Multi-sensor fusion mechanical arm joint module force control system and debugging method

Abstract

The invention relates to the field of mechanical arms, and discloses a multi-sensor fusion mechanical arm joint module force control system and a debugging method, and the multi-sensor fusion mechanical arm joint module force control system comprises a sensing module which is used for collecting state data of a mechanical arm joint module in real time through multiple sensors; the processing module is used for preprocessing the state data by adopting a federated learning algorithm and fusing the preprocessed state data by utilizing an extended Kalman filtering algorithm to obtain processed data; and the decision module is used for constructing a multi-objective optimization function based on a mechanical arm dynamic model according to the processing data. The method comprises the following steps: acquiring multi-dimensional data through multiple sensors, preprocessing the data by utilizing federated learning, fusing and outputting an accurate state estimation value in combination with extended Kalman filtering, constructing a multi-objective optimization function based on fused data, identifying kinetic parameters in real time through a recursive least square method, and dynamically adjusting a weighting factor by utilizing a particle swarm optimization algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of robotic arms, and specifically to a force control system and debugging method for a robotic arm joint module with multi-sensor fusion. Background Art

[0002] As a core component of intelligent manufacturing and intelligent home appliances, the force control technology of robotic arms is crucial in fields such as industrial assembly and service robots. In the smart home scenario, robotic arms need to perform fine operations such as food ingredient grasping, tableware placement, and cleaning tasks, which pose higher requirements for force control accuracy, environmental adaptability, and human-robot collaboration safety. Existing robotic arm force control systems mainly rely on single sensors (such as strain force sensors) or fixed-parameter control algorithms, and it is difficult to cope with the challenges of diverse object materials (such as glass, ceramics, and flexible food ingredients), dynamic load changes (such as grasping objects of different weights), and unstructured scenarios (such as operating on a cluttered tabletop) in the home environment.

[0003] Traditional robotic arm force control systems usually adopt an architecture of "single-sensor feedback + PID control". For example, industrial robotic arms collect the end contact force by installing six-axis force sensors, and combine with a PID controller to adjust the joint torque to achieve a force control closed-loop. However, the parameters of the PID controller need to be manually preset. When the robotic arm grasps objects of different weights, the change of the inertia matrix will cause a significant increase in force control error; it is impossible to identify the dynamic parameters of the robotic arm in real time, and it is difficult to cope with model mismatch problems caused by load changes, joint wear, etc. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the present invention provides a force control system and debugging method for a robotic arm joint module with multi-sensor fusion, which solves the problems that the traditional robotic arm force control system cannot identify the dynamic parameters of the robotic arm in real time and is difficult to cope with model mismatch caused by load changes, joint wear, etc.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: A force control system for a robotic arm joint module with multi-sensor fusion, including:

[0006] A sensing module: used to collect the state data of the robotic arm joint module in real time through multi-sensors;

[0007] A processing module: used to preprocess the state data by using a federated learning algorithm and fuse the preprocessed state data by using an extended Kalman filter algorithm to obtain processed data;

[0008] A decision-making module: used to construct a multi-objective optimization function based on the robotic arm dynamics model according to the processed data, and cooperate with the recursive least squares method and the particle swarm optimization algorithm for parameter identification and optimization to generate control instructions;

[0009] Execution module: Used to drive the robotic arm joint module to move according to the control instructions.

[0010] By adopting the above technical solution, multi-dimensional data is collected through multiple sensors to provide rich inputs for the model. Federated learning is used to preprocess the data, and the extended Kalman filter is combined to fuse the output to obtain accurate state estimation values. A multi-objective optimization function is constructed based on the fused data. The dynamic parameters are identified in real time through the recursive least squares method, and the weighted factor is dynamically adjusted using the particle swarm optimization algorithm to balance multiple objectives such as accuracy and energy consumption. The joint motion is driven according to the optimized instructions, and the state monitoring and emergency braking are realized through human-machine interaction. The real-time online identification of the robotic arm dynamic parameters, the adaptive adjustment of the multi-objective control strategy, and the generation of control instructions are achieved, solving the problem that the traditional robotic arm force control system cannot identify the robotic arm dynamic parameters in real time and is difficult to cope with the model mismatch caused by load changes, joint wear, etc.

