Adaptive position and force control method and system for teleoperation robot in complex scene

Abstract

A teleoperation robot adaptive position control and force control method and system in a complex scene, acquires multi-source perception data and carries out space-time alignment preprocessing; task state and contact state are identified based on the preprocessed data; when a new task is identified, the control parameters are updated based on the model-independent meta-learning framework and the basic parameters in the meta database and the new task samples; the position control weight and the force control weight are continuously adjusted according to the task and the contact state; the dynamic parameters are identified online, and the load inertia compensation term of the position control loop is updated based on the parameters; the control amount increments output by the position control and the force control are fused according to the weight to generate the final control amount by using a parallel double-loop weighted fusion architecture. The present application avoids the hard switching mutation of position control and force control through continuous weight adjustment, realizes dynamic load inertia compensation through online identification, and combines high-precision positioning and soft contact in a small space, and realizes rapid adaptive migration of multiple maintenance tasks through meta-learning.

Description

Technical Field

[0001] This invention belongs to the field of robot control technology, and in particular relates to an adaptive position and force control method and system for remotely operated robots in complex scenarios. Background Technology

[0002] In special power operations (such as substation equipment maintenance, and inspection and maintenance of enclosed cable tunnels), remotely operated robotic arms are often used to replace manual labor in dangerous environments with strong electromagnetic fields, narrow spaces, and high dynamic interference. To ensure operational accuracy and safety, the slave robotic arm usually needs to implement position control in free space to achieve precise positioning. After contacting the equipment, it switches to force control or impedance control to achieve compliant contact. A typical solution can be found in the published patent document CN109968361B, which discloses that impedance control enables the robotic arm to have a certain degree of compliance when contacting the environment. However, its impedance parameters are mostly fixed or only roughly adjusted according to the environment. The position control and force control usually adopt a binary hard switching strategy based on the contact force threshold.

[0003] Existing technologies (including the hybrid control scheme based on selection matrix switching disclosed in CN120326607A) have the following shortcomings in practical applications, which limit their reliability in complex power maintenance scenarios: First, there are defects in the decoupling of position control and force control, as well as hard switching. Existing hybrid position / force control or impedance control systems mostly use position control during the free motion phase and switch to force control or fixed impedance mode during the contact phase. The switching condition is usually whether the contact force exceeds a set threshold, which is a discrete mode switching. In confined equipment compartments, the robotic arm end effector is prone to premature or delayed contact when approaching the operating port. Hard switching can cause abrupt changes in the control law, resulting in end-effector position oscillations or sudden increases in contact force, which can easily cause scratches on the sealing surface or equipment collisions. Under dynamic loads, it is difficult to simultaneously achieve both positioning accuracy and contact compliance.

[0004] Second, fixed dynamic parameters lead to poor adaptability to dynamic working conditions. Existing solutions mostly use fixed dynamic parameters (inertia, friction) calibrated at the factory. When the robotic arm changes the end effector or grasps objects of different masses, the actual load does not match the model. The position control loop is prone to tracking lag or overshoot, and the compliance characteristics of the force control loop deviate from the design value. In live maintenance operations with frequent load changes, the control accuracy drops significantly.

[0005] Third, the anomaly detection and emergency response capabilities are insufficient. Existing emergency strategies mostly rely on a single sensor (such as relying solely on a six-dimensional force sensor exceeding its limit) to determine faults and trigger actions, which are susceptible to electromagnetic interference, leading to misjudgments or missed detections. Furthermore, the link between anomaly identification and the execution of emergency actions is relatively long, making it difficult to meet the rapid response requirements for high-risk operations such as circuit breaker tripping, which require millisecond-level response.

[0006] Fourth, the system has poor task generalization ability and relies on manual parameter configuration. The optimal impedance parameters and PID gains required for different maintenance tasks (such as GIS gas replenishment, enclosure opening and closing, and disconnector operation) vary greatly. The existing system requires operators to manually readjust the control parameters for different tasks, which is time-consuming and cannot achieve rapid adaptive migration under multiple operating scenarios.

[0007] Therefore, there is an urgent need for an adaptive positioning control system that can overcome the above-mentioned shortcomings and is suitable for teleoperated robots in complex scenarios. Force control collaborative control scheme. Summary of the Invention

[0008] To address the problems of abrupt changes in control laws caused by hard switching between position control and force control, and poor control accuracy under dynamic loads due to fixed dynamic parameters in existing technologies, this invention utilizes a robot position control and force control collaborative control method, system, terminal, and medium. It continuously adjusts the position control and force control weights and merges them into a single output based on the task state and contact state. Simultaneously, it identifies dynamic parameters online to update load compensation, achieving a continuous and smooth transition between position control and force control, and maintaining high-precision positioning and compliant contact under dynamic loads.

[0009] The present invention adopts the following technical solution.

[0010] An adaptive position and force control method for teleoperated robots in complex scenarios includes: Acquire multi-source sensing data, and perform time-alignment and spatial-alignment preprocessing on the multi-source sensing data; Based on preprocessed multi-source sensing data, identify the current task status and contact status; Based on the task status and the contact status, continuously adjust the position control weight and force control weight; The dynamic parameters are identified online, and the load inertia compensation term of the position control loop is updated based on the dynamic parameters; A parallel dual-ring weighted fusion architecture is adopted, which weights and fuses the control increment output by the position control ring and the control increment output by the force control ring according to the position control weight and the force control weight to generate the final control increment, and drives the robot to perform actions based on the final control increment.

[0011] More preferably, the preprocessing of temporal and spatial alignment of the multi-source sensing data includes: A sliding window dynamic time warping algorithm is adopted, using the torque sensor as the reference clock, to synchronize the time of sensor data with different sampling frequencies, and obtain time-aligned data. By employing hand-eye calibration and laser-torque joint registration, the time-aligned data is mapped onto the robot's base coordinate system to obtain preprocessed multi-source perception data.

[0012] More preferably, the identification of the current task state and contact state includes: The distance between the end effector and the target point, the end effector velocity, and the end effector contact torque are extracted as feature parameters. The feature parameters are input into a pre-trained support vector machine classifier, which outputs task type labels and contact state labels.

[0013] More preferably, the continuous adjustment of the position control weight and the force control weight includes: When the task type label is non-contact positioning, the initial value of the position control weight is set to the first upper limit value, and decreases linearly as the ratio of the end contact torque to the torque safety threshold increases; the initial value of the force control weight is set to the first lower limit value, and increases linearly as the ratio of the end contact torque to the torque safety threshold increases. When the task type label is contact operation, the initial value of the position control weight is set to the second lower limit value, and it increases linearly with the increase of the ratio of the difference between the current position of the end point and the target position to the target position tolerance; the initial value of the force control weight is set to the second upper limit value, and it decreases linearly with the increase of the ratio. The sum of the position control weight and the force control weight is 1.

[0014] More preferably, when a new task is identified, the control parameters are updated based on the model-independent meta-learning framework, using the basic parameters in the pre-stored meta-database and the new task samples; The update includes: using the basic parameters as initial values, using the weighted sum of the position error and force error of the new task sample as the loss function, and using the gradient descent method to iteratively update the basic parameters to obtain the updated control parameters; The updated control parameters are used to guide the continuous adjustment of the position control weights and the force control weights.

