performance.

[0073] Reference Figure 6As shown, this application further proposes that when the predicted result of the assistive behavior meets the dynamic operation requirements, a sequence of control commands is generated and output to the actuator of the flexible exoskeleton for assistive control. The dynamic operation requirements include: minimum assist efficiency requirements in flat walking mode, maximum assist torque requirements in uphill climbing mode, braking damping requirements in downhill descent mode, foot lift height requirements in obstacle crossing mode, and lumbar support torque requirements in load-bearing transport mode.

[0074] Specifically, dynamic operational requirements refer to the specific performance indicators or parameter ranges that flexible exoskeleton devices must meet to effectively assist human movement in different operational scenarios or movement modes. These requirements are determined comprehensively based on factors such as the nature of the work, environmental characteristics, and the physiological load on the human body. For example, the minimum assist efficiency requirement in flat walking mode aims to ensure that the exoskeleton can minimize energy consumption while maintaining gait smoothness when providing assistance; the maximum assist torque requirement in uphill climbing mode emphasizes that the exoskeleton can provide sufficient driving force to assist the human body in completing climbing movements when facing gravity challenges; the braking damping requirement in downhill descent mode focuses on providing stable counterforce during descent to prevent the human body from accelerating out of control; the foot lift height requirement in obstacle crossing mode ensures that the feet can safely cross obstacles and avoid tripping; and the lumbar support torque requirement in load-bearing carrying mode aims to reduce the burden on the lumbar region when carrying heavy objects and protect the spine. These requirement parameters can be pre-set using experimental data, biomechanical models, or expert experience and stored in the system.

[0075] To ensure that the predicted assistive behavior accurately meets these dynamic operational needs, the method of this application includes identifying the current operational mode based on the current scene description information and retrieving the corresponding requirement parameters from the dynamic operational requirements. The identification of the current operational mode can be achieved by analyzing various sensor data, such as terrain slope, human movement speed, and muscle activation level, from the current scene description information. For example, when the system detects a continuous uphill slope, high muscle activation level, and a specific gait pattern, it can determine that the current state is an "uphill climbing mode." Once the operational mode is identified, the system will accurately retrieve the requirement parameters corresponding to that mode from a pre-set dynamic operational requirement library, such as the maximum assist torque requirement in the uphill climbing mode.

[0076] Subsequently, the expected assist torque, expected assist timing, and expected assist duration from the assist behavior prediction results are compared with the aforementioned demand parameters. This is a crucial verification step designed to ensure that the assist behavior predicted by the enhancement strategy model not only aligns with human movement intentions but also quantitatively meets the specific requirements of the current work mode. For example, the system compares the predicted expected assist torque with the maximum assist torque requirement in an uphill climbing mode to check if it falls within the allowable range. Similarly, a similar comparison is made with the expected assist timing and expected assist duration to ensure that assistance is provided at the right time for the appropriate duration. When the expected assist torque, expected assist timing, and expected assist duration all fall within the allowable range defined by the demand parameters, the system determines that the assist behavior prediction results meet the dynamic work requirements. Only when all these conditions are met will the system generate the corresponding control command sequence and output it to the actuators of the flexible exoskeleton for assist control.

[0077] To more clearly illustrate the specific implementation process of the technical solution of this invention, a detailed description is provided below in conjunction with a specific application scenario. The scenario is set as follows: a construction worker is at a power transmission line construction site in a mountainous area, wearing a flexible exoskeleton device, carrying a 20kg tower material on his back, and walking uphill on a gravel slope with a gradient of 18 degrees. The target work point is located at a position 15 meters away horizontally and approximately 4.8 meters above the current height.