[0011] [[ID=~]]Preferably, the sensing module includes a data acquisition unit and a data transmission unit. The data acquisition unit is used to collect the state data of the robotic arm joint module in real time through multiple sensors. The multiple sensors include a force sensor, a motion sensor, an environmental sensor, and a vision sensor. The state data includes the motion state information, force information, environmental information, and target object information of the robotic arm joint module. The data transmission unit is used to transmit the various data through a communication protocol. The communication protocol includes SPI, CAN, and Ethernet.

[0012] Preferably, the processing module includes a preprocessing unit and a fusion unit. The preprocessing unit is used to preprocess the data of each sensor at the edge computing node using the federated learning algorithm. The preprocessing includes data filtering, denoising, feature extraction, and normalization processing. The fusion unit is used to fuse the preprocessed state data using the extended Kalman filter algorithm to obtain the processed data. The extended Kalman filter algorithm includes a state prediction equation, a covariance prediction equation, and a measurement update equation.

[0013] Preferably, the state prediction equation is where is the predicted state at time k based on the information at time k - 1, f is the non-linear state transition function, is the state estimation value at time k - 1, U k-1 is the control input at time k - 1;

[0014] The covariance prediction equation is where P k|k-1 is the predicted covariance, F k-1 is the Jacobian matrix of the state transition matrix, P k-1|k-1 is the covariance estimation value at time k - 1, Qk-1 is the process noise covariance matrix;

[0015] The measurement update equation includes and P k|k = (I - K k H k )P k|k-1 , where K k is the Kalman gain, H k is the Jacobian matrix of the measurement matrix, R k is the measurement noise covariance matrix, z k is the measurement value at time k, h is the nonlinear measurement function, and I is the identity matrix.

[0016] Preferably, the decision-making module includes a predictive control unit, an adaptive control unit, a scheduling management unit, and an instruction generation unit. The predictive control unit is used to construct a multi-objective optimization function based on the manipulator dynamics model according to the processed data, including force error, energy consumption, and velocity smoothness. The manipulator dynamics model is a relationship model between joint torque and motion state established based on the Lagrangian equation. The adaptive control unit is used to cooperate with the recursive least squares method and the particle swarm optimization algorithm to identify and optimize the parameters of the manipulator dynamics model to obtain an identification result. The iterative variables of the recursive least squares method include parameter estimation values, gain matrices, and covariance matrices. The recursive least squares method is used to continuously update the parameters for parameter estimation. The particle swarm optimization algorithm is used to globally search and optimize the weighting factors λ1, λ2, and λ3 when the parameter estimation converges to a local optimum during the parameter identification process of the recursive least squares method. The scheduling management unit is used to allocate computing resources based on the task priority and system resource occupancy using a scheduling strategy that combines time slice rotation and priority preemption. The instruction generation unit is used to generate control instructions according to the optimized multi-objective function and the parameter identification result.

[0017] Preferably, the multi-objective optimization function is where N is the prediction horizon length, λ1, λ2, and λ3 are weighting factors, f d (k) is the desired force vector at time k, f(k) is the predicted actual force vector at time k, τ(k) is the joint torque vector at time k, is the joint velocity change vector at time k.

[0018] Preferably, the execution module includes a drive control unit and a human-machine interaction unit. The drive control unit is used to adopt a servo driver based on field-oriented control to drive the robotic arm joint module to move according to control instructions. The human-machine interaction unit is used to perform human-machine interaction through a touch screen and a graphical interface, including displaying the system status, receiving user instructions, and performing emergency braking operations.

[0019] A debugging method for force control of a multi-sensor fusion robotic arm joint module is applied to the above-mentioned multi-sensor fusion robotic arm joint module force control system, and includes the following steps:

[0020] Perception: Real-time collect the status data of the robotic arm joint module through multi-sensors;

[0021] Processing: Use the federated learning algorithm to preprocess the status data, and use the extended Kalman filter algorithm to fuse the preprocessed status data to obtain processed data;

[0022] Decision-making: Construct a multi-objective optimization function according to the processed data, and use the recursive least squares method and the particle swarm optimization algorithm in cooperation to perform parameter identification and optimization, and generate control instructions;

[0023] Execution: Drive the robotic arm joint module to move according to the control instructions.