[0015] More preferably, the online identification of dynamic parameters includes: Construct the robot joint torque equation, and rewrite the joint torque equation into a linear parametric form by multiplying the regression matrix with the parameter vector to be identified; The recursive least squares method with a forgetting factor is adopted. In each control cycle, the gain matrix is ​​calculated based on the parameter estimates of the previous cycle, the regression matrix of the current cycle, and the joint torque measurement of the current cycle. The gain matrix is ​​then used to update the parameter vector to be identified to obtain the dynamic parameters of the current cycle. The dynamic parameters include load inertia parameters and friction force parameters.

[0016] More preferably, the method for calculating the increment of the control quantity output by the position control loop is as follows: The product of proportional gain and position error, the product of integral gain and integral of position error, and the product of derivative gain and derivative of position error are added together, and the load inertia compensation term is added to obtain the control quantity increment output by the position control loop. The load inertia compensation term is obtained by multiplying the inertia compensation coefficient by the ratio of the load inertia parameter to the robot's unloaded inertia, and then multiplying the ratio by the joint angular acceleration.

[0017] More preferably, the control quantity increment output by the force control loop is obtained in the following way: An impedance control model is used to map the external contact force into end displacement; By using the pseudo-inverse of the Jacobian matrix, the end displacement is converted into the control increment of the joint space, thus obtaining the control increment output by the force control loop.

[0018] More preferably, the method further includes: Force anomaly features, position anomaly features, and vibration anomaly features are extracted from the preprocessed multi-source sensing data; Using the improved DS evidence theory, the basic trust assignment of each sensor output is first discounted according to the reliability weight of each sensor, and then the discounted basic trust assignment is synthesized by Dempster to obtain the fused trust degree of each anomaly type. When the fusion trust level of a high-risk fault type exceeds a preset threshold, the pre-stored emergency trajectory is invoked to execute an emergency action.

[0019] This invention also proposes an adaptive position and force control system for teleoperated robots in complex scenarios, comprising a data alignment unit, a state recognition unit, a weight allocation unit, a parameter identification unit, and a control fusion unit: A data alignment unit is used to acquire multi-source sensing data and perform time alignment and spatial alignment preprocessing on the multi-source sensing data. The state recognition unit is used to identify the current task state and contact state based on preprocessed multi-source sensing data; The weight allocation unit is used to continuously adjust the position control weight and force control weight according to the task state and the contact state. The parameter identification unit is used to identify dynamic parameters online and update the load inertia compensation term of the position control loop based on the dynamic parameters. The control fusion unit is used to employ a parallel dual-loop weighted fusion architecture to weight and fuse the control increment output by the position control loop and the control increment output by the force control loop according to the position control weight and the force control weight, generate the final control increment, and drive the robot to perform actions based on the final control increment.

[0020] More preferably, the system further includes: The parameter adaptation unit is used to update the control parameters based on the model-independent meta-learning framework, using the basic parameters in the pre-stored meta-database and the new task samples when a new task is identified. Anomaly monitoring and emergency response unit is used to extract force anomaly features, position anomaly features, and vibration anomaly features from preprocessed multi-source sensing data. The improved DS evidence theory is used to fuse the force anomaly features, position anomaly features, and vibration anomaly features according to the reliability weight of each sensor to obtain the fusion confidence level of each anomaly type. When the fusion confidence level of a high-risk fault type exceeds a preset threshold, the pre-stored emergency trajectory is invoked to execute emergency actions.

[0021] The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is used to perform the steps of the above method according to the instructions.

[0022] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. It achieves a smooth and continuous transition between position control and force control, significantly improving the stability of operation in confined spaces. Existing solutions typically employ a hard switching strategy between position control and force control, which can cause abrupt changes in control mode during the switching instant, leading to end-effector position oscillations or sudden increases in contact force. This invention avoids such abrupt changes by using a dynamic weighting function to continuously adjust the contribution ratio of position control and force control based on the task type and real-time contact state. This continuous transition mechanism enables the robot to maintain high-precision positioning during the non-contact phase and exhibit good compliance during the contact phase, effectively solving the operational instability problem caused by excessive rigidity or excessive compliance in confined space operations.

[0024] 2. Online dynamic parameter identification ensures high-precision control under dynamic loads. Existing methods often rely on fixed dynamic parameters, leading to a significant decrease in control accuracy when changing the end effector or grasping objects of different masses. This invention improves the recursive least squares algorithm by updating inertia and friction parameters in real time within the control cycle, ensuring that the load inertia compensation term of the position control loop always matches the actual working conditions. This method can capture changes in load characteristics in real time, ensuring that the robot maintains stable trajectory tracking performance under dynamic load conditions and avoiding control lag or overshoot caused by parameter mismatch.

[0025] 3. This millisecond-level anomaly detection and emergency response mechanism significantly reduces the risk of high-risk operations. Existing emergency response strategies often rely on single sensor threshold judgments, which are susceptible to noise interference, leading to false alarms or missed detections. This invention employs an improved DS evidence theory that integrates multimodal features and combines sensor reliability weights for comprehensive analysis, overcoming the vulnerability of single sensors to interference. A tiered triggering mechanism ensures that when a high-risk fault is identified, the system can skip the conventional process and directly execute emergency actions, minimizing response time and effectively preventing equipment damage and personnel safety risks.

[0026] 4. Meta-learning-driven task generalization capabilities significantly reduce the workload of manual configuration. Traditional multi-task robots require manual tuning of control parameters for each task, resulting in a time-consuming switching process. This invention introduces a model-independent meta-learning framework, which allows for rapid parameter adaptation with only a small number of new task samples, enabling the robot to be deployed to new task cycles without manual intervention. This design significantly improves task switching efficiency and ensures that the robot maintains its optimal control performance range across different task scenarios, enhancing the system's versatility and adaptability to various situations. Attached Figure Description

[0027] Figure 1 This is a flowchart of an adaptive position and force control method for a teleoperated robot in complex scenarios according to the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0029] like Figure 1 As shown, this invention proposes an adaptive position and force control method for teleoperated robots in complex scenarios, specifically including the following steps: Acquire multi-source sensing data, and perform time-alignment and spatial-alignment preprocessing on the multi-source sensing data; A sliding window dynamic time warping algorithm is adopted, using the torque sensor as the reference clock, to synchronize the time of sensor data with different sampling frequencies, and obtain time-aligned data. By employing hand-eye calibration and laser-torque joint registration, the time-aligned data is mapped onto the robot's base coordinate system to obtain preprocessed multi-source perception data.

[0030] Based on preprocessed multi-source sensing data, identify the current task status and contact status; The distance between the end effector and the target point, the end effector velocity, and the end effector contact torque are extracted as feature parameters. The feature parameters are input into a pre-trained support vector machine classifier, which outputs task type labels and contact state labels.

[0031] When a new task is identified, the control parameters are updated based on the model-independent meta-learning framework, using the basic parameters in the pre-stored meta-database and the new task samples. The update includes: using the basic parameters as initial values, using the weighted sum of the position error and force error of the new task sample as the loss function, and using the gradient descent method to iteratively update the basic parameters to obtain the updated control parameters; The updated control parameters are used to guide the continuous adjustment of the position control weights and the force control weights.

