Collaborative robot joint adaptive control system and method based on multi-sensor fusion

Abstract

The invention discloses a cooperative robot joint adaptive control system and method based on multi-sensor fusion, and relates to the technical field of robot joint control. According to the invention, through an improved extended Kalman filtering and deep neural network compensation mechanism integrated by the multi-source data fusion unit, sensor noise is effectively suppressed, and multi-source data conflicts are solved; the sensor reliability evaluation unit is based on a dynamic weight distribution mechanism of Bayesian reasoning, and can reduce the fusion weight of the sensor in real time when the sensor is detected to be abnormal, and improve the contribution degree of reliable sensor data. The adaptive control unit is combined with the self-learning optimization unit to form a closed-loop optimization architecture, so that the performance limitation caused by fixed traditional PID parameters is thoroughly solved; through adaptive weighted fusion of multi-source sensor data, the joint state estimation precision is improved; and a deep neural network residual error compensation mechanism is utilized to reduce a torque control steady-state error.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot joint control, in particular to a collaborative robot joint adaptive control system and method based on multi-sensor fusion. BACKGROUND

[0002] In the field of robot technology, joint control has always been a key link to determine the motion accuracy and interaction ability of robots. Traditional robot joint control mainly adopts feedback control methods based on a single sensor, such as position control based on angle encoders or torque control based on current sensors. With the development of technology, various sensors have been applied to robot systems, including angle sensors, current sensors, acceleration sensors, and force / torque sensors. There have been some attempts to fuse multi-sensor data in the prior art, such as state estimation methods based on extended Kalman filtering, sensor fusion methods based on neural networks, and joint torque estimation techniques based on recursive Newton-Euler algorithms. These methods have shown certain advantages in specific scenarios, especially in structured environments, and can provide relatively stable control performance.

[0003] However, the existing multi-sensor fusion control methods still have many shortcomings. First, most methods lack a dynamic evaluation mechanism for sensor reliability, and when a sensor fails or is disturbed, the performance of the entire control system will decrease sharply. Second, existing methods usually use fixed parameter controllers, which are difficult to adapt to environmental changes and load fluctuations, resulting in insufficient control accuracy in complex working conditions. Third, many methods rely on accurate system models, but in actual applications, it is extremely difficult to model accurately due to nonlinear factors such as joint friction and elastic deformation. In addition, the lack of self-learning optimization mechanism makes it difficult for the system to improve performance from actual operation experience, and the tuning of control parameters relies on human experience, which is time-consuming and laborious and difficult to achieve optimal state. In human-robot interaction scenarios, when the robot comes into contact with the environment or people, these shortcomings are particularly evident, affecting control accuracy and posing safety hazards. Therefore, we propose a collaborative robot joint adaptive control system and method based on multi-sensor fusion. SUMMARY

[0004] The purpose of the present application is to solve the problems mentioned in the background art. The present application provides a collaborative robot joint adaptive control system and method based on multi-sensor fusion.

[0005] To achieve the above purpose, the present application specifically adopts the following technical solutions:

[0006] A collaborative robot joint adaptive control system based on multi-sensor fusion, comprising:

[0007] A sensor data acquisition unit integrated in the robot for real-time acquisition of real-time running state signals of the robot;

[0008] A multi-source data fusion unit connected to the sensor data acquisition unit, which receives the multi-source signals collected by the sensor data acquisition unit in real time, and uses an improved extended Kalman filter combined with a deep neural network to fuse sensor data to generate accurate joint state estimation;

[0009] A sensor reliability evaluation unit connected to the sensor data acquisition unit, which dynamically evaluates the running state of each sensor based on Bayesian inference algorithm, and automatically adjusts the weight when detecting faults or interference;

[0010] An adaptive control unit connected to the multi-source data fusion unit and the sensor reliability evaluation unit, including a core unit based on a model reference adaptive controller, which can optimize control parameters in real time according to load fluctuations and environmental changes;

[0011] A self-learning optimization unit integrated with reinforcement learning algorithm, which extracts experience from historical operation data, automatically sets control parameters and improves system performance;

[0012] An output execution unit that sends control signals to joint motors through a drive interface to ensure high-precision position and torque control;

[0013] A communication bus unit connected between each unit to realize data interaction and synchronization between units, supporting real-time feedback loops.

[0014] Further, the sensor data acquisition unit includes:

[0015] An angle sensor for measuring the angular position change of the robot joint;

[0016] A current sensor for monitoring the real-time current signal of the joint drive motor;

[0017] An acceleration sensor for detecting the acceleration dynamic response of the joint;

[0018] A force / torque sensor for sensing the force or torque applied to the joint.

[0019] Further, the fusion of the multi-source data fusion unit includes:

[0020] Preprocessing of multi-source signals transmitted by the sensor data acquisition unit, including signal filtering to remove noise, timestamp alignment to ensure data synchronization, and normalization processing to adapt to subsequent algorithm input;

[0021] The improved extended Kalman filter algorithm is applied to predict the real-time state (such as angle, angular velocity) of the joint based on the preset robot joint dynamics model, and correct the prediction error through the state update equation;

[0022] The integrated deep neural network module learns the nonlinear mapping relationship of sensor data using the trained network structure, analyzes the residual signal and outputs the compensation value to enhance the robustness and accuracy of state estimation;

[0023] The preliminary estimation results of the improved extended Kalman filter and the compensation output of the deep neural network are fused to generate the final joint state estimation value, including angle position, acceleration dynamic response and applied torque size, by using a weighted fusion strategy.