[0024] The present invention provides a multi-sensor fusion robotic arm joint module force control system and a debugging method. It has the following beneficial effects:

[0025] 1. The present invention collects multi-dimensional data through multi-sensors, preprocesses the data using federated learning, combines the extended Kalman filter to fuse and output accurate state estimation values, constructs a multi-objective optimization function based on the fused data, real-time identifies the dynamic parameters through the recursive least squares method, and uses the particle swarm optimization algorithm to dynamically adjust the weighting factor, and drives the joint movement according to the optimized instructions, realizing the real-time online identification of the robotic arm dynamic parameters, the adaptive adjustment of the multi-objective control strategy and the generation of control instructions, and solving the problem that the traditional robotic arm force control system cannot real-time identify the robotic arm dynamic parameters and is difficult to cope with model mismatches caused by load changes, joint wear, etc.

[0026] 2. The present invention real-time iteratively updates the dynamic parameters through the recursive least squares method, combines the particle swarm optimization algorithm to globally search for the weighting factors of the multi-objective function, forms a closed-loop feedback of "parameter identification - objective optimization", enables the robotic arm to automatically correct the model parameters and balance multiple objectives such as force control accuracy, energy consumption, and motion smoothness when grasping objects of different weights or when joint performance decays, ensures that the control strategy dynamically evolves with the change of the environment, and enhances the robustness of the system in unstructured scenarios.

[0027] 3. The present invention constructs a multi-objective optimization function based on the manipulator dynamics model, ensures the real-time operation of the algorithm through the time slice rotation and priority preemption strategies, enables the system to complete the full-process calculation from the solution of the objective function to the generation of instructions, and realizes multi-objective collaborative optimization by dynamically adjusting the weighting factors, ensuring both the control of the end contact force, reducing the motor energy consumption and suppressing the sudden change of the joint speed, enabling the manipulator to output efficient and stable control instructions in different task scenarios, and improving the overall operation efficiency and equipment reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 FIG. is the system architecture diagram of the multi-sensor fusion manipulator joint module force control system proposed by the present invention;

[0029] Figure 2 FIG. is the method flowchart of the debugging method for the multi-sensor fusion manipulator joint module force control proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0031] Please refer to the attached Figure 1 , the embodiment of the present invention provides a multi-sensor fusion manipulator joint module force control system, including:

[0032] Perception module: used to collect the state data of the manipulator joint module in real time through multi-sensors; the perception module includes a data acquisition unit and a data transmission unit. The data acquisition unit is used to collect the state data of the manipulator joint module in real time through multi-sensors. The multi-sensors include force sensors, motion sensors, environmental sensors, and vision sensors. The state data includes the motion state information, force information, environmental information, and target object information of the manipulator joint module. The data transmission unit is used to transmit each item of data through communication protocols, and the communication protocols include SPI, CAN, and Ethernet.

[0033] Specifically, a "perception nervous system" of the manipulator is constructed through multi-type sensors. The force sensor captures the details of the contact force, the motion sensor monitors the joint dynamics, the vision sensor constructs the environmental map, and the environmental sensor senses variables such as temperature and humidity. After the data is preprocessed by edge computing, it is transmitted to the processing module with low latency and high reliability, providing multi-dimensional data support for subsequent control.

[0034] The main sensor of the force sensor, ATINano17, adopts the strain measurement principle. It converts the force on the joint into a voltage signal through a Wheatstone bridge, converts it into a digital quantity through a 24-bit ADC, with a sampling rate of 1kHz, and can directly output six-dimensional force / torque data. The redundant FBG torque sensor utilizes the fiber Bragg grating wavelength drift characteristic (sensitivity 1pm / N·m), and calculates the torque in real time through a demodulator (such as MicronOpticssm125), avoiding system failures caused by the failure of a single sensor.

[0035] The IMU module of the motion sensor fuses the gyroscope and accelerometer data through complementary filtering and outputs the joint attitude quaternion (update rate 100Hz), with a drift rate <0.1° / h; the absolute encoder uses a binary code disk, supports power-off position memory, and the repeat positioning accuracy ≤ ±0.001°.

[0036] The binocular camera of the vision sensor focuses on the target object through the ROI (region of interest) algorithm, combines the stereoBM algorithm of OpenCV to calculate the disparity map, and the three-dimensional reconstruction error ≤ 1mm; the structured light camera generates 10 frames of point cloud data per second, performs surface fitting through the PCL library, and constructs a three-dimensional model of the object.

[0037] The data transmission unit performs communication protocol layering. High-speed channels (SPI / CAN): Transmit real-time data such as force sensors and encoders. The CAN bus baud rate is set to 1Mbps to meet the control cycle (1ms) requirement; Low-speed channel (Ethernet): Transmit vision point cloud data, using the TCP / IP protocol, with a bandwidth reservation of 100Mbps. And an anti-interference design is adopted. The sensor cable uses a twisted shielded wire, the CAN bus is equipped with a 120Ω terminal resistor, and the Ethernet interface integrates a magnetic isolation module to ensure data transmission in an industrial environment.