[0032] Based on the task status and the contact status, continuously adjust the position control weight and force control weight; When the task type label is non-contact positioning, the initial value of the position control weight is set to the first upper limit value, and decreases linearly as the ratio of the end contact torque to the torque safety threshold increases; the initial value of the force control weight is set to the first lower limit value, and increases linearly as the ratio of the end contact torque to the torque safety threshold increases. When the task type label is contact operation, the initial value of the position control weight is set to the second lower limit value, and it increases linearly with the increase of the ratio of the difference between the current position of the end point and the target position to the target position tolerance; the initial value of the force control weight is set to the second upper limit value, and it decreases linearly with the increase of the ratio. The sum of the position control weight and the force control weight is 1.

[0033] The dynamic parameters are identified online, and the load inertia compensation term of the position control loop is updated based on the dynamic parameters; The online identification of dynamic parameters includes: Construct the robot joint torque equation, and rewrite the joint torque equation into a linear parametric form by multiplying the regression matrix with the parameter vector to be identified; The recursive least squares method with a forgetting factor is adopted. In each control cycle, the gain matrix is ​​calculated based on the parameter estimates of the previous cycle, the regression matrix of the current cycle, and the joint torque measurement of the current cycle. The gain matrix is ​​then used to update the parameter vector to be identified to obtain the dynamic parameters of the current cycle. The dynamic parameters include load inertia parameters and friction force parameters.

[0034] A parallel dual-ring weighted fusion architecture is adopted, which weights and fuses the control increment output by the position control ring and the control increment output by the force control ring according to the position control weight and the force control weight to generate the final control increment, and drives the robot to perform actions based on the final control increment.

[0035] The method for calculating the increment of the control quantity output by the position control loop is as follows: The product of proportional gain and position error, the product of integral gain and integral of position error, and the product of derivative gain and derivative of position error are added together, and the load inertia compensation term is added to obtain the control quantity increment output by the position control loop. The load inertia compensation term is obtained by multiplying the inertia compensation coefficient by the ratio of the load inertia parameter to the robot's unloaded inertia, and then multiplying the ratio by the joint angular acceleration.

[0036] The incremental control output of the force control loop is obtained in the following way: An impedance control model is used to map the external contact force into end displacement; By using the pseudo-inverse of the Jacobian matrix, the end displacement is converted into the control increment of the joint space, thus obtaining the control increment output by the force control loop.

[0037] The method further includes: Force anomaly features, position anomaly features, and vibration anomaly features are extracted from the preprocessed multi-source sensing data; Using the improved DS evidence theory, the basic trust assignment of each sensor output is first discounted according to the reliability weight of each sensor, and then the discounted basic trust assignment is synthesized by Dempster to obtain the fused trust degree of each anomaly type. When the fusion trust level of a high-risk fault type exceeds a preset threshold, the pre-stored emergency trajectory is invoked to execute an emergency action.

[0038] This invention also proposes an adaptive position and force control system for teleoperated robots in complex scenarios, comprising a data alignment unit, a state recognition unit, a weight allocation unit, a parameter identification unit, and a control fusion unit: A data alignment unit is used to acquire multi-source sensing data and perform time alignment and spatial alignment preprocessing on the multi-source sensing data. The state recognition unit is used to identify the current task state and contact state based on preprocessed multi-source sensing data; The weight allocation unit is used to continuously adjust the position control weight and force control weight according to the task state and the contact state. The parameter identification unit is used to identify dynamic parameters online and update the load inertia compensation term of the position control loop based on the dynamic parameters. The control fusion unit is used to employ a parallel dual-loop weighted fusion architecture to weight and fuse the control increment output by the position control loop and the control increment output by the force control loop according to the position control weight and the force control weight, generate the final control increment, and drive the robot to perform actions based on the final control increment.

[0039] More preferably, the system further includes: The parameter adaptation unit is used to update the control parameters based on the model-independent meta-learning framework, using the basic parameters in the pre-stored meta-database and the new task samples when a new task is identified. Anomaly monitoring and emergency response unit is used to extract force anomaly features, position anomaly features, and vibration anomaly features from preprocessed multi-source sensing data. The improved DS evidence theory is used to fuse the force anomaly features, position anomaly features, and vibration anomaly features according to the reliability weight of each sensor to obtain the fusion confidence level of each anomaly type. When the fusion confidence level of a high-risk fault type exceeds a preset threshold, the pre-stored emergency trajectory is invoked to execute emergency actions.

[0040] The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is used to perform the steps of the above method according to the instructions.

[0041] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0042] Example 1 This invention proposes an adaptive position and force control method for teleoperated robots in complex scenarios, comprising: Step 1: Spatiotemporal alignment and preprocessing of multi-source sensing data To address the issues of sensor data jumps and inconsistent timestamps from multiple data sources under strong electromagnetic interference, the raw data from the lidar, six-dimensional torque sensor, and vision camera are first preprocessed: Step 1.1, Time Alignment An improved sliding window dynamic time warping (DTW) algorithm is used to synchronize sensor data with different sampling frequencies. Let the lidar sampling period be... =100ms, the torque sensor sampling period is =1ms, the visual camera sampling period is =33ms, using the torque sensor as the reference clock, for the raw data of any sensor at time t. Its time-aligned data The calculation formula is:

[0043] in, For sensors At the original moment The output value, These correspond to laser, torque, and vision sensors, respectively. The width of the sliding window; For sensors The noise standard deviation was obtained through offline calibration: the noise standard deviation of the lidar. =0.02m, noise standard deviation of torque sensor =0.05 N·m, the noise standard deviation of the visual camera =0.01m. This formula can eliminate the phase deviation caused by asynchronous sampling and suppress impulse noise through Gaussian weighted fusion.

[0044] In this invention, the sliding window half-width The value is determined based on: The reference clock period is 1ms, and the window width should cover the sensor's maximum sampling jitter range. The maximum sampling jitter of the lidar is approximately ±5ms, therefore, we take... =5, which can suppress noise while avoiding the introduction of excessive phase lag.

[0045] Noise Standard Deviation The following results were obtained through offline static calibration experiments: 10,000 sets of sensor data were continuously collected under static conditions, and the standard deviation of their normal distribution was calculated as [the standard deviation]. The final value of is determined by this method, which is well-known in the field and will not be described in detail here.

[0046] Step 1.2, Spatial Alignment Hand-eye calibration combined with laser-torque joint registration is used to map all preprocessed multi-source sensing data to the robot's base coordinate system. Below. Assume the vision sensor is in the end-effector coordinate system. The pose transformation matrix is ​​as follows The pose transformation matrix of the lidar in the base coordinate system is: Then visual feature points Coordinates in the base coordinate system for:

[0047] in, Let A be the homogeneous transformation matrix from coordinate system A to coordinate system B, including the rotation matrix. Translation vector ,Right now ; The values ​​were obtained through calibration using a lidar and an end flange calibration plate, and are stored offline as constants. This formula enables unified spatial reference across multiple sensors, avoiding pose calculation errors.

[0048] in, The specific method of obtaining it is as follows: A calibration plate is installed at the robot's end effector. The robot is driven to move to N≥20 different poses. The poses of the calibration plate observed by the lidar and the robot's forward poses are recorded. The hand-eye calibration equations are solved using the least squares method.

[0049] in, for The first group of observation poses. This process is a standard calibration procedure in this field and will not be elaborated upon here.

[0050] Time alignment eliminates phase deviations caused by differences in sampling frequencies between different sensors, ensuring consistency of multi-source data across the time axis and preventing control commands from malfunctioning due to data delays. Spatial alignment maps measurements from all sensors to the same base coordinate system, eliminating inconsistencies in reference systems caused by differences in installation location and attitude, thus providing a reliable sensing foundation for subsequent task identification and collaborative control.