[0024] The accurate joint state estimation is output in real time to the adaptive control unit for dynamic adjustment of the current command of the joint drive motor, realizing smooth motion and force control adaptation of the collaborative robot.

[0025] Further, the Bayesian inference algorithm is based on the following formula:

[0026]

[0027] In the formula, P(failure) represents the prior probability of sensor failure, which is initialized based on historical data and a preset model;

[0028] P(observation|failure) is a likelihood function, which calculates the probability distribution under the failure condition through real-time sensor data (such as abnormal fluctuations in angle, current or acceleration);

[0029] P(observation) is a normalization factor to ensure that the sum of posterior probabilities is 1; it is used to dynamically evaluate the running state of each sensor.

[0030] Further, the evaluation of the sensor reliability evaluation unit includes:

[0031] Real-time acquisition of sensor data, extraction of characteristic parameters (such as signal deviation or noise level);

[0032] Application of Bayesian formula to update posterior probability to identify potential failures or disturbances (such as when P(failure|observation) exceeds the threshold value of 0.8);

[0033] Based on the posterior probability result, automatically adjust the sensor weights in the multi-source data fusion unit (for example, reduce the weight of the faulty sensor to 0.2 and increase the weight of the reliable sensor to 0.8);

[0034] The weight adjustment result is fed back to the fusion process to ensure the robustness of joint state estimation;

[0035] Periodically calibrate the prior probability model to cope with environmental changes or long-term drift.

[0036] Further, the model in the adaptive control unit is a reference dynamics model of the joint, which is constructed based on Newton-Euler equation or Lagrange mechanics, and preset mass, inertia and friction parameters of the joint; the reference dynamics model calculates tracking error by comparing reference output (such as ideal angle trajectory) with actual sensor feedback (state estimation from the multi-source data fusion unit) in real time, and dynamically adjusts proportional coefficient, integral time and differential gain of the PID controller.

[0037] Further, the historical operation data includes joint angle trajectory, actual tracking error, motor current instruction, applied torque, environmental disturbance characteristics and key operation information such as PID parameters (proportional coefficient Kp, integral time Ti and differential gain Td) adjusted by the adaptive control unit.

[0038] Further, the output execution unit includes:

[0039] The drive amplifier module is used for receiving control current instruction generated by the adaptive control unit and converting it into high-power analog signal to directly drive the joint motor to realize accurate position and torque output;

[0040] The real-time feedback monitoring unit collects actual operation data of the motor through built-in sensors (such as Hall effect sensor), compares it with instruction signal to correct deviation, and ensures closed-loop control accuracy;

[0041] The safety protection mechanism integrates overcurrent, overheat and short circuit protection circuit, automatically triggers soft stop or alarm when detecting abnormal working condition, and prevents equipment damage;

[0042] The communication interface module supports industrial standard protocol (such as EtherCAT or CAN bus), realizes low-delay data transmission with the joint driver, and guarantees real-time execution of control instruction and system collaboration.

[0043] Further, the communication bus unit includes:

[0044] The high-speed real-time communication module adopts EtherCAT or CANopen industrial bus protocol, supports ≤1ms cycle synchronous transmission, and ensures unblocked interaction of instruction and feedback data among the multi-source data fusion unit, the sensor reliability evaluation unit and the adaptive control unit;

[0045] The topology management unit constructs master-slave ring network topology, realizes cascading communication among joint control units, aligns data frames through hardware time stamp mechanism, and eliminates timing error caused by transmission delay;

[0046] Data time-sharing multiplexing interface, independent communication time slots are allocated for sensor data acquisition units (angle, current, acceleration, force / torque sensors) to ensure preemptive transmission of high-priority state signals (such as joint overload alarm);

[0047] Redundant communication channel, dual-channel hot backup bus architecture is deployed, when the main channel error rate exceeds 10 -5 , automatically switch to the standby channel to maintain system reliability;

[0048] Protocol conversion gateway, compatible with RS485, Modbus and other traditional device interfaces, realizes data intercommunication with external monitoring system, and protects core control data safety through firewall isolation layer;

[0049] Bandwidth dynamic allocation mechanism, based on control task criticality (such as adaptive control unit current instruction> sensor raw data), real-time adjustment of bandwidth proportion to ensure real-time requirements of closed-loop control;

[0050] Bus state monitoring unit, continuously detects transmission error rate, load rate and node online state, when communication delay> 2ms or node disconnection, triggers system degradation operation mode and reports fault code.