[0038] The processing module: It is used to preprocess the state data using the federated learning algorithm and fuse the preprocessed state data using the extended Kalman filter algorithm to obtain the processed data; The processing module includes a preprocessing unit and a fusion unit. The preprocessing unit is used to preprocess the data of each sensor at the edge computing node using the federated learning algorithm. The preprocessing includes data filtering, denoising, feature extraction, and normalization processing. The fusion unit is used to fuse the preprocessed state data using the extended Kalman filter algorithm to obtain the processed data. The extended Kalman filter algorithm includes a state prediction equation, a covariance prediction equation, and a measurement update equation.

[0039] The state prediction equation is Among them, is the predicted state at time k based on the information at time k-1, f is the non-linear state transition function, is the state estimate value at time k-1, Uk-1 is the control input at time k-1;

[0040] The covariance prediction equation is where P k|k-1 is the predicted covariance, F k-1 is the Jacobian matrix of the state transition matrix, P k-1|k-1 is the covariance estimate at time k-1, Q k-1 is the process noise covariance matrix;

[0041] The measurement update equation includes and P k|k =(I-K k H k )P k|k-1 , where K k is the Kalman gain, H k is the Jacobian matrix of the measurement matrix, R k is the measurement noise covariance matrix, z k is the measurement value at time k, h is the non-linear measurement function, and I is the identity matrix.

[0042] Specifically, the processing module realizes distributed feature extraction through federated learning, reducing data transmission pressure and privacy risks; and completes multi-source data fusion with the help of the Extended Kalman Filter (EKF), constructing a unified state space including joint position, velocity, force, and temperature, and providing an accurate real-time state estimate of the robotic arm for the decision-making module.

[0043] The preprocessing unit constructs a federated learning network based on edge computing nodes (such as NVIDIA Jetson Nano). Each sensor node (force sensor, vision camera, etc.) acts as an independent client, completes data cleaning and feature extraction locally, and only uploads the feature vector to the central server, avoiding the leakage of raw data and transmission bandwidth pressure.

[0044] The fusion unit is the core component of the processing module, and realizes the optimal estimation of multi-source sensor data based on the Extended Kalman Filter (EKF) algorithm, providing real-time and accurate robotic arm state information for the decision-making module. The input data of the Extended Kalman Filter algorithm includes the measurement value of the force sensor, the joint position and velocity of the motion sensor data, the vision measurement value, and the joint torque command of the control input. The output data is the state estimate value: a 12-dimensional vector including joint position, velocity, and end force, with an update frequency of 1 kHz; and the covariance matrix, which reflects the uncertainty of the state estimate and is used for parameter adjustment of the decision-making module.

[0045] In the state prediction stage of the fusion unit: the formula is Input the state estimate value at the previous moment Control input Uk-1 , the processing process is to predict, calculate, and force predict through the robotic arm dynamics model, and output the predicted state of the robotic arm joint module at time k based on the information at time k-1, so as to predict the state (such as joint velocity, end force) through the dynamics model and control input, realize the preliminary estimation of the motion state of the robotic arm at the next moment, provide a basic prediction value for subsequent precise control, and enable the robotic arm to plan the motion trend in advance according to the current state and control instructions.

[0046] In the covariance prediction stage: the formula is Input the covariance matrix P at the previous moment k-1|k-1 , the process noise matrix Q k-1 , the processing process is the state transition matrix, simplified to the identity matrix, and calculate the covariance, output the predicted covariance of the robotic arm joint module, so as to quantify the prediction uncertainty (such as the variance of velocity / force), reflect the model confidence, realize the evaluation of the reliability of the state prediction, clarify the fluctuation range of the prediction result, provide a standard for measuring the prediction error for the subsequent measurement update stage, and assist the decision-making module to judge whether and how to adjust the control strategy.

[0047] In the measurement update stage: the formula is Input the measurement value z k , including the data collected by the force sensor, the data of the encoder, the data collected by the vision camera, the measurement function Directly map the force and position prediction values, and the processing process is to calculate the residual, Kalman gain, and state update, and output the corrected state estimation value of the robotic arm joint module, so as to correct the prediction value using the sensor data (such as adjusting the end force estimation with the actual measured value of the force sensor), realize the fusion of the actual sensor measurement data and the model prediction data, correct the prediction deviation, improve the accuracy and reliability of the robotic arm state estimation, and enable the robotic arm to adjust the motion state in time according to the actual situation to meet the high-precision control requirements.