[0078] In this scenario, after the construction workers activated the flexible exoskeleton, the system began collecting data from multiple sensors in real time. Inertial measurement units (IMUs) were installed at various joints of the human torso and exoskeleton, collecting data at a frequency of 150Hz. At the current moment, the IMU output the following attitude angle data: torso pitch angle 18.5 degrees (forward tilt), roll angle 2.1 degrees (slightly tilted to the left), yaw angle 0 degrees (facing the direction of travel); joint angle data were: hip flexion 25 degrees, knee flexion 15 degrees, and ankle dorsiflexion 8 degrees. Electromyography (EMG) sensors were attached to major muscle groups such as the rectus femoris, gastrocnemius, and erector spinae, collecting muscle activation signals at a frequency of 1000Hz. After preprocessing, the muscle activation level of the rectus femoris was 0.72 (normalized value, range 0-1), the gastrocnemius activation level was 0.45, and the erector spinae activation level was 0.68. The depth camera is mounted on the waist of the exoskeleton and collects three-dimensional point cloud data of the terrain in front at a frequency of 30Hz. After point cloud segmentation and plane fitting, the terrain features within a 3-meter range in front are extracted: the ground slope angle is 18.2 degrees, the road surface smoothness variance is 0.03 meters, and there is a protruding rock about 8 centimeters high 2 meters in front.

[0079] Before fusion processing, the system first performs time synchronization and spatial alignment of the multi-source data. Since the sampling frequencies of each sensor are different, a hardware-triggered method is used, with the unified clock of the exoskeleton control system as the reference, to timestamp each frame of sensor data. Linear interpolation is then used to interpolate the low-frequency sensor data to the sampling time of the high-frequency sensor. Specifically, the time interval between two adjacent frames of depth camera data is 33.3 milliseconds. Based on the timestamp of the inertial measurement unit (IMU), the system calculates the terrain parameters of the interpolation points using the time ratio of the two adjacent depth frames. For spatial alignment, a pre-calibrated coordinate system transformation matrix is ​​used to transform the IMU data from the sensor coordinate system to the world coordinate system with the human body's center of mass as the origin, and the depth camera's point cloud data from the camera coordinate system to the same world coordinate system.

[0080] After time synchronization and spatial alignment are completed, multi-source data is input into a Kalman filter for state estimation. The state vector of the Kalman filter includes the three-dimensional position and velocity of the human body's center of mass, the attitude angle and angular velocity of the human torso, and parameters such as the slope angle of the terrain in front, the height of the steps, and the distance to obstacles. The initial state is set as follows: starting position (0,0,0), starting velocity (0.8m / s,0,0.2m / s), attitude angle (0,18.5°,0), and terrain parameters (α=18.2°,h=0m,d=2m). In the prediction step, the filter uses the state equation, based on the optimal state estimate of the previous moment, combined with the exoskeleton assist torque as the control input, to predict the state at the current moment. Assuming that the center of mass velocity at the previous moment was 0.78m / s, the acceleration was 0.05m / s², and the time step Δt=0.0067 seconds (corresponding to 150Hz), the predicted current position is px=0.78×0.0067≈0.0052 meters. In the update step, the filter incorporates the actual measurements from each sensor at the current moment and calculates the Kalman gain. In this scenario, prediction uncertainty is low, observation noise is moderate, and the Kalman gain is approximately 0.6, meaning the system assigns about 60% confidence to the observations. Subsequently, the actual and predicted observations at the current moment are compared, the observation residuals are calculated, and the prior estimate is weighted and corrected using the Kalman gain to obtain the posterior optimal state estimate. The corrected optimal state estimate for the current moment is output, which includes the following scene description information: human center of mass position (0.132, 0, 0.051) meters, center of mass velocity (0.82 m / s, 0, 0.21 m / s), torso attitude angles (roll 2.0 degrees, pitch 18.4 degrees, yaw 0 degrees), terrain slope angle ahead 18.3 degrees, and a bulge 8.2 cm high located 2.1 meters ahead.