[0051] A cooperative robot joint adaptive control method based on multi-sensor fusion, including the control system of any one of the above, comprising the following steps:

[0052] S1, system initialization stage, load the reference dynamic model parameters of the joint, including the preset mass, inertia and friction coefficient, calibrate the initial settings of the sensor data acquisition unit, and configure the synchronization protocol of the communication bus unit;

[0053] S2, real-time acquisition of multi-source signals by sensor data acquisition unit, including position change of angle sensor, driving current of current sensor, dynamic response of acceleration sensor and applied force value of force / torque sensor;

[0054] S3, apply sensor reliability evaluation unit, calculate the posterior probability of each sensor based on Bayesian inference algorithm, when P(failure|observation) exceeds threshold value 0.8, automatically adjust weight, reduce the weight of faulty sensor to 0.2, and increase the weight of reliable sensor to 0.8;

[0055] S4, in the multi-source data fusion unit, pre-process the collected signals, including signal filtering to remove noise, time stamp alignment to ensure data synchronization, normalization processing to adapt to algorithm input, then apply improved extended Kalman filter to predict joint state, integrate deep neural network to analyze residual error and output compensation value, and use weighted fusion strategy to generate accurate joint state estimation value (such as angle position, acceleration dynamic response and applied torque size);

[0056] S5. In the adaptive control unit, compare the ideal output of the reference dynamic model with the actual state estimate, calculate the tracking error, and dynamically adjust the proportional coefficient Kp, integral time Ti, and derivative gain Td of the PID controller to optimize the control parameters.

[0057] S3. The optimized control current command is sent to the joint motor through the output execution unit, and the drive amplifier module converts the high-power analog signal to achieve high-precision position and torque output. At the same time, the deviation is corrected through the real-time feedback monitoring unit.

[0058] S7. Monitor the execution process and record historical operation data, including joint angle trajectory, tracking error, motor current command, applied torque and adaptive parameter adjustment records;

[0059] S8. Utilize a self-learning optimization unit to integrate reinforcement learning algorithms to analyze historical data, extract empirical patterns, and automatically tune control parameters (such as Kp, Ti, Td) to improve system robustness and performance.

[0060] S9. Data interaction between units is achieved through the communication bus unit. EtherCAT or CANopen protocol is used to ensure synchronous transmission with a cycle of ≤1ms. Real-time feedback loop is supported, and bandwidth is dynamically allocated to prioritize the processing of adaptive control commands.

[0061] S10. Repeat steps two to nine to achieve continuous adaptive control. When a communication delay >2ms or a node is lost, trigger a degraded operation mode and report a fault code to ensure system reliability.

[0062] The beneficial effects of this invention are as follows:

[0063] 1. This invention effectively suppresses sensor noise (such as high-frequency interference of current signals and angle quantization error) and solves multi-source data conflicts (such as the contradiction between the readings of the accelerometer and torque sensor under impact load) by integrating an improved extended Kalman filter and deep neural network (DNN) compensation mechanism through a multi-source data fusion unit.

[0064] 2. The sensor reliability assessment unit of this invention, based on a dynamic weight allocation mechanism using Bayesian inference, can, in real time, reduce the fusion weight (e.g., to 0.2) and increase the contribution of reliable sensor data (e.g., to 0.8) when a sensor anomaly is detected (such as accelerometer drift causing P(fault|observation) > 0.8). This mechanism ensures that even if a single sensor fails or is subjected to strong electromagnetic interference, the system can still maintain the continuity of joint state estimation based on the remaining valid data, avoiding joint jitter or loss of control caused by control command disorder, and significantly improving the system's survivability and control stability under abnormal operating conditions.

[0065] 3、The adaptive control unit of the application combines the closed-loop optimization architecture of the self-learning optimization unit, which completely solves the performance limitations caused by the fixed parameters of the traditional PID. The model reference adaptive controller (MRAC) adjusts the PID parameters (such as the proportional coefficient Kp and the integral time Ti) online according to the dynamic error between the real-time state estimation of the joint (from the multi-source data fusion unit) and the reference model, and quickly responds to load mutations (such as grabbing workpieces of different masses); at the same time, the reinforcement learning algorithm continuously accumulates experience and automatically tunes the control rule library by analyzing historical operation data (such as the optimal Kp / Ti combination under different loads), so that the system can adapt to known task changes.

[0066] 4、The application can respond to changes in joint dynamic load in real time, effectively suppress tracking errors caused by external disturbances; through adaptive weighted fusion of multi-source sensor data, the accuracy of joint state estimation is improved; using the residual compensation mechanism of deep neural network, the steady-state error of torque control is reduced; based on the parameter self-tuning function of reinforcement learning, the system continuously optimizes the control response speed in repeated operations, shortens the tuning period; when sensor failure or communication anomaly is detected, the system automatically switches to a degraded mode to maintain basic motion functions, ensuring that the interruption time in collaborative tasks does not exceed 500ms. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 is the system module diagram of the application;

[0068] Figure 2 is the method flowchart of the application. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely with reference to the drawings in the embodiments of the application.

[0070] Please refer to Figure 1 The application provides a collaborative robot joint adaptive control system based on multi-sensor fusion, which comprises:

[0071] A sensor data acquisition unit is integrated in the robot for real-time acquisition of real-time running state signals of the robot, i.e., can acquire multiple signals such as joint angle, torque, temperature, vibration and position.

[0072] A multi-source data fusion unit is connected to the sensor data acquisition unit and receives the multi-source signals collected by the sensor data acquisition unit in real time, and uses an improved extended Kalman filter combined with a deep neural network to fuse sensor data to generate accurate joint state estimation, i.e., can effectively handle sensor noise and data conflicts, and improve the robustness and real-time performance of state estimation.