[0048] Through three-stage iteration, the fusion unit realizes a closed loop from model prediction to data correction, ensuring the force control accuracy in the grasping task.

[0049] Decision-making module: It is used to construct a multi-objective optimization function based on the robotic arm dynamics model according to the processed data, and use the recursive least squares method and the particle swarm optimization algorithm in cooperation to perform parameter identification and optimization, and generate control instructions. The decision-making module includes a predictive control unit, an adaptive control unit, a scheduling and management unit, and an instruction generation unit. The predictive control unit is used to construct a multi-objective optimization function based on the robotic arm dynamics model according to the processed data, including force error, energy consumption, and velocity smoothness. The robotic arm dynamics model is a relationship model between joint torque and motion state established based on the Lagrange equation. The adaptive control unit is used to cooperate with the recursive least squares method and the particle swarm optimization algorithm to identify and optimize the parameters of the robotic arm dynamics model to obtain the identification result. The iterative variables of the recursive least squares method include parameter estimation values, gain matrices, and covariance matrices. The recursive least squares method is used to continuously update parameters for parameter estimation. The particle swarm optimization algorithm is used to perform a global search and optimization of the weighting factors λ1, λ2, and λ3 when the parameter estimation converges to a local optimum during the parameter identification process using the recursive least squares method. The scheduling and management unit is used to allocate computing resources based on the task priority and system resource occupancy using a scheduling strategy that combines time slice rotation and priority preemption. The instruction generation unit is used to generate control instructions according to the optimized multi-objective function and the parameter identification result.

[0050] The multi-objective optimization function is where N is the prediction horizon length, λ1, λ2, and λ3 are weighting factors, and f d (k) is the desired force vector at time k, f(k) is the predicted actual force vector at time k, τ(k) is the joint torque vector at time k, is the joint velocity change vector at time k.

[0051] Specifically, the decision-making module constructs a multi-objective optimization function that integrates force error, energy consumption, and velocity smoothness based on the robotic arm dynamics model. Based on the multi-objective optimization function, it comprehensively considers multiple objectives such as the force control accuracy, motion smoothness, and energy consumption of the robotic arm, outputs joint torque instructions, drives the robotic arm to execute tasks, real-time identifies model parameters through the recursive least squares method (RLS), combines the particle swarm optimization (PSO) to break through the local optimum, and dynamically generates optimal control instructions to achieve the adaptive evolution of the control strategy.

[0052] The predictive control unit constructs a multi-objective optimization function based on the robotic arm dynamics model and the fused state data, and outputs an optimized control target sequence. The fused state data output by its input processing module includes a 12-dimensional state vector such as joint position, velocity, and end force; task target parameters such as the desired force and motion trajectory points in the grasping task. Its core processing process: Use the formula Among them: Force error term: Calculate the difference between the fused value of the force sensor f(k) and the desired force f d (k) to ensure the accuracy of the contact force of the end effector (for example, the error is ≤0.5N when grasping an egg); Energy consumption term: Minimize the motor energy consumption based on the sum of squares of the joint torques τ(k) (for example, the energy consumption is reduced by 15% during continuous operation); Velocity smoothing term: Suppress the motion impact through the change in joint velocity (for example, the velocity fluctuation is ≤5% when placing on a tray). Receding horizon optimization solution: Use the model predictive control (MPC) framework to solve the optimization function recursively within the prediction horizon N (set to 20 steps, corresponding to 20ms), and output the control quantity sequence for a period of time in the future. Output: Optimized multi-objective function parameters (weighting factors λ1, λ2, λ3) and the control quantity sequence, which are used for parameter identification and command generation of the adaptive control unit.

[0053] For example: In the "grasping a glass" task, input the position of the glass detected by vision and the desired grasping force f d = 1.5N, the optimization function automatically adjusts λ1 = 0.7 (force control priority), and outputs the joint torque sequence to make the end force rise smoothly to the target value, avoiding damage to the glass due to excessive force.