[0081] The system determines safety constraint parameters based on the current scene description information. In this embodiment, based on the construction worker's body parameters (height 175 cm, weight 75 kg) and the design parameters of the exoskeleton device, the following safety constraint limits are preset: For joint angle limiting parameters, the maximum hip flexion is 120 degrees and the minimum extension is -10 degrees; the maximum knee flexion is 135 degrees and the minimum extension is 0 degrees; the maximum ankle dorsiflexion is 20 degrees and the minimum plantarflexion is 15 degrees. For joint torque limiting parameters, the maximum assist torque for the hip joint is 120 Nm, the maximum assist torque for the knee joint is 100 Nm, and the maximum assist torque for the ankle joint is 60 Nm. For human body center of gravity stability boundary parameters, the safety margin of the supporting polygon boundary is 5 cm, meaning the center of gravity projection is at least 5 cm from the boundary. For foot landing safety zone parameters, the minimum safe distance between the landing point and the terrain edge or obstacle is 10 cm. For assist output response time parameters, the maximum allowable delay from intention recognition to assist output is 150 milliseconds.

[0082] The system calculates various risk coefficients based on current scene information (terrain slope 18.3 degrees, human movement speed 0.82 m / s, and rectus femoris muscle activation 0.72). For the joint angle limit parameter, the current hip flexion is 25 degrees, 95 degrees different from the limit of 120 degrees, indicating sufficient safety margin; the risk coefficient calculated using the exponential mapping function is 0.05. For the joint torque limit parameter, the estimated assist torque required for uphill walking based on the exoskeleton dynamics model is approximately 35 Nm, significantly different from the maximum limit of 120 Nm; the calculated risk coefficient is 0.12. For the human body center of gravity stability boundary parameter, the current human body center of gravity projection point is located within the supporting polygon, 12 cm from the front boundary, 20 cm from the rear boundary, and 15 cm from each of the left and right boundaries. The preset safety margin is 5 cm, and the current minimum distance is 12 cm; the calculated risk coefficient is 0.15. Regarding the foot-landing safety zone parameter, the predicted next foot-landing point is approximately 0.6 meters ahead, 15 centimeters from the edge of the nearest pile of loose rocks. The preset safety distance is 10 centimeters, resulting in a calculated risk coefficient of 0.08. For the assist output response time parameter, the current system's actual response delay is 120 milliseconds, with a preset maximum allowable delay of 150 milliseconds. Using a ratio, the calculated risk coefficient is 0.8. However, due to the urgent need for assistance and high muscle activation (0.72) when walking uphill, the system weights and amplifies the risk coefficient, ultimately correcting it to 0.85.

[0083] The five risk coefficients are ranked from highest to lowest: assist output response time parameter (0.85), human body center of gravity stability boundary parameter (0.15), joint torque limit parameter (0.12), foot landing safe zone parameter (0.08), and joint angle limit parameter (0.05). Therefore, the assist output response time parameter is set as the highest priority constraint, requiring the exoskeleton to complete the entire process from intent recognition to assist output within 150 milliseconds. The system checks for conflicting constraint instructions. The current assist output response time parameter requires rapid assist output (expected delay ≤ 150 milliseconds), while the joint torque limit parameter requires a smooth torque output increase to avoid impact (requiring approximately 200 milliseconds of rise time), creating a contradiction. Based on the highest priority (response time), lower priority constraints are reduced: the rise time requirement for the joint torque limit parameter is reduced from 200 milliseconds to 130 milliseconds, while the torque change rate is limited to a safe range by adjusting the torque gradient. After three iterations, the constraint instructions are compatible, forming an optimized set of safe constraints.