[0073] The sensor reliability evaluation unit is connected to the sensor data acquisition unit and dynamically evaluates the operation state of each sensor based on a Bayesian inference algorithm, automatically adjusts the weight when a fault or interference is detected, that is, dynamically optimizes the weight distribution of multi-sensor fusion, and improves the robustness and accuracy of the system under abnormal working conditions.

[0074] The adaptive control unit is connected to the multi-source data fusion unit and the sensor reliability evaluation unit, includes a core unit based on a model reference adaptive controller, can optimize control parameters in real time according to load fluctuations and environmental changes, that is, can generate adaptive control instructions based on joint state estimation, and then dynamically adjust PID parameters according to environmental changes and task requirements, to realize accurate tracking of joint torque and position.

[0075] The self-learning optimization unit integrates a reinforcement learning algorithm, extracts experience from historical operation data, automatically sets control parameters and improves system performance, that is, can extract optimal control strategies from historical operation data through a reinforcement learning algorithm, automatically set control parameters such as PID gain and integral time, improve the response speed of the system, reduce torque and position tracking error, and accumulate experience in long-term operation, realize continuous performance optimization and self-adaptation ability, and cope with complex tasks and environmental changes.

[0076] The output execution unit sends control signals to joint motors through a drive interface to ensure high-precision position and torque control, that is, can respond to instructions from the adaptive control unit in real time, maintain the smoothness and stability of joint motion under dynamic load changes and environmental interference, and improve the operation precision and safety of the collaborative robot.

[0077] The communication bus unit is connected between each unit to realize data interaction and synchronization between units, supports real-time feedback loop, that is, can realize synchronous transmission, support real-time feedback loop, dynamically allocate bandwidth to prioritize adaptive control instructions.

[0078] In this embodiment, preferably, the sensor data acquisition unit includes:

[0079] An angle sensor for measuring the angular position change of the robot joint;

[0080] A current sensor for monitoring the real-time current signal of the joint drive motor;

[0081] An acceleration sensor for detecting the acceleration dynamic response of the joint;

[0082] A force / torque sensor for sensing the force or torque applied to the joint.

[0083] The joint dynamics characteristics can be comprehensively captured through the cooperation of the above sensors, and real-time feedback is provided to the output execution unit. The joint abnormal vibration mode and external impact load can be effectively identified through the joint data of the acceleration sensor and the force / torque sensor, and key fault diagnosis basis is provided for the reliability evaluation unit. The high-resolution position signal of the angle sensor and the motor load information provided by the current sensor are mutually checked, which improves the confidence of joint state estimation, thereby supporting the adaptive control unit to complete the dynamic adjustment of control parameters within milliseconds, and the cooperation mechanism not only enhances the data redundancy, but also maintains the core function of the system through the weight redistribution of the reliability evaluation unit in the event of sensor failure, and ensures the safe and stable operation of the collaborative robot in the scene of precision assembly and human-machine cooperation.

[0084] In the embodiment, preferably, the fusion of the multi-source data fusion unit includes:

[0085] The multi-source signals transmitted by the sensor data acquisition unit are preprocessed, including signal filtering to remove noise, timestamp alignment to ensure data synchronization, and normalization processing to adapt to subsequent algorithm input;

[0086] An improved extended Kalman filter algorithm is applied to predict the real-time state (such as angle, angular velocity) of the joint based on a preset robot joint dynamics model, and to correct the prediction error through a state update equation;

[0087] A deep neural network module is integrated to learn the nonlinear mapping relationship of sensor data using a trained network structure, analyze residual signals and output compensation values to enhance the robustness and accuracy of state estimation;

[0088] The preliminary estimation results of the improved extended Kalman filter and the compensation output of the deep neural network are fused, and a weighted fusion strategy is used to generate the final joint state estimation value, including angle position, acceleration dynamic response and applied torque size;

[0089] The accurate joint state estimation is output in real time to the adaptive control unit for dynamic adjustment of the current command of the joint driving motor, and smooth motion and force control adaptation of the collaborative robot are achieved.

[0090] Through effective fusion of data, sensor noise interference can be reduced, data conflict problems can be solved, and high-precision joint state estimation values can be generated, including real-time angle position, acceleration dynamic response, and applied torque size. Not only does this improve the robustness and real-time performance of state estimation, but it also provides reliable input for the adaptive control unit, ensuring accurate tracking of joint torque and position under load fluctuations and environmental changes, while reducing estimation errors to enhance the overall control performance of collaborative robots, supporting safe and efficient operation in precision assembly and human-robot collaboration applications. In addition, the fused data can also assist the sensor reliability evaluation unit in dynamically adjusting the weight, optimizing the response capability of the system under abnormal conditions, and further improving the adaptability and stability of the control system.

[0091] In this embodiment, preferably, the Bayesian inference algorithm is based on the following formula:

[0092]

[0093] In the formula, P(failure) represents the prior probability of sensor failure, which is initialized based on historical data and a preset model;

[0094] P(observation|failure) is the likelihood function, which calculates the probability distribution under the failure condition through real-time sensor data such as abnormal fluctuations in angle, current, or acceleration;

[0095] P(observation) is the normalization factor, which ensures that the sum of posterior probabilities is 1; it is used to dynamically evaluate the operating state of each sensor.