[0054] The adaptive control unit identifies the dynamic model parameters of the robotic arm in real time, optimizes the weighting factors of the multi-objective function, and improves the system robustness. Its input: The control quantity sequence τ(k) output by the predictive control unit; The fused state data and the original sensor measurement values of the processing module (such as encoder position, original force value of the force sensor). Its core processing process: Recursive least squares (RLS) parameter identification is used to estimate the dynamic model parameters (inertia matrix, Coriolis force matrix, gravity vector) in real time and perform iterations; Particle swarm optimization (PSO) weighting factor adjustment: Its trigger condition is when the parameter update amplitude identified by RLS < 1% (i.e., falling into a local optimum), the PSO algorithm is started; Minimize the multi-objective function value J, and the particle positions correspond to the weighting factors λ1, λ2, λ3. Search for the global optimal solution through iterations (such as reallocating weights when the load changes suddenly), and output the identified dynamic model parameters and the optimized weighting factors λ1, λ2, λ3.

[0055] For example: When the robotic arm grasps ingredients of different weights (such as switching from grasping an egg to grasping a can), RLS updates the inertia matrix parameters in real time, and PSO synchronously adjusts λ2 (energy consumption weight) from 0.3 to 0.5 to balance the grasping force and the motor load.

[0056] The scheduling management unit dynamically allocates computing resources based on task priorities and system resources to ensure real-time performance. It dynamically allocates computing resources based on task priorities and system resource status. By defining three levels of priorities (real-time control tasks are level 1, non-real-time tasks are level 3) and adopting a time-slice rotation and priority preemption strategy, it ensures that the control cycle is stable within 1 ms. After inputting the task requests of each unit and the system resource status, it first sorts the tasks by priority, then allocates equal time slices to tasks with the same priority. When a high-priority task arrives, it immediately interrupts the low-priority task. Finally, it outputs the task execution sequence and the resource allocation table, and preferentially guarantees the CPU resources for generating control instructions in a multi-task concurrent scenario to avoid control delays.

[0057] The instruction generation unit converts the optimized control parameters into executable physical instructions to drive the movement of the robotic arm. It converts the optimized control parameters into executable physical instructions. After inputting the optimized control quantity sequence of the predictive control unit, the parameter identification results of the adaptive control unit, and the execution priority instructions of the scheduling management unit, it first converts the joint torque instruction into a 12-bit precision PWM signal, then calculates the feedforward compensation term according to the identified dynamic parameters to counteract the inertial force and Coriolis force, and at the same time performs a physical limit check on the instruction (such as joint torque ≤ 20 N·m). Finally, it outputs the drive control instruction and the safety status signal. For example, in the "stirring ingredients" task, it accurately generates the PWM signal and checks the torque safety threshold to ensure uniform stirring force and equipment safety.

[0058] The execution module: is used to drive the robotic arm joint module to move according to the control instruction. The execution module includes a drive control unit and a human-machine interaction unit. The drive control unit is used to use a servo drive based on field-oriented control to drive the robotic arm joint module to move according to the control instruction. The human-machine interaction unit is used to perform human-machine interaction through a touch screen and a graphical interface, including displaying the system status, receiving user instructions, and performing an emergency braking operation.

[0059] Specifically, the execution module converts digital instructions into mechanical motion. The servo drive unit realizes the accurate output of joint torque. The human-machine interaction unit constructs a safe and convenient operation interface through the touch screen and the emergency braking mechanism to ensure the reliability and usability of the robotic arm in a human-robot collaboration scenario.

[0060] The drive control unit receives the torque instruction τ(k), realizes torque closed-loop through the current loop (bandwidth 1 kHz), and receives the encoder feedback, performs trajectory tracking through a PID controller, and compensates for the centrifugal force and Coriolis force during the high-speed movement of the robotic arm in real time through a feedforward control algorithm.

[0061] The human-machine interaction unit adopts a safety mechanism. When the axial force of any axis exceeds 30 N, an emergency brake is triggered, and the braking distance ≤ 10 mm (when the speed ≤ 0.1 m / s); the distance of personnel is monitored by a ToF sensor (accuracy ± 1 cm), and when < 20 cm, it automatically switches to the low-speed mode (speed ≤ 0.05 m / s). The interactive interface uses the Qt framework to develop the GUI, including a 3D simulation interface: real-time rendering of the manipulator motion trajectory, error visualization (force error is represented by a color scale, red > 0.5 N, green ≤ 0.2 N); a data dashboard: dynamically displays parameters such as joint temperature, motor current, and multi-objective function values.