[0084] In the path planning phase, the system constructs a terrain grid map covering a 10m x 2m area ahead, with each grid cell measuring 0.1m x 0.1m. Each grid cell stores terrain slope (range 15-20 degrees), drivability (slope less than 25 degrees is drivable), and obstacle information (a protruding rock 2.1m ahead is marked as a avoidance zone). The current foot position is at coordinates (0,0), and the target work point is located at (15,4.8) (horizontal distance 15 meters, vertical elevation 4.8 meters). A dynamic programming method is used to search for the optimal trajectory. In the cost function, path length cost has a weight of 0.3, terrain cost has a weight of 0.4 (smaller slope changes are better), and safety cost has a weight of 0.3 (the farther the foot landing point is from obstacles, the better). Dynamic programming expands the search layer by layer from the starting point. The possible landing area for each step is limited by stride length (maximum stride 0.8 meters, minimum stride 0.3 meters). The cumulative cost of each candidate state is calculated, and the minimum value is retained. After approximately 1500 state expansions, the search was completed, and the optimal trajectory was obtained through backtracking. This trajectory included 15 foot landing points with an average stride length of 0.5 meters. The specific landing point coordinate sequence was: (0.5, 0.15), (1.0, 0.32), (1.5, 0.48)... up to (15, 4.8). Simultaneously, joint drive timing was generated. Based on the gait model, the expected angle curves for the hip, knee, and ankle joints were calculated for each landing point. For example, the peak hip angle for the first landing point was 30 degrees, and the peak knee angle was 20 degrees. The assist torque output timing was 100 milliseconds for the rise phase, 80 milliseconds for the peak maintenance phase, and 100 milliseconds for the fall phase.

[0085] During the execution of the assist control, the system continuously monitors real-time terrain point cloud data. After the construction personnel complete step 5, the system detects a deviation between the actual foot landing position and the expected landing point: the expected landing point is (2.5, 0.80), and the actual landing point is (2.55, 0.78), with a longitudinal deviation of 5 cm and a lateral deviation of 2 cm. The preset safety deviation thresholds are 3 cm longitudinally and 2 cm laterally. The current longitudinal deviation of 5 cm exceeds the threshold, triggering local path replanning. The system uses the current actual position (2.55, 0.78) as the starting point and re-executes dynamic planning, covering the remaining 10 expected landing points, generating a corrected local path containing 10 new landing point coordinates. The corrected path is spline interpolated and stitched with the remaining initial path, maintaining positional continuity and first derivative continuity at the connection point to form a new feasible assist path sequence.

[0086] In the enhancement strategy model construction phase, the system collects human motion intention feedback signals. Features are extracted using a 200-millisecond time window: peak activation of the rectus femoris muscle is 0.75, mean 0.68, rise slope 3.2 / second, duration 180 ms; peak activation of the gastrocnemius muscle is 0.48, mean 0.42, rise slope 2.1 / second, duration 160 ms; peak activation of the erector spinae muscle is 0.71, mean 0.65, rise slope 2.8 / second, duration 190 ms, forming a muscle activation feature vector with a dimension of 18 (6 muscle groups × 3 features). The plantar pressure sensor collects the following data: peak pressure in the heel area is 8.5 N / cm², pressure center trajectory offset is 3 mm, and ground contact time accounts for 40%; peak pressure in the arch area is 3.2 N / cm², pressure center trajectory offset is 2 mm, and ground contact time accounts for 20%; peak pressure in the forefoot area is 12.5 N / cm², pressure center trajectory offset is 5 mm, and ground contact time accounts for 40%, forming a plantar pressure feature vector with a dimension of 9 (3 areas × 3 features). The inertial measurement unit (IMU) collects the following data on the centroid offset: forward offset of 6.2 cm, left offset of 1.8 cm, and centroid offset rate of change of 0.15 m / s² in the forward direction and 0.03 m / s² in the left direction, forming a centroid motion feature vector with a dimension of 4. These three feature vectors are then concatenated to form a fused feature vector with a total dimension of 31 (18+9+4).

[0087] The multilayer perceptron network structure consists of: 31 nodes in the input layer, 16 nodes in the first hidden layer, 8 nodes in the second hidden layer, and 3 nodes in the output layer (corresponding to the expected assist torque, expected assist timing, and expected assist duration, respectively). The network has been pre-trained, with a training dataset containing 5000 samples, covering 1000 samples each for five different working conditions: flat ground, uphill, downhill, traversing, and carrying load. The fused feature vector is input into the network and calculated via forward propagation: a weighted sum is applied from the input layer to the first hidden layer, followed by a ReLU activation function; the same process is applied from the first hidden layer to the second hidden layer; and linear activation is applied from the second hidden layer to the output layer. The output results are: expected assist torque 48 Nm for the hip joint, 32 Nm for the knee joint, and 15 Nm for the ankle joint; the expected assist timing is output 120 milliseconds from the current moment; and the expected assist duration is 380 milliseconds.