[0051] Step 2: Task type identification and position-force dynamic weight allocation To address the varying operational requirements of different maintenance tasks, a multi-feature fusion classifier is first used to identify the current task type. Then, the weight ratio of position control and force control is dynamically adjusted to achieve a continuous transition from "precise positioning primarily using position control" to "contact operation primarily using force control," rather than a hard switch.

[0052] Step 2.1, Task Type Identification: Based on the preprocessed multi-source sensing data, identify the current task state and contact state. Extract the three types of features for the current operation: (Distance between the endpoint and the target point) (End-effector velocity) (End contact torque), input to a pre-trained support vector machine (SVM) classifier, output task type label. The corresponding relationships are: 0 = idle standby, 1 = non-contact fine positioning (such as moving above the GIS valve port), 2 = contact operation (such as turning the valve or pressing the trip button).

[0053] The training and multi-class classification methods of the support vector machine classifier are well-known pattern recognition techniques in the field, and the selection and training process of its kernel function will not be described in detail in this invention.

[0054] In this embodiment, the SVM uses an RBF core. Input 3D features First, the training set is Min-Max normalized to [0,1]. The training set is obtained through offline teaching, with 200 samples for each task type (idle standby, non-contact fine positioning, and contact operation), and the labels are automatically annotated by the instructor's operation log; the penalty coefficient C=10. The above training process is a well-known configuration in the field of SVM pattern recognition, and the classifier can be directly trained according to this configuration in the scenario of this application.

[0055] By using a multi-feature fusion classifier to identify the current task type, the system can distinguish between different working conditions such as non-contact positioning and contact operation, providing a reliable perception basis for subsequent dynamic weight allocation and avoiding the risk of misjudgment under complex interference by a single feature.

[0056] Specifically, for different tasks such as container opening and closing and GIS air replenishment, a meta-learning-driven control parameter transfer architecture is constructed, eliminating the need for manual parameter readjustment: In a preferred embodiment of this invention, a typical task metadata database is constructed, which categorizes eight typical power maintenance tasks, storing three types of metadata for each task: The characteristics of the task are: ( For target location tolerance, For maximum permissible force, (Maximum speed of motion); Initial control parameters: ; Successful trajectory samples: more than 100 sets of joint angle-torque-position paired data for this task.

[0057] The eight typical power maintenance tasks described are merely examples, including but not limited to: GIS gas replenishment, enclosure opening and closing, disconnector operation, circuit breaker tripping, grounding switch operation, cable termination installation and removal, insulator cleaning, and foreign object removal.

[0058] The metadata database supports continuous expansion. New tasks can be added to the system by collecting no fewer than 20 successful samples, and are not limited to the task types mentioned above.

[0059] The metadata database supports continuous expansion. When a new task is identified, the control parameters are updated using the pre-stored basic parameters in the metadata database and the new task samples, based on the meta-learning-driven control parameter migration architecture.

[0060] Specifically, when switching to a new task, the Model Independent Meta-Learning (MAML) framework is used to quickly update the control parameters based on a small number of new task samples:

[0061] in, The basic parameters, shared by all tasks, are obtained through offline meta-training. =0.01, which is the learning rate; Let be the loss function, which is the weighted sum of position error and force error: ; =20, meaning only 20 sets of samples are needed to complete the adaptation for a new task. This invention can complete the parameter adaptation for a new task in just 3 seconds, without any manual intervention.

[0062] This application's model-independent meta-learning framework operates on the control parameter vector. It uses linear updates (including PID gain, impedance parameters, etc.) rather than deep neural network weight updates, therefore it lacks network structures such as convolutional layers or fully connected layers. Its "inner loop" calculates the loss function L using 20 sets of samples from the new task. The gradient is calculated, and a learning rate step is performed. Gradient descent at 0.01 yields The "outer loop" calculates the second gradient of the multi-task loss during the offline meta-training phase to update the... This linear MAML is a direct application of the meta-learning framework to parameterized control laws. Those skilled in the art can execute it based on the loss function, learning rate, and sample size given in this embodiment.

[0063] By employing a meta-learning framework to enable rapid transfer of control parameters, the system can adapt to new task scenarios based on a small number of samples without manual readjustment. This method improves task switching efficiency and ensures that different tasks maintain optimal control performance, thereby enhancing the robot's versatility and adaptability to various situations. The updated control parameters obtained from the meta-learning parameter adaptation are then passed to the subsequent weight allocation step to guide the continuous adjustment of position control weights and force control weights.

[0064] Step 2.2: Based on the task status, contact status, and updated control parameters, continuously adjust the position control weight and force control weight. Let the position control weight be Force control weight is ,satisfy The weights are dynamically adjusted based on the task type and real-time contact status.

[0065] in, As a preferred embodiment of the present invention, the torque safety threshold is taken as... =5 N·m, exceeding this value is considered as contact occurring; For the target location tolerance, preferably, the present invention takes =0.005m (5mm), meaning that the force control is initiated when the end enters within 5mm of the target. This represents the current position of the endpoint in the base coordinate system. The location of the target operation point.

[0066] Specifically, torque safety threshold The setting is based on: In typical electrical contact operations, the normal operating force range is 0.5–3 N·m. When it exceeds 5 N·m, there is a risk of jamming or collision. Therefore, this value is set as a safety threshold.

[0067] Target location tolerance The setting is based on the fact that the size of the operating port of power equipment is usually 10-20mm, with a 5mm safety margin reserved to ensure a smooth transition from position control to force control.

[0068] During the non-contact positioning stage, the position control weight is up to 90%. If there is accidental contact (torque increases), the position control weight is gradually reduced and the force control weight is increased. During the contact operation stage, the position control weight is at least 10%. As the end approaches the target, the position control weight is gradually increased to ensure operational accuracy and avoid position drift caused by excessively soft force control.

[0069] The basis for setting the upper limit of 90% and the lower limit of 10% for position control weight is as follows: Excessive position control weight can lead to excessive contact rigidity, while excessively low weight can cause position drift. Experiments have shown that 90% and 10% can achieve the optimal balance between accuracy and compliance.

[0070] By identifying the current task type through a multi-feature fusion classifier, the system can distinguish between different working conditions such as non-contact positioning and contact operation, and dynamically adjust the weight ratio of position control and force control accordingly. This continuous weight adjustment mechanism avoids the abrupt changes in control mode that occur in traditional hard switching strategies, thereby improving the compliance of the contact process while ensuring positioning accuracy, and effectively reducing operational instability caused by excessive rigidity or excessive compliance.

[0071] Step 3: Online identification of load inertia and friction (solving the problem of dynamic parameter drift) This step identifies dynamic parameters online based on motion information from preprocessed multi-source sensing data and updates the load inertia compensation term of the position control loop based on these parameters. Existing solutions often use fixed dynamic parameters, leading to decreased control accuracy after load changes. This invention identifies load inertia and friction parameters online every 10ms and updates the control law in real time. Step 3.1, Constructing a dynamic model The equation for the joint torque of the robot is:

[0072] This formula is used to describe the dynamic model of a multi-degree-of-freedom robot and calculate the driving torque required for each joint. Wherein, To provide torque to the motor, current is passed through it. Conversion: ; The torque constant of the motor is preferably 0.12 N·m / A; This is the joint inertia matrix, which includes the robot's body inertia and the load inertia. ; Coriolis force / centrifugal force matrix The gravitational moment matrix; External disturbance torque; The frictional torque is obtained using an improved LuGre model:

[0073] in, For the average deformation of the bristle, =800N / m, which is the bristle stiffness, determining the magnitude of frictional hysteresis; =0.5 N·s / m, which is the micro damping coefficient, characterizing the energy dissipation of the bristle; =0.3N·s / m is the viscous friction coefficient, which characterizes linear friction in the high-speed region.