[0096] In this embodiment, preferably, the evaluation of the sensor reliability evaluation unit includes:

[0097] Real-time acquisition of sensor data and extraction of characteristic parameters (such as signal deviation or noise level);

[0098] Application of Bayesian formula to update posterior probability to identify potential failures or disturbances (such as when P(failure|observation) exceeds the threshold value of 0.8);

[0099] Based on the posterior probability result, automatically adjust the sensor weights in the multi-source data fusion unit (for example, reduce the weight of the faulty sensor to 0.2 and increase the weight of the reliable sensor to 0.8);

[0100] Feedback the weight adjustment result to the fusion process to ensure the robustness of joint state estimation;

[0101] Periodically calibrate the prior probability model to cope with environmental changes or long-term drift.

[0102] The evaluation can dynamically optimize sensor weight allocation, effectively cope with abnormal working conditions, and improve the adaptability and accuracy of the control system. At the same time, combined with a real-time calibration mechanism, it can reduce the impact of long-term drift and ensure the continuity of joint state estimation when faults or disturbances occur, thereby enhancing the safety and operational stability of collaborative robots in precision tasks.

[0103] In this embodiment, preferably, the model in the adaptive control unit is a reference dynamic model of the joint, constructed based on the Newton-Euler equations or Lagrange mechanics, with preset mass, inertia and friction parameters of the joint; the reference dynamic model calculates the tracking error by comparing the reference output (such as the ideal angle trajectory) with the actual sensor feedback (state estimation from the multi-source data fusion unit) in real time, and dynamically adjusts the proportional coefficient, integral time and derivative gain of the PID controller.

[0104] This operation can reduce torque and position tracking errors, enhance the response speed of joints under dynamic load changes, and ensure that collaborative robots can achieve high-precision motion control in scenarios such as precision assembly and human-robot collaboration.

[0105] In this embodiment, preferably, the historical operation data includes key operational information such as joint angle trajectory, actual tracking error, motor current command, applied torque, environmental interference characteristics, and PID parameters (proportional coefficient Kp, integral time Ti, derivative gain Td) adjusted by the adaptive control unit. Combining historical operation data allows for the training of a reinforcement learning model, the construction of a state-action value function mapping relationship, and the dynamic generation of the optimal control strategy.

[0106] In this embodiment, preferably, the output execution unit includes:

[0107] The drive amplifier module is used to receive the control current command generated by the adaptive control unit and convert it into a high-power analog signal to directly drive the joint motor to achieve precise position and torque output.

[0108] The real-time feedback monitoring unit collects actual motor operating data through built-in sensors (such as Hall effect sensors), compares it with command signals to correct deviations, and ensures closed-loop control accuracy.

[0109] The safety protection mechanism integrates overcurrent, overheat and short circuit protection circuits. When abnormal operating conditions are detected, it automatically triggers soft stop or alarm to prevent equipment damage.

[0110] The communication interface module supports industry standard protocols (such as EtherCAT or CAN bus) to achieve low-latency data transmission with the joint driver, ensuring real-time execution of control commands and system coordination.

[0111] The coordinated operation of these modules ensures the smoothness and stability of joint movement under dynamic load changes and environmental disturbances. Specifically, the high bandwidth of the drive amplifier module guarantees a rapid response to current commands, while the real-time feedback monitoring unit reduces position tracking lag through millisecond-level data sampling and correction. The safety protection mechanism can trigger a soft stop within 5 milliseconds of detecting abnormal torque (such as exceeding a threshold of 20%), effectively preventing joint overload damage. The low-latency transmission (<1ms) of the communication interface module ensures real-time synchronization between adaptive control commands and joint actuators, providing a fundamental guarantee for the safe and stable operation of collaborative robots in demanding scenarios such as precision assembly and human-robot collaboration.

[0112] In this embodiment, preferably, the communication bus unit includes:

[0113] The high-speed real-time communication module adopts the EtherCAT or CANopen industrial bus protocol and supports synchronous transmission with a cycle of ≤1ms, ensuring unblocked interaction of commands and feedback data between the multi-source data fusion unit, the sensor reliability assessment unit and the adaptive control unit.

[0114] The topology management unit constructs a master-slave ring network topology, enabling cascaded communication between various joint control units. It aligns data frames through a hardware timestamp mechanism to eliminate timing errors caused by transmission delays.

[0115] The data time-division multiplexing interface allocates independent communication time slots to sensor data acquisition units (angle, current, acceleration, force / torque sensors) to ensure preemptive transmission of high-priority status signals (such as joint overload alarms);

[0116] Redundant communication channels are used, and a dual-channel hot backup bus architecture is deployed. When the main channel bit error rate exceeds 10... -5 Automatically switch to backup channel to maintain system reliability;

[0117] The protocol conversion gateway is compatible with traditional device interfaces such as RS485 and Modbus, enabling data communication with external monitoring systems while protecting the security of core control data through a firewall isolation layer.

[0118] The bandwidth dynamic allocation mechanism adjusts the bandwidth ratio in real time based on the criticality of the control task (e.g., adaptive control unit current command > sensor raw data) to ensure the real-time requirements of closed-loop control.