[0088] The system identifies the operating mode based on the current scene description: terrain slope 18.3 degrees (greater than 10 degrees), human movement speed 0.82 m / s (moderate), high muscle activation (rectus femoris 0.72), and trunk forward lean 18.4 degrees, comprehensively judging it as "uphill climbing mode". From the uphill climbing mode requirement parameter library, the following parameters are retrieved: maximum assist torque requirement: hip joint ≤ 110 Nm, knee joint ≤ 90 Nm, ankle joint ≤ 55 Nm; minimum assist efficiency requirement ≥ 65%; allowable assist timing range: -20 milliseconds to +50 milliseconds (relative to the expected gait phase); allowable assist duration range: 300 milliseconds to 500 milliseconds.

[0089] The predicted results were compared with the required parameters: the expected assist torque was 48 Nm ≤ 110 Nm for the hip joint, 32 Nm ≤ 90 Nm for the knee joint, and 15 Nm ≤ 55 Nm for the ankle joint, all of which were met; the expected assist timing was 120 milliseconds, the ideal timing corresponding to the gait phase was 100 milliseconds, and the deviation was +20 milliseconds, which was within the range of -20 milliseconds to +50 milliseconds, thus meeting the requirements; the expected assist duration was 380 milliseconds, which was within the range of 300 milliseconds to 500 milliseconds, also meeting the requirements. Since all three parameters fell within the allowable range of the required parameters, the system determined that the predicted assist behavior met the dynamic operation requirements.

[0090] The system generates a control command sequence based on this: target torques for each joint are 48 Nm for the hip joint (incline type, ascent time 20 ms), 32 Nm for the knee joint (step type, instantaneous response), and 15 Nm for the ankle joint (exponential ascent type, time constant 15 ms); the timing commands are triggered at t=0 ms, with the hip joint starting output at t=20 ms, the knee joint starting output at t=25 ms, and the ankle joint starting output at t=18 ms; the commands are updated every 50 ms control cycle. The control command sequence is output to the actuators of each joint of the exoskeleton, and the exoskeleton provides assistance according to the commands to help the construction worker complete the uphill walking.

[0091] In the closed-loop feedback loop, the system continuously collects feedback data during the execution of control commands by the actuator. At the end of the current control cycle, the actual output torque at the hip joint is 50 Nm (expected 48 Nm, deviation +2 Nm), the actual joint angle is the peak hip joint angle of 32 degrees (expected 30 degrees, deviation +2 degrees), and the actual response time is 125 milliseconds (expected 120 milliseconds, deviation +5 milliseconds). The preset deviation thresholds are: torque deviation not exceeding 5 Nm, angle deviation not exceeding 3 degrees, and time deviation not exceeding 10 milliseconds. Since all deviations are currently within the allowable range, the system determines there is no significant deviation, does not trigger a reassessment, and continues to execute according to the current strategy.

[0092] If the actual response time delay reaches 165 milliseconds in another gait cycle, exceeding the preset threshold of 10 milliseconds, the system will immediately trigger a reassessment of the safety constraint set. When recalculating the various risk coefficients, due to the excessive response delay, the risk coefficient of the assist output response time parameter rises to 0.95, becoming the highest priority constraint. Based on this constraint, the system reduces conflicting low-priority constraints (such as joint torque limit parameters), forcing subsequent assist response times to be within 150 milliseconds, while appropriately adjusting the torque output curve to ensure safety. The corrected constraint parameters are fed back to the main control loop, and subsequent assist behavior is adjusted based on these corrected constraints. Through this closed-loop feedback mechanism, the system can continuously monitor the execution effect, promptly detect deviations, and adaptively correct them, ensuring an optimal balance between safety and efficiency under different operating conditions.