[0062] Through the collaborative operation of the perception module, processing module, decision-making module, and execution module, a multi-sensor fusion force control system for the manipulator joint module is constructed, so as to realize the real-time identification of the dynamic parameters of the manipulator and the dynamic optimization of the model.

[0063] The perception module uses multi-source devices such as force sensors, motion sensors, and vision sensors to collect data such as joint position, force, ambient temperature, and target object coordinates in real time, and transmits the data to the processing module through SPI / CAN / Ethernet communication protocols, providing multi-dimensional input data for the dynamic model. For example, when grasping ingredients of different weights, the force sensor real-time feedbacks the change of contact force, and the vision camera identifies the shape and position of the object in advance, providing a basis for model parameter identification.

[0064] The processing module uses the federated learning algorithm to filter, denoise, and extract features from the sensor data at the edge node, and then fuses the data through the extended Kalman filter (EKF) to output accurate state estimation values including joint states, end forces, etc. Among them, the state prediction equation of the EKF is constructed based on the manipulator dynamics model (Lagrange equation), and the input accuracy of the model is improved by fusing multi-source data, laying a foundation for subsequent parameter identification.

[0065] The prediction control unit in the decision-making module constructs a multi-objective optimization function including force error, energy consumption, and speed smoothness based on the state estimation value of the processing module, clarifying the control objectives and parameter optimization directions; the adaptive control unit uses the recursive least squares method (RLS) to identify the dynamic parameters in real time, and dynamically matches the manipulator load change by iteratively updating the parameter estimation value, gain matrix, and covariance matrix; at the same time, the particle swarm optimization algorithm (PSO) performs a global search on the weighting factors of the multi-objective function when the RLS falls into a local optimum, balancing objectives such as force control accuracy and energy consumption; the scheduling management unit ensures that the RLS / PSO algorithm completes parameter updates within a 1 ms control cycle through the time slice rotation and priority preemption strategies, avoiding calculation delays from affecting real-time performance; the instruction generation unit generates joint torque instructions according to the optimized model parameters and control objectives, and outputs them to the execution module after feedforward compensation and safety verification.

[0066] The execution module drives the joint motion through a servo drive based on field-oriented control and uses the human-machine interaction unit to monitor the status in real time. For example, when joint wear causes changes in the inertia matrix, the RLS corrects the model parameters in real time, and the PSO adjusts the weighting factor to give priority to ensuring the force control accuracy. The servo drive accurately outputs the torque according to the updated instructions to ensure stable grasping force. At the same time, the emergency braking mechanism (such as force overrun triggering braking) is used to ensure human-machine safety.

[0067] Thus, the full-process dynamic optimization from data acquisition, fusion, parameter identification to control instruction generation is realized, improving the adaptability of the robotic arm to scenarios such as load changes and component wear, and solving the problem of model mismatch such as load changes and joint wear that the traditional robotic arm force control system cannot handle.

[0068] Please refer to the appendix Figure 2 , a debugging method for the force control of a multi-sensor fusion robotic arm joint module, which is applied to the above-mentioned multi-sensor fusion robotic arm joint module force control system, and includes the following steps:

[0069] Perception: Real-time collect the status data of the robotic arm joint module through multiple sensors;

[0070] Processing: Use the federated learning algorithm to preprocess the status data, and use the extended Kalman filter algorithm to fuse the preprocessed status data to obtain the processed data;

[0071] Decision-making: Construct a multi-objective optimization function according to the processed data, and use the recursive least squares method and the particle swarm optimization algorithm in cooperation to perform parameter identification and optimization to generate control instructions;

[0072] Execution: Drive the robotic arm joint module to move according to the control instructions.

[0073] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A force control system for a robotic arm joint module with multi-sensor fusion, characterized in that, Including: Perception module: used to collect the status data of the robotic arm joint module in real time through multiple sensors; Processing module: used to preprocess the status data using the federated learning algorithm and fuse the preprocessed status data using the extended Kalman filter algorithm to obtain processed data; Decision-making module: used to construct a multi-objective optimization function based on the robotic arm dynamics model according to the processed data, and cooperate with the recursive least squares method and the particle swarm optimization algorithm for parameter identification and optimization to generate control instructions; Execution module: used to drive the robotic arm joint module to move according to the control instructions.