[0119] The bus status monitoring unit continuously monitors the transmission error rate, load rate, and node online status. When the communication delay is greater than 2ms or a node becomes disconnected, the system is triggered to degrade its operating mode and a fault code is reported.

[0120] The coordinated operation of these units facilitates efficient collaborative operation and real-time response capabilities under dynamic load changes and complex environments. This collaborative mechanism not only enhances data redundancy and fault tolerance but also improves the control accuracy, operational stability, and safety of collaborative robots in demanding scenarios such as precision assembly and human-robot collaboration. For example, in the event of abnormal vibration or external impact, core functions are maintained through weight redistribution, ensuring accurate tracking of joint torque and position.

[0121] The above system can solve the three core problems in the joint control of traditional collaborative robots mentioned in the background technology: 1) Insufficient sensor accuracy and noise interference lead to deviations in joint state estimation; 2) System control instability is caused by single sensor failure or environmental interference; 3) Fixed control parameters are difficult to adapt to dynamic load changes and complex task requirements.

[0122] Specifically:

[0123] 1. Addressing sensor accuracy and noise issues: This system effectively suppresses sensor noise (such as high-frequency interference in current signals and angle quantization errors) and resolves multi-source data conflicts (such as discrepancies in readings from the accelerometer and torque sensor under impact loads) through an improved extended Kalman filter and deep neural network (DNN) compensation mechanism integrated by a multi-source data fusion unit.

[0124] 2. Addressing robustness issues under sensor failure and interference: The sensor reliability assessment unit employs a dynamic weight allocation mechanism based on Bayesian inference. When a sensor anomaly is detected (e.g., accelerometer drift causing P(failure|observation) > 0.8), its fusion weight is adjusted in real-time (e.g., reduced to 0.2), while the contribution of reliable sensor data is increased (e.g., increased to 0.8). This mechanism ensures that even if a single sensor fails or is subjected to strong electromagnetic interference, the system can still maintain the continuity of joint state estimation based on the remaining valid data, avoiding joint jitter or loss of control caused by control command disorder. This significantly improves the system's survivability and control stability under abnormal operating conditions.

[0125] 3. Addressing the dynamic adaptability issue: The closed-loop optimization architecture, composed of an adaptive control unit and a self-learning optimization unit, completely solves the performance limitations caused by fixed parameters in traditional PID controllers. The Model Reference Adaptive Controller (MRAC) adjusts PID parameters (such as proportional coefficient Kp and integral time Ti) online based on the dynamic error between the real-time joint state estimation (from the multi-source data fusion unit) and the reference model, enabling rapid response to load changes (such as gripping workpieces of different weights). Simultaneously, the reinforcement learning algorithm continuously accumulates experience and automatically tunes the control rule base by analyzing historical operation data (such as the optimal Kp / Ti combination under different loads), enabling the system to adapt not only to known task changes.

[0126] Through the above-mentioned technological innovations, this system has significantly improved the control accuracy, anti-interference ability, and task adaptability of collaborative robot joints in precision operation, human-machine collaboration, and unstructured environments.

[0127] Please see Figure 2 The present invention also provides a collaborative robot joint adaptive control method based on multi-sensor fusion, including the control system of any of the above, comprising the following steps:

[0128] S1. During the system initialization phase, load the reference dynamic model parameters of the joint, including the preset mass, inertia and friction coefficient, calibrate the initial settings of the sensor data acquisition unit, and configure the synchronization protocol of the communication bus unit.

[0129] S2. Real-time acquisition of multi-source signals through the sensor data acquisition unit, including position changes of the angle sensor, drive current of the current sensor, dynamic response of the acceleration sensor, and applied force value of the force / torque sensor;

[0130] S3. The sensor reliability assessment unit calculates the posterior probability of each sensor based on the Bayesian inference algorithm. When P(fault|observation) exceeds the threshold of 0.8, the weights are automatically adjusted, reducing the weight of faulty sensors to 0.2 and increasing the weight of reliable sensors to 0.8.

[0131] S4. In the multi-source data fusion unit, the acquired signals are preprocessed, including signal filtering to remove noise, timestamp alignment to ensure data synchronization, normalization to adapt to the algorithm input, and then an improved extended Kalman filter is applied to predict the joint state. A deep neural network is integrated to analyze the residuals and output compensation values. A weighted fusion strategy is used to generate accurate joint state estimates (such as angular position, acceleration dynamic response and applied torque magnitude).

[0132] S5. In the adaptive control unit, compare the ideal output of the reference dynamic model with the actual state estimate, calculate the tracking error, and dynamically adjust the proportional coefficient Kp, integral time Ti, and derivative gain Td of the PID controller to optimize the control parameters.

[0133] S3. The optimized control current command is sent to the joint motor through the output execution unit, and the drive amplifier module converts the high-power analog signal to achieve high-precision position and torque output. At the same time, the deviation is corrected through the real-time feedback monitoring unit.

[0134] S7. Monitor the execution process and record historical operation data, including joint angle trajectory, tracking error, motor current command, applied torque and adaptive parameter adjustment records;

[0135] S8. Utilize a self-learning optimization unit to integrate reinforcement learning algorithms to analyze historical data, extract empirical patterns, and automatically tune control parameters (such as Kp, Ti, Td) to improve system robustness and performance.