[0093] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for adaptive human-robot collaborative control with fusion safety constraints, characterized in that, The application to flexible exoskeleton devices includes the following steps: Acquire real-time environmental data and human motion status information, and generate current scene description information through sensor fusion processing; Based on the current scenario description information, determine at least one security constraint parameter, and perform priority evaluation and conflict resolution on each security constraint parameter to generate an optimized security constraint set; Based on the optimized set of safety constraints, an initial assistance path is generated through path planning, and the initial assistance path is iteratively adjusted according to the real-time terrain conditions to form a sequence of feasible assistance paths. Based on the feasible assistance path sequence, combined with human-computer interaction signals and human movement intention feedback, an enhancement strategy model is constructed through fusion processing to output the prediction results of assistance behavior; When the prediction results of the assistive behavior meet the dynamic operation requirements, a sequence of control instructions is generated and output to the actuator of the flexible exoskeleton for assistive control. The system acquires feedback data from the actuators and compares it with the predicted results of the assist behavior. If there is a discrepancy between the two, it triggers a reassessment and correction of the safety constraint set and feeds the corrected constraint parameters back to the main control loop for dynamic adjustment of subsequent assist behaviors. 2.The method of claim 1, wherein: Real-time environmental data and human motion status information are acquired and processed through sensor fusion to generate current scene description information, including: The inertial measurement unit collects the attitude angle data and angular velocity data of the human torso and limbs, the electromyography sensor collects the muscle activation signals of the main muscle groups of the human body, and the depth camera or lidar collects the three-dimensional point cloud data of the terrain in front and the ground slope information. The attitude angle data, angular velocity data, muscle activation signals, 3D point cloud data and ground slope information are input into the Kalman filter for time synchronization and data fusion, and output a multi-dimensional state vector representing the current human pose, motion trend and terrain features as the current scene description information. 3.The method of claim 1, wherein: Safety constraint parameters include: joint angle limit parameters, joint torque limit parameters, human body center of gravity stability boundary parameters, foot landing point safety zone parameters, and assist output response time parameters; Among them, the joint angle limiting parameter is used to limit the maximum flexion and extension angle of each joint of the exoskeleton, the joint torque limiting parameter is used to limit the maximum assist torque output by each joint of the exoskeleton, the human body center of gravity stability boundary parameter is used to characterize the safe offset range of the human body center of gravity within the supporting polygon, the foot landing point safe area parameter is used to characterize the safe distance of the foot landing point relative to the terrain edge or obstacle, and the assist output response time parameter is used to characterize the maximum allowable delay time from intent recognition to assist output.

4. The adaptive human-robot collaborative control method fusing safety constraints according to claim 3, characterized in that: Prioritize and resolve conflicts among the various security constraint parameters to generate an optimized set of security constraints, including: Based on the terrain slope, human movement speed, and muscle activation level in the current scene description information, calculate the real-time risk coefficients of each safety constraint parameter; The safety constraint parameters are sorted from high to low according to the real-time risk coefficient, and the safety constraint parameter with the highest risk coefficient is set as the highest priority constraint. When there is a contradiction between the control instructions of any two safety constraint parameters, the control instruction of the highest priority constraint shall be used as the basis for execution, and the control instructions of the lower priority constraints shall be reduced or delayed until the control instructions of all constraints are compatible with each other, thus forming an optimized set of safety constraints.

5. The adaptive human-robot collaborative control method fusing safety constraints according to claim 3, characterized in that: Based on the optimized safety constraint set, an initial assist path is generated through path planning, including: Based on the current position of the human foot and the target work point in the current scene description information, and combined with the joint angle limit parameters and foot landing safety area parameters in the optimized safety constraint set, a dynamic programming method is used to search for the optimal trajectory from the current position to the target work point on the terrain grid map. The optimal trajectory aims to minimize path length, terrain slope variation, and foot landing safety. It generates a sequence containing a series of foot landing coordinates and corresponding joint drive timings as the initial assist path.