[0074] Among them, internal state The average deformation of the mane is represented by the state evolution equation in the standard form of the LuGre model:

[0075] For the Stribeck function, The Coulomb friction torque is... For the maximum static friction torque, For Stribeck velocity; the above evolution equation is an inherent, well-known part of the LuGre friction model. Parameters , , and , , The method for obtaining this information is as follows: A ramp-step torque excitation is applied to a single joint of the target robotic arm, with the joint rotation speed... Using the input as the input and minimizing the error between the motor current-calculated torque and the model output torque as the criterion, the torque is obtained using least squares fitting; this embodiment provides... =800N / m =0.5 N·s / m =0.3N·s / m is the calibration result of a certain seven-degree-of-freedom robotic arm under rated load. Different models of robotic arms can obtain the parameters again by following the same calibration process described above.

[0076] Step 3.2, Online Recursive Least Squares Identification Rewrite the robot joint torque equation in linear parametric form:

[0077] in, For the regression matrix, , where is the parameter vector to be identified.

[0078] Parameters are updated using recursive least squares (RLS) with a forgetting factor:

[0079]

[0080]

[0081] in, For the first Estimated values ​​of identification parameters for each control cycle; This is the gain matrix; The forgetting factor, set to 0.995, represents the system's rate of forgetting historical data. A value closer to 1 indicates a more robust system to slowly varying loads; a smaller value results in a faster response to sudden load changes. Experiments have shown that 0.995 achieves optimal stability in dynamic power operation scenarios. Let covariance be the matrix, initialized to... =10 6 • I. To prevent the covariance matrix of weak excitation segments from being... Too small a gain Approaching zero, identification stagnation, this invention addresses Set a lower limit for protection: when traces When it is less than the preset lower limit, let ; or to Adding regularization terms to the recursive expression This treatment is a standard numerical stability method with a forgetting factor (RLS), ensuring stability in each cycle. Non-zero, parameters are continuously micro-updated.

[0082] By identifying and updating the dynamic model parameters online in real time, the control law is always adapted to the current load state, avoiding the degradation of control performance caused by fixed parameters not matching the actual situation. This method can capture changes in load inertia and friction characteristics, improving the stability and accuracy of the system under dynamic conditions such as tool changes and grasping different objects.

[0083] Step 4: Design of Adaptive Position Control-Force Control Cooperative Control Law A parallel dual-ring weighted fusion architecture is adopted. Based on position control weight and force control weight, the control increment output by the position control ring and the control increment output by the force control ring are weighted and fused according to the position control weight and force control weight to generate the final control increment.

[0084] This invention employs a parallel dual-ring weighted fusion architecture, with the position control ring outputting the joint angle increment. Force control ring outputs joint angle increment The final control quantity is a weighted sum of the two:

[0085] It should be noted that the position control ring output of this application... With force control ring output These are all joint space angle increments (unit: rad), not a direct sum of absolute position and relative offset in Cartesian space. The position control ring uses position error... The joint angle increment is directly obtained through PID control and inertia compensation; the force control loop first obtains the Cartesian compliant displacement from the impedance model, and then through the Jacobian pseudo-inverse. Mapped to joint angle increments of the same dimension (See step 4.2). Therefore It is a convex combination of two joint increments of the same dimension. The actual end-effector motion is the result of the collaboration between the two loops: when the position control points forward and the force control pushes back due to contact, the weighted result shows a compliant convergence of "position control approaching the target and force control limiting the contact force," rather than the cancellation of the forces from the two loops. This parallel weighted architecture is a well-known structure in the field of hybrid position / force control for robots, and those skilled in the art, in conjunction with the information provided in this application, will understand its significance. The stability of a continuous function can be predicted by its impedance parameter, without the need for a separate Lyapunov proof.

[0086] Step 4.1, Position Control Loop Design (Improved Adaptive PID) To address the issue of decreased position control accuracy caused by dynamic load changes, a load inertia compensation term is introduced into the traditional PID controller:

[0087] in, This refers to the positional error; For the PID gain, the critical gain and oscillation period are obtained through classic PID tuning methods such as the Ziegler-Nichols method, and then the PID gain parameters are calculated. Based on this, the present invention adds a load inertia compensation term to offset the tracking lag caused by dynamic load changes. =0.3 is the inertia compensation coefficient, which is calibrated experimentally; The real-time load inertia is obtained through online identification in step 4; =0.8kg·m2 is the robot's no-load inertia, a factory constant; The joint angular acceleration is obtained differentially through the joint encoder. When the load increases, the feedforward compensation is automatically increased to offset the tracking lag caused by inertia, ensuring that the position control accuracy remains stable at the 0.1mm level.

[0088] Step 4.2, Force Control Loop Design (Improved Impedance Control) To address the compliance requirements in contact operations, impedance control is employed to achieve dynamic "position-force" mapping. The target impedance model is as follows:

[0089] in, =0.5kg =20 N·s / m =150N / m are the target inertia, damping and stiffness coefficients, respectively, and their preferred values ​​are derived from the force-position coupling characteristics of typical power operation tasks; The actual Cartesian position at the end. For the desired position; The external contact force is measured by a six-dimensional torque sensor. When the external force increases, the end automatically generates a small displacement to achieve compliant contact and avoid rigid collisions.

[0090] The force-potential coupling characteristics of the typical power operation tasks include, but are not limited to: GIS air replenishment: The contact surface is made of soft sealing material, and the stiffness is taken as... =80N / m, to avoid scratches; Enclosure opening and closing: The contact surface is metal, and the rigidity is taken as... =150N / m; Emergency tripping: requires rapid response, damping should be used. =40 N·s / m.

[0091] In the actual system, the corresponding parameters are obtained by looking up the table in the metadata database based on the task type label, so as to achieve dynamic adjustment.

[0092] Convert the impedance model into joint space control variables. Through the Jacobian matrix Mapping:

[0093] in, This is the pseudo-inverse of the Jacobian matrix. , used to convert Cartesian space motion into joint space motion; =0.001s is the control period. This invention converts the compliance requirement in Cartesian space into a joint angle adjustment amount, realizing a force control closed loop.

[0094] The Jacobian matrix and its pseudo-inverse calculation method are common knowledge in the field of kinematic mapping of serial robotic arms, and the derivation process will not be elaborated in this invention.

[0095] The parallel dual-loop weighted fusion architecture enables position control and force control to participate in decision-making within the same control cycle, overcoming the problem of error propagation from the previous stage to the next stage in a serial structure. The position control loop introduces load inertia compensation, which can maintain trajectory tracking performance when the load changes; the force control loop uses impedance control to achieve dynamic mapping between position and force, enabling the robot to automatically adjust its end effector displacement according to external forces during contact, thereby achieving compliant operation and avoiding rigid collisions.