[0136] S9. Data interaction between units is achieved through the communication bus unit. EtherCAT or CANopen protocol is used to ensure synchronous transmission with a cycle of ≤1ms. Real-time feedback loop is supported, and bandwidth is dynamically allocated to prioritize the processing of adaptive control commands.

[0137] S10. Repeat steps two to nine to achieve continuous adaptive control. When a communication delay >2ms or a node is lost, trigger a degraded operation mode and report a fault code to ensure system reliability.

[0138] The aforementioned adaptive control method can respond to changes in joint dynamic load in real time and effectively suppress tracking errors caused by external disturbances; it improves the accuracy of joint state estimation through adaptive weighted fusion of multi-source sensor data; it reduces torque control steady-state error by utilizing a deep neural network residual compensation mechanism; and it enables the system to continuously optimize control response speed and shorten tuning cycle during repetitive operations based on reinforcement learning parameter self-tuning function. When a sensor fault or communication anomaly is detected, it autonomously switches to a degradation mode to maintain basic motion functions, ensuring that the collaborative task interruption time does not exceed 500ms.

[0139] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A collaborative robot joint adaptive control system based on multi-sensor fusion, characterized in that, include: The sensor data acquisition unit, integrated inside the robot, is used to collect real-time operating status signals of the robot. The multi-source data fusion unit is connected to the sensor data acquisition unit and receives multi-source signals acquired by the sensor data acquisition unit in real time. It uses an improved extended Kalman filter combined with a deep neural network to fuse sensor data to generate accurate joint state estimates. The sensor reliability assessment unit is connected to the sensor data acquisition unit. It dynamically assesses the operating status of each sensor based on the Bayesian inference algorithm and automatically adjusts the weights when a fault or interference is detected. The adaptive control unit, connected to the multi-source data fusion unit and the sensor reliability assessment unit, includes the core unit of the model reference adaptive controller, which can optimize control parameters in real time according to load fluctuations and environmental changes. The self-learning optimization unit integrates reinforcement learning algorithms to extract experience from historical operation data, automatically tune control parameters, and improve system performance. The output execution unit sends control signals to the joint motor through the drive interface to ensure high-precision position and torque control; The communication bus unit connects all units to enable data interaction and synchronization between them, and supports real-time feedback loops.

2. The collaborative robot joint adaptive control system based on multi-sensor fusion according to claim 1, characterized in that, The sensor data acquisition unit includes: Angle sensors are used to measure changes in the angular position of robot joints; A current sensor is used to monitor the real-time current signal of the joint drive motor. Accelerometer, used to detect the dynamic acceleration response of the joint; Force / torque sensors are used to sense the magnitude of force or torque applied to a joint.

3. The collaborative robot joint adaptive control system based on multi-sensor fusion according to claim 1, characterized in that, The fusion of the multi-source data fusion unit includes: The multi-source signals transmitted by the sensor data acquisition unit are preprocessed, including signal filtering to remove noise, timestamp alignment to ensure data synchronization, and normalization to adapt to subsequent algorithm inputs. An improved extended Kalman filter algorithm is applied to predict the real-time state of the joint (such as angle and angular velocity) based on a preset robot joint dynamics model, and the prediction error is corrected by the state update equation. An integrated deep neural network module is used to learn the nonlinear mapping relationship of sensor data using a trained network structure, analyze residual signals and output compensation values ​​to enhance the robustness and accuracy of state estimation. The preliminary estimation results of the improved extended Kalman filter are combined with the compensation output of the deep neural network, and a weighted fusion strategy is used to generate the final joint state estimate, including angular position, acceleration dynamic response and applied torque magnitude. The precise joint state estimate is output to the adaptive control unit in real time to dynamically adjust the current command of the joint drive motor, thereby enabling smooth movement and force control adaptation of the collaborative robot.

4. The collaborative robot joint adaptive control system based on multi-sensor fusion according to claim 1, characterized in that, The Bayesian inference algorithm is based on the following formula: In the formula, P(fault) represents the prior probability of sensor failure, which is initialized based on historical data and a preset model; P(observation|fault) is a likelihood function, which calculates the probability distribution under fault conditions using real-time sensor data (such as abnormal fluctuations in angle, current, or acceleration). P(observation) is a normalization factor that ensures that the sum of the posterior probabilities is 1. Used to dynamically evaluate the operating status of each sensor.

5. The collaborative robot joint adaptive control system based on multi-sensor fusion according to claim 1, characterized in that, The evaluation by the sensor reliability evaluation unit includes: Real-time acquisition of sensor data and extraction of characteristic parameters (such as signal deviation or noise level); Apply Bayes' theorem to update the posterior probability and identify potential faults or disturbances (e.g., when P(fault|observation) exceeds the threshold of 0.8). Based on the posterior probability results, the sensor weights in the multi-source data fusion unit are automatically adjusted (e.g., the weight of faulty sensors is reduced to 0.2, and the weight of reliable sensors is increased to 0.8). The weight adjustment results are fed back into the fusion process to ensure the robustness of joint state estimation; Regularly calibrate the prior probability model to cope with environmental changes or long-term drift.