6. The adaptive human-robot collaborative control method fusing safety constraints according to claim 5, characterized in that: The initial assistance path is iteratively adjusted based on real-time terrain conditions to form a sequence of feasible assistance paths, including: During the execution of assisted control, real-time terrain point cloud data and the actual foot placement position of the human body are continuously monitored; The actual foot landing position is compared with the expected landing position of the corresponding gait in the initial assist path, and the landing deviation is calculated. If the landing point deviation exceeds the preset safety deviation threshold, the path planning will be re-executed based on the current measured terrain conditions, starting from the current actual foot landing point position, to generate a corrected local assistance path. The corrected local assistance path will then be combined with the remaining unexecuted initial assistance path to form a feasible assistance path sequence.

7. The adaptive human-robot collaborative control method fusing safety constraints according to claim 1, characterized in that: By combining human-computer interaction signals and human motion intention feedback, an enhanced strategy model is constructed through fusion processing to output assisted behavior prediction results, including: Collect human movement intention feedback signals, including: muscle activation timing signals collected by electromyography sensors, plantar pressure distribution signals collected by plantar pressure sensors, and human center of mass offset direction signals collected by inertial measurement units. Human-computer interaction signals and human motion intention feedback signals are input into a multilayer perceptron network. The multilayer perceptron network uses historical motion data as training samples to establish a mapping relationship between the current human motion intention and the expected assistance output. The multilayer perceptron network calculates the expected assist torque, expected assist timing, and expected assist duration in real time based on the current input signal, and uses these as the result of assist behavior prediction. 8.The method of claim 7, wherein: Before inputting the human-computer interaction signals and the human motion intention feedback signals into the multilayer perceptron network, it is necessary to construct an input feature vector, including: The muscle activation time-series signal acquired by the electromyography sensor is segmented according to the acquisition time window. Within each time window, the peak value, mean value, rising slope and duration of muscle activation of each muscle group are extracted to form a muscle activation feature vector. The plantar pressure distribution signal collected by the plantar pressure sensor is divided into the heel area, arch area and forefoot area according to the foot region. The pressure peak, pressure center trajectory and ground contact time ratio of each area are calculated to form a plantar pressure feature vector. The human body center of mass offset direction signal collected by the inertial measurement unit is decomposed into forward and backward offset and left and right offset, and combined with the rate of change of the human body center of mass offset direction signal, a center of mass motion feature vector is formed. The muscle activation feature vector, plantar pressure feature vector, and center of mass motion feature vector are concatenated to form a fused feature vector, which serves as the input layer node data for the multilayer perceptron network.

9. The adaptive human-machine collaboration control method fusing safety constraints according to claim 8, characterized in that: The fused feature vectors are input into a multilayer perceptron network to establish a mapping relationship between the current human movement intention and the desired assist output, including: The multilayer perceptron network includes an input layer, at least two hidden layers, and an output layer. The number of nodes in the input layer is the same as the dimension of the fused feature vector. The number of nodes in the hidden layer decreases layer by layer. The number of nodes in the output layer is three, corresponding to the expected assist torque, the expected assist timing, and the expected assist duration, respectively. The multilayer perceptron network is pre-trained by collecting human motion data and corresponding optimal assist parameters under various terrain conditions. During the training process, the backpropagation algorithm is used to adjust the network weights to minimize the mean square error between the expected assist parameters output by the network and the optimal assist parameters in the training samples.

10. The adaptive human-machine collaboration control method fusing safety constraints according to claim 7, characterized in that: Dynamic operation requirements include: minimum assist efficiency requirement in flat walking mode, maximum assist torque requirement in uphill climbing mode, braking damping requirement in downhill descent mode, foot lift height requirement in obstacle crossing mode, and waist support torque requirement in load-bearing mode. When the prediction results of the assistive behavior meet the dynamic operation requirements, a sequence of control instructions is generated, including: Identify the current work mode based on the current scenario description information, and retrieve the requirement parameters corresponding to the current work mode from the dynamic work requirements; The expected assist torque, expected assist timing, and expected assist duration in the assist behavior prediction results are compared with the demand parameters. When the expected assist torque, expected assist timing, and expected assist duration all fall within the allowable range defined by the demand parameters, it is determined that the assist behavior prediction results meet the dynamic operation requirements.

Images