[0096] Step 5: Millisecond-level anomaly identification and priority triggering (solving emergency response delay issues) Force anomaly features, position anomaly features, and vibration anomaly features are extracted from the preprocessed multi-source sensing data. The improved DS evidence theory is used to fuse the force anomaly features, position anomaly features, and vibration anomaly features according to the reliability weight of each sensor to obtain the fusion confidence level of each anomaly type. When the fusion confidence level of a high-risk fault type exceeds a preset threshold, the pre-stored emergency trajectory is invoked to execute emergency actions.

[0097] For emergency tasks such as circuit breaker tripping, a three-level anomaly tracing and rapid triggering mechanism is designed to compress the response delay to less than 100ms: Step 5.1, Multimodal Anomaly Feature Extraction Extract three types of abnormal features: Abnormal force: (Exceeding the emergency torque threshold, used to distinguish between normal contact and dangerous jamming); Bit anomaly: (A positional deviation exceeding 2cm is considered four times the typical equipment operating tolerance and is thus deemed out of control.) Vibration anomaly: (A sudden change in angular acceleration, three times the peak value of normal motion acceleration, is used to identify impacts.)

[0098] Step 5.2, Anomaly Type Fusion Based on Improved DS Evidence Theory The DS evidence theory is improved by using reliability weights for different sensors, and the anomaly probabilities of three types of features are fused to avoid misjudgment by a single sensor. Let the recognition framework be , =No abnormalities =Stuck, =Collision, =Equipment failure (emergency tripping required).

[0099] Basic trust assignment for each sensor output First, apply a discount to eliminate the conflict:

[0100] in, For sensors The reliability weight is determined based on the long-term statistical accuracy of the sensor; the reliability weight of the lidar is... =0.9, the reliability weight of the torque sensor is =0.95, the reliability weight of the visual camera is =0.85.

[0101] Then synthesize Dempster:

[0102]

[0103] Among them, B, C, and D represent the three sensors—LiDAR, torque sensor, and vision camera—in the recognition framework, respectively. The focal length of the sensor. Basic trust allocation for each sensor. The construction method is as follows: taking force anomalies as an example, define Gaussian membership degree. ,when hour For positional anomalies and vibration anomalies, membership degrees are constructed based on corresponding thresholds and standard deviations, and confidence values ​​are assigned. Finally, normalization is performed. This mapping from physical quantities to basic trust assignments is a standard preliminary step in the application of DS evidence theory.

[0104] Step 5.3, Rapid Execution of Emergency Actions After an emergency is triggered, the normal task process is skipped, and the pre-stored circuit breaker tripping trajectory library is directly invoked:

[0105] in, The emergency action time is preset to ensure that the tripping operation is completed within 100ms.

[0106] Based on the fusion of multimodal anomaly characteristics and improved D-S evidence theory, the system can comprehensively identify anomaly types by integrating information from different sensors, overcoming the vulnerability of single sensors to noise interference. A tiered triggering mechanism ensures that in the event of a high-risk fault, the system can quickly skip routine procedures and directly execute emergency actions, thereby reducing potential damage and improving overall safety.

[0107] Example 2 This embodiment provides a robot position control and force control collaborative control system. This system is isomorphic to the method embodiment but differs in dimension, focusing on the structural connections and data flow between modules. It should be understood that this embodiment is merely illustrative and not restrictive; those skilled in the art can merge or split the modules according to actual hardware deployment requirements.

[0108] Example 2 This embodiment provides a robot position and force control collaborative control system. This system is isomorphic to the method embodiment, focusing on the structural connections and data flow between modules, and is used to execute the method described in Embodiment 1. It should be understood that this embodiment is merely illustrative and not restrictive; those skilled in the art can merge or split the modules according to actual hardware deployment requirements.

[0109] The system includes: a data alignment unit, a status identification unit, a parameter adaptation unit, a weight allocation unit, a parameter identification unit, a control fusion unit, and an anomaly monitoring and emergency response unit.

[0110] The data alignment unit is used to acquire multi-source sensing data and perform temporal and spatial alignment preprocessing on the multi-source sensing data. Specifically, the data alignment unit receives raw data streams from LiDAR, a six-dimensional torque sensor, and a vision camera. It completes time alignment using a sliding window dynamic time warping algorithm with the torque sensor as the reference clock, and completes spatial alignment through hand-eye calibration and laser-torque joint registration. It outputs a spatiotemporally aligned sensing data packet uniformly mapped to the robot's base coordinate system and pushes this data packet to the state recognition unit and parameter identification unit.

[0111] The state recognition unit is used to identify the current task state and contact state based on preprocessed multi-source sensing data. Specifically, the state recognition unit receives spatiotemporally aligned sensing data packets from the data alignment unit, extracts the distance between the end effector and the target point, the end effector's motion velocity, and the end effector's contact torque as feature parameters, outputs task type labels and contact state labels through a pre-trained support vector machine classifier, and transmits the output results to the parameter adaptation unit and the weight allocation unit.

[0112] The parameter adaptation unit, when a new task is identified, updates the control parameters based on a model-independent meta-learning framework, utilizing pre-stored basic parameters from the metadata database and new task samples. Specifically, the parameter adaptation unit receives the task type label output by the state recognition unit. When it is determined that a switch to a new task has occurred, it retrieves the corresponding task features and initial control parameters from the metadata database. Using the weighted sum of position error and force error as the loss function, it iteratively updates the basic parameters using gradient descent to obtain the updated control parameters, which are then transmitted to the weight allocation unit and the control fusion unit.

[0113] The weight allocation unit continuously adjusts the position control weight and force control weight based on the task status, contact status, and updated control parameters. Specifically, the weight allocation unit receives the task type label, contact status label, and updated control parameters, calculates the position control weight and force control weight in real time using a continuous weight function, ensures that the sum of the two is 1 and that they transition smoothly and continuously with the status, and then sends the weight parameters to the control fusion unit.

[0114] The parameter identification unit is used to identify dynamic parameters online and update the load inertia compensation term of the position control loop based on the dynamic parameters. Specifically, the parameter identification unit receives motion state data and motor current feedback from the data alignment unit, uses a recursive least squares algorithm with a forgetting factor to update the load inertia parameters and friction parameters in each control cycle, and transmits the identified load inertia parameters to the control fusion unit.

[0115] The control fusion unit employs a parallel dual-loop weighted fusion architecture to weight and fuse the control increments output from the position control loop and the force control loop according to their respective weights, generating the final control increment. Specifically, the position control loop module and the force control loop module operate in parallel within the control fusion unit: the position control loop module outputs a position control increment with inertia compensation based on the updated load inertia compensation term and the desired trajectory; the force control loop module outputs a force control increment based on an impedance model based on the external contact force and the desired position; the control fusion unit receives weight parameters transmitted by the weight allocation unit, weights and sums the output increments of the two parallel loops, generates the final control increment, and sends it to the robot drive layer.

[0116] The anomaly monitoring and emergency response unit extracts force anomaly features, position anomaly features, and vibration anomaly features from preprocessed multi-source sensing data. It uses an improved DS evidence theory to fuse these features according to the reliability weights of each sensor. When the fusion confidence level of a high-risk fault type exceeds a preset threshold, a pre-stored emergency trajectory is invoked to execute an emergency action. Specifically, the anomaly monitoring and emergency response unit receives spatiotemporally aligned sensing data packets output by the data alignment unit, performs multimodal fusion judgment using the improved DS evidence theory, and when a high-risk fault is identified, directly invokes the pre-stored emergency trajectory and sends it to the robot drive layer, skipping the conventional control process.