2. The force control system for the robotic arm joint module with multi-sensor fusion according to claim 1, characterized in that: The perception module includes a data acquisition unit and a data transmission unit. The data acquisition unit is used to collect the status data of the robotic arm joint module in real time through multiple sensors. The multiple sensors include a force sensor, a motion sensor, an environment sensor, and a vision sensor. The status data includes the motion state information, force information, environment information, and target object information of the robotic arm joint module. The data transmission unit is used to transmit each item of data through a communication protocol, and the communication protocol includes SPI, CAN, and Ethernet.

3. The force control system for the robotic arm joint module with multi-sensor fusion according to claim 1, characterized in that: The processing module includes a preprocessing unit and a fusion unit. The preprocessing unit is used to preprocess the data of each sensor at the edge computing node using the federated learning algorithm. The preprocessing includes data filtering, denoising, feature extraction, and normalization processing. The fusion unit is used to fuse the preprocessed status data using the extended Kalman filter algorithm to obtain processed data. The extended Kalman filter algorithm includes a state prediction equation, a covariance prediction equation, and a measurement update equation.

4. The force control system for the robotic arm joint module with multi-sensor fusion according to claim 3, characterized in that: The state prediction equation is where is the predicted state at time k based on the information at time k-1, f is the non-linear state transition function, is the state estimate at time k-1, U k-1 is the control input at time k-1; The covariance prediction equation is where P k|k-1 is the predicted covariance, F k-1 is the Jacobian matrix of the state transition matrix, P k-1|k-1 is the covariance estimate at time k-1, Q k-1 is the process noise covariance matrix; The measurement update equation includes and P k|k =(I - K k H k )P k|k-1 , where K k is the Kalman gain, H k is the Jacobian matrix of the measurement matrix, R k is the measurement noise covariance matrix, z k is the measurement value at time k, h is the nonlinear measurement function, and I is the identity matrix.

5. The force control system for the robotic arm joint module with multi-sensor fusion according to claim 1, characterized in that: The decision-making module includes a predictive control unit, an adaptive control unit, a scheduling and management unit, and an instruction generation unit. The predictive control unit is used to construct a multi-objective optimization function based on the robotic arm dynamics model according to the processed data, including force error, energy consumption, and speed smoothness. The robotic arm dynamics model is a relationship model between joint torque and motion state established based on the Lagrangian equation. The adaptive control unit is used to cooperate with the recursive least squares method and the particle swarm optimization algorithm to identify and optimize the parameters of the robotic arm dynamics model to obtain an identification result. The iterative variables of the recursive least squares method include parameter estimation values, gain matrices, and covariance matrices. The recursive least squares method is used to continuously update the parameters for parameter estimation. The particle swarm optimization algorithm is used to globally search and optimize the weighting factors λ1, λ2, and λ3 when the parameter estimation converges to a local optimum during the parameter identification process of the recursive least squares method. The scheduling and management unit is used to allocate computing resources based on the task priority and system resource occupancy using a scheduling strategy that combines time slice rotation and priority preemption. The instruction generation unit is used to generate control instructions according to the optimized multi-objective function and the parameter identification result.

6. The force control system of the robotic arm joint module with multi-sensor fusion according to claim 5, characterized in that: The multi-objective optimization function is where N is the prediction horizon length, λ1, λ2, and λ3 are weighting factors, and f d (k) is the desired force vector at time k, f(k) is the predicted actual force vector at time k, τ(k) is the joint torque vector at time k, and is the joint velocity change vector at time k.

7. The force control system of the robotic arm joint module with multi-sensor fusion according to claim 1, characterized in that: The execution module includes a drive control unit and a human-machine interaction unit. The drive control unit is used to adopt a servo driver based on field-oriented control to drive the robotic arm joint module to move according to control instructions. The human-machine interaction unit is used to perform human-machine interaction through a touch screen and a graphical interface, including displaying the system status, receiving user instructions, and performing emergency braking operations.

8. Debugging method for force control of robotic arm joint module with multi-sensor fusion, characterized in that: Applied to the multi-sensor fusion robotic arm joint module force control system according to any one of claims 1-7, it includes the following steps: Perception: Real-time collect the status data of the robotic arm joint module through multi-sensors; Processing: Preprocess the status data using a federated learning algorithm, and fuse the preprocessed status data using an extended Kalman filter algorithm to obtain processed data; Decision-making: Construct a multi-objective optimization function according to the processed data, and use the recursive least squares method and the particle swarm optimization algorithm in cooperation to perform parameter identification and optimization, and generate control instructions; Execution: Drive the robotic arm joint module to move according to the control instructions.

Images