[0117] It should be understood that the above-mentioned units can be implemented as independent hardware modules or integrated into a system-on-a-chip or a logic function module in an industrial control computer, as long as they can realize the functional logic of multi-source data spatiotemporal alignment, new task parameter adaptation, continuous weight fusion and allocation, online parameter identification and compensation, parallel dual-loop weighted output and abnormal emergency response.

[0118] Example 3 This embodiment provides a terminal and a computer-readable storage medium to provide physical support for the method of Embodiment 1 and the system of Embodiment 2 from the dimensions of hardware carrier and program storage.

[0119] On the one hand, a terminal includes a processor and storage media: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method described in Embodiment 1, specifically including: Invoke the data alignment command to acquire multi-source sensing data and perform time and spatial alignment preprocessing on the multi-source sensing data; Invoke the status recognition command to identify the current task status and contact status based on preprocessed multi-source sensing data; When a new task is identified, the parameter adaptation instruction is invoked, and the control parameters are updated based on the model-independent meta-learning framework using the basic parameters in the pre-stored meta-database and the new task samples. Invoke the weight allocation instruction to continuously adjust the position control weight and force control weight based on the task status, contact status, and updated control parameters. Call the parameter identification command to identify dynamic parameters online and update the load inertia compensation term of the position control loop; The control fusion instruction is invoked, and a parallel dual-loop weighted fusion architecture is used to weight and fuse the control increments output by the position control loop and the force control loop according to the position control weight and the force control weight, generate the final control increment, and send it to the robot drive layer. During the control execution process, the abnormal monitoring command is invoked to extract force abnormality features, position abnormality features and vibration abnormality features from the preprocessed multi-source sensing data. The improved DS evidence theory is used for fusion judgment. When the fusion trust level of the high-risk fault type exceeds the preset threshold, the pre-stored emergency trajectory is invoked to execute the emergency action.

[0120] It should be understood that the specific form of the terminal includes, but is not limited to, industrial control computers, embedded microcontrollers, edge computing gateways, or cloud servers, as long as its processor can complete the collaborative control process disclosed in Embodiment 1 according to the above instruction logic.

[0121] On the other hand, a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0122] Computer-readable storage media are tangible physical media, including but not limited to portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, static random access memory, portable compressed disk read-only memory, digital multifunction disks, memory sticks, or floppy disks. The computer program stored in the storage medium includes program code modules corresponding to the steps of Embodiment 1. When the processor reads and executes these code modules, it can virtualize the data alignment unit, state recognition unit, parameter adaptation unit, weight allocation unit, parameter identification unit, control fusion unit, and anomaly monitoring and emergency response unit described in Embodiment 2 on the terminal hardware platform, thereby fully realizing the collaborative control functions of continuous weight fusion, online identification of load compensation, new task parameter adaptation, and anomaly emergency response.

[0123] It should be understood that computer-readable storage media do not include transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, but must rely on physical physical media to store instructions used by instruction execution devices.

[0124] Example 4 This embodiment uses the gas replenishment operation of GIS equipment in a 500kV substation as an example to illustrate the application process of the method of this application in a practical engineering project. It should be understood that this embodiment is only illustrative and not restrictive, and those skilled in the art can make adaptive adjustments to the parameters and procedures according to the actual operation scenario.

[0125] The operating environment is a GIS equipment compartment with strong electromagnetic interference and narrow operating space. The air supply port is located below the compartment wall, surrounded by busbars and secondary cable trays. The remotely operated robot uses a seven-degree-of-freedom serial robotic arm with a six-dimensional torque sensor and gas supply connector integrated at the end effector. The base is equipped with a lidar and an industrial camera.

[0126] Step S100: Spatiotemporal alignment and preprocessing of multi-source sensing data In the strong electromagnetic environment on site, the sampling frequencies of the various sensors differ: the sampling period of the lidar is 100ms, the sampling period of the torque sensor is 1ms, and the sampling period of the vision camera is 33ms.

[0127] The system uses a torque sensor as the reference clock and employs a sliding window dynamic time warping algorithm for time alignment. The sliding window half-width is set to 5 reference periods (5ms), covering the maximum sampling jitter range of the LiDAR (±5ms). The noise standard deviation is obtained based on offline static calibration experiments: 0.02m for the LiDAR, 0.05 N·m for the torque sensor, and 0.01m for the vision camera.

[0128] During the spatial alignment stage, hand-eye calibration and laser-torque joint registration are adopted to map visual feature points and laser point clouds to the robot base coordinate system in a unified manner, ensuring that the detection error of the air inlet position is controlled within 0.5mm, providing a reliable benchmark for subsequent fine operations.

[0129] Step S200: Task Status Identification The system extracts the distance between the end effector and the air inlet, the end effector's velocity, and the end effector's contact torque as feature parameters. These parameters are then input into a pre-trained support vector machine classifier, which identifies the current task as non-contact fine localization. This identification result serves as the basis for subsequent weight allocation and parameter adaptation decisions.

[0130] Step S300: Adapting new task parameters After this air replenishment operation is completed, the robot needs to switch to the GIS valve opening task in the adjacent interval. The system calls the pre-stored task characteristics (target position tolerance, maximum allowable force, maximum movement speed) and initial control parameters (PID gain, impedance parameters) in the metadata database.

[0131] Based on a model-independent meta-learning framework, the system collects 20 successful valve opening operation samples on-site. Using the weighted sum of position and force errors as the loss function, the system iteratively updates the basic parameters using gradient descent. The entire process takes approximately 3 seconds and requires no manual intervention, thus achieving rapid adaptation of control parameters.

[0132] Step S400: Position Control-Force Control Dynamic Weight Allocation During the non-contact fine positioning stage, the system sets the position control weight to 0.9 and the force control weight to 0.1 to ensure high-precision approximation of the air inlet at the end.

[0133] When the end effector enters within 5mm of the air inlet and a contact torque exceeding 5 N·m (torque safety threshold) is detected, the weighting function is automatically and continuously adjusted: the position control weight is linearly reduced to 0.5, and the force control weight is linearly increased to 0.5. This continuous transition mechanism avoids contact impact caused by hard switching and achieves a smooth transition from positioning to compliant docking.

[0134] Step S500: Online dynamic parameter identification and compensation During the tightening of the air supply connector, the robot's load inertia changes with the connector's position. The system performs online identification every 10ms: a recursive least squares algorithm with a forgetting factor (0.995) is used to estimate the load inertia and friction parameters in real time.

[0135] The identified real-time load inertia is used to update the load inertia compensation term of the position control loop, which offsets the tracking lag caused by changes in hose drag force and connector weight, ensuring stable tightening torque.

[0136] Step S600: Parallel dual-loop weighted fusion control The system adopts a parallel dual-ring architecture: The position control loop adopts adaptive PID control with inertia compensation, with PID gains of 120, 5, and 20, and an inertia compensation coefficient of 0.3. The force control ring uses impedance control, and the target stiffness is set to 80 N / m and the damping is set to 20 N·s / m for the characteristics of the soft sealing surface.

[0137] The two control signals are weighted and fused according to their current weights (0.5 / 0.5) to generate the final control signal, which is then sent to the joint actuator to achieve the control effect of "precise positioning and smooth contact without damaging the parts".

[0138] Step S700: Anomaly Monitoring and Emergency Response During operation, the system monitors three types of abnormal characteristics in real time: contact torque exceeding 10 N·m (force abnormality), end position deviation exceeding 2 cm (position abnormality), and joint angular acceleration exceeding 500 rad / s² (vibration abnormality).

[0139] The system employs an improved DS evidence theory for fusion judgment: it discounts the evidence based on the long-term statistical accuracy of lidar (reliability weight 0.9), torque sensor (0.95), and vision camera (0.85). When the fusion trust level determines a high-risk fault (such as equipment jamming) and the confidence level exceeds 0.8, the system immediately skips the routine process, calls the pre-stored emergency evacuation trajectory, and removes the terminal from the danger zone within 100ms to prevent damage to the sealing surface.

[0140] This embodiment verifies the effectiveness of the present invention in environments with strong electromagnetic fields, narrow spaces, and high dynamic interference. It can achieve millisecond-level safety emergency response and rapid switching of multiple tasks while ensuring sub-millimeter-level operational accuracy.

[0141] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0142] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0143] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0144] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. An adaptive position and force control method for a teleoperated robot in complex scenarios, characterized in that, include: Acquire multi-source sensing data, and perform time-alignment and spatial-alignment preprocessing on the multi-source sensing data; Based on preprocessed multi-source sensing data, identify the current task status and contact status; Based on the task state and the contact state, continuously adjust the position control weight and force control weight, including: When the task type label corresponding to the task status is non-contact positioning, the initial value of the position control weight is set to the first upper limit value, and it decreases linearly as the ratio of the end contact torque to the torque safety threshold increases; the initial value of the force control weight is set to the first lower limit value, and it increases linearly as the ratio of the end contact torque to the torque safety threshold increases. When the task type label is contact operation, the initial value of the position control weight is set to the second lower limit value, and it increases linearly with the increase of the ratio of the difference between the current position of the end point and the target position to the target position tolerance; the initial value of the force control weight is set to the second upper limit value, and it decreases linearly with the increase of the ratio. The sum of the position control weight and the force control weight is 1; When a new task is identified, the control parameters are updated based on the model-independent meta-learning framework, using the basic parameters in the pre-stored meta-database and the new task samples. The update includes: using the basic parameters as initial values, using the weighted sum of the position error and force error of the new task sample as the loss function, and using the gradient descent method to iteratively update the basic parameters to obtain the updated control parameters; The updated control parameters are used to guide the continuous adjustment of the position control weights and the force control weights; The dynamic parameters are identified online, and the load inertia compensation term of the position control loop is updated based on the dynamic parameters; The online identification of dynamic parameters includes: Construct the robot joint torque equation, and rewrite the joint torque equation into a linear parametric form by multiplying the regression matrix with the parameter vector to be identified; The recursive least squares method with a forgetting factor is adopted. In each control cycle, the gain matrix is ​​calculated based on the parameter estimates of the previous cycle, the regression matrix of the current cycle, and the joint torque measurement of the current cycle. The gain matrix is ​​then used to update the parameter vector to be identified to obtain the dynamic parameters of the current cycle. The dynamic parameters include load inertia parameters and friction force parameters; A parallel dual-ring weighted fusion architecture is adopted, which weights and fuses the control increment output by the position control ring and the control increment output by the force control ring according to the position control weight and the force control weight to generate the final control increment, and drives the robot to perform actions based on the final control increment. The method further includes: Force anomaly features, position anomaly features, and vibration anomaly features are extracted from the preprocessed multi-source sensing data; Using the improved DS evidence theory, the basic trust assignment of each sensor output is first discounted according to the reliability weight of each sensor, and then the discounted basic trust assignment is synthesized by Dempster to obtain the fused trust degree of each anomaly type. When the fusion trust level of a high-risk fault type exceeds a preset threshold, the pre-stored emergency trajectory is invoked to execute an emergency action.

2. The adaptive position and force control method for a teleoperated robot in complex scenarios according to claim 1, characterized in that: The preprocessing of the multi-source sensing data, including temporal and spatial alignment, includes: A sliding window dynamic time warping algorithm is adopted, using the torque sensor as the reference clock, to synchronize the time of sensor data with different sampling frequencies, and obtain time-aligned data. By employing hand-eye calibration and laser-torque joint registration, the time-aligned data is mapped onto the robot's base coordinate system to obtain preprocessed multi-source perception data.

3. The adaptive position and force control method for a teleoperated robot in complex scenarios according to claim 1, characterized in that: The identification of the current task state and contact state includes: The distance between the end effector and the target point, the end effector velocity, and the end effector contact torque are extracted as feature parameters. The feature parameters are input into a pre-trained support vector machine classifier, which outputs task type labels and contact state labels.

4. The adaptive position and force control method for a teleoperated robot in complex scenarios according to claim 1, characterized in that: The method for calculating the increment of the control quantity output by the position control loop is as follows: The product of proportional gain and position error, the product of integral gain and integral of position error, and the product of derivative gain and derivative of position error are added together, and the load inertia compensation term is added to obtain the control quantity increment output by the position control loop. The load inertia compensation term is obtained by multiplying the inertia compensation coefficient by the ratio of the load inertia parameter to the robot's unloaded inertia, and then multiplying the ratio by the joint angular acceleration.

5. The adaptive position and force control method for a teleoperated robot in complex scenarios according to claim 1, characterized in that: The incremental control output of the force control loop is obtained in the following way: An impedance control model is used to map the external contact force into end displacement; By using the pseudo-inverse of the Jacobian matrix, the end displacement is converted into the control increment of the joint space, thus obtaining the control increment output by the force control loop.

6. An adaptive position and force control system for a teleoperated robot in complex scenarios using the method described in any one of claims 1-5, characterized in that, It includes a data alignment unit, a state recognition unit, a weight allocation unit, a parameter identification unit, and a control fusion unit: A data alignment unit is used to acquire multi-source sensing data and perform time alignment and spatial alignment preprocessing on the multi-source sensing data. The state recognition unit is used to identify the current task state and contact state based on preprocessed multi-source sensing data; An anomaly monitoring and emergency response unit is used to extract force anomaly features, position anomaly features, and vibration anomaly features from preprocessed multi-source sensing data. The improved DS evidence theory is used to fuse the force anomaly features, position anomaly features, and vibration anomaly features according to the reliability weight of each sensor to obtain the fusion confidence level of each anomaly type. When the fusion confidence level of a high-risk fault type exceeds a preset threshold, the pre-stored emergency trajectory is invoked to execute emergency actions. The weight allocation unit is used to continuously adjust the position control weight and force control weight according to the task state and the contact state. The parameter adaptation unit is used to update the control parameters based on the model-independent meta-learning framework, using the basic parameters in the pre-stored meta-database and the new task samples when a new task is identified. The parameter identification unit is used to identify dynamic parameters online and update the load inertia compensation term of the position control loop based on the dynamic parameters. The control fusion unit is used to employ a parallel dual-loop weighted fusion architecture to weight and fuse the control increment output by the position control loop and the control increment output by the force control loop according to the position control weight and the force control weight, generate the final control increment, and drive the robot to perform actions based on the final control increment.

7. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-5.

Images