6. The collaborative robot joint adaptive control system based on multi-sensor fusion according to claim 1, characterized in that, The model in the adaptive control unit is a reference dynamic model of the joint, constructed based on the Newton-Euler equations or Lagrange mechanics, with preset mass, inertia and friction parameters of the joint; the reference dynamic model calculates the tracking error by comparing the reference output (such as the ideal angle trajectory) with the actual sensor feedback (state estimation from the multi-source data fusion unit) in real time, and dynamically adjusts the proportional coefficient, integral time and derivative gain of the PID controller.

7. The collaborative robot joint adaptive control system based on multi-sensor fusion according to claim 1, characterized in that, The historical operation data includes key operating information such as joint angle trajectory, actual tracking error, motor current command, applied torque, environmental interference characteristics, and PID parameters (proportional coefficient Kp, integral time Ti, derivative gain Td) adjusted by the adaptive control unit.

8. The collaborative robot joint adaptive control system based on multi-sensor fusion according to claim 1, characterized in that, The output execution unit includes: The drive amplifier module is used to receive the control current command generated by the adaptive control unit and convert it into a high-power analog signal to directly drive the joint motor to achieve precise position and torque output. The real-time feedback monitoring unit collects actual motor operating data through built-in sensors (such as Hall effect sensors), compares it with command signals to correct deviations, and ensures closed-loop control accuracy. The safety protection mechanism integrates overcurrent, overheat and short circuit protection circuits. When abnormal operating conditions are detected, it automatically triggers soft stop or alarm to prevent equipment damage. The communication interface module supports industry standard protocols (such as EtherCAT or CAN bus) to achieve low-latency data transmission with the joint driver, ensuring real-time execution of control commands and system coordination.

9. The collaborative robot joint adaptive control system based on multi-sensor fusion according to claim 1, characterized in that, The communication bus unit includes: The high-speed real-time communication module adopts the EtherCAT or CANopen industrial bus protocol and supports synchronous transmission with a cycle of ≤1ms, ensuring unblocked interaction of commands and feedback data between the multi-source data fusion unit, the sensor reliability assessment unit and the adaptive control unit. The topology management unit constructs a master-slave ring network topology, enabling cascaded communication between various joint control units. It aligns data frames through a hardware timestamp mechanism to eliminate timing errors caused by transmission delays. The data time-division multiplexing interface allocates independent communication time slots to sensor data acquisition units (angle, current, acceleration, force / torque sensors) to ensure preemptive transmission of high-priority status signals (such as joint overload alarms); Redundant communication channels are used, and a dual-channel hot backup bus architecture is deployed. When the main channel bit error rate exceeds 10... -5 Automatically switch to backup channel to maintain system reliability; The protocol conversion gateway is compatible with traditional device interfaces such as RS485 and Modbus, enabling data communication with external monitoring systems while protecting the security of core control data through a firewall isolation layer. The bandwidth dynamic allocation mechanism adjusts the bandwidth ratio in real time based on the criticality of the control task (e.g., adaptive control unit current command > sensor raw data) to ensure the real-time requirements of closed-loop control. The bus status monitoring unit continuously monitors the transmission error rate, load rate, and node online status. When the communication delay is greater than 2ms or a node becomes disconnected, the system is triggered to degrade its operating mode and a fault code is reported.

10. A collaborative robot joint adaptive control method based on multi-sensor fusion, comprising the control system described in any one of claims 1-9, characterized in that, Includes the following steps: S1. During the system initialization phase, load the reference dynamic model parameters of the joint, including the preset mass, inertia and friction coefficient, calibrate the initial settings of the sensor data acquisition unit, and configure the synchronization protocol of the communication bus unit. S2. Real-time acquisition of multi-source signals through the sensor data acquisition unit, including position changes of the angle sensor, drive current of the current sensor, dynamic response of the acceleration sensor, and applied force value of the force / torque sensor; S3. The sensor reliability assessment unit calculates the posterior probability of each sensor based on the Bayesian inference algorithm. When P(fault|observation) exceeds the threshold of 0.8, the weights are automatically adjusted, reducing the weight of faulty sensors to 0.2 and increasing the weight of reliable sensors to 0.

8. S4. In the multi-source data fusion unit, the acquired signals are preprocessed, including signal filtering to remove noise, timestamp alignment to ensure data synchronization, normalization to adapt to the algorithm input, and then an improved extended Kalman filter is applied to predict the joint state. A deep neural network is integrated to analyze the residuals and output compensation values. A weighted fusion strategy is used to generate accurate joint state estimates (such as angular position, acceleration dynamic response and applied torque magnitude). S5. In the adaptive control unit, compare the ideal output of the reference dynamic model with the actual state estimate, calculate the tracking error, and dynamically adjust the proportional coefficient Kp, integral time Ti, and derivative gain Td of the PID controller to optimize the control parameters. S6. The optimized control current command is sent to the joint motor through the output execution unit, and the drive amplifier module converts the high-power analog signal to achieve high-precision position and torque output. At the same time, the deviation is corrected through the real-time feedback monitoring unit. S7. Monitor the execution process and record historical operation data, including joint angle trajectory, tracking error, motor current command, applied torque and adaptive parameter adjustment records; S8. Utilize a self-learning optimization unit to integrate reinforcement learning algorithms to analyze historical data, extract empirical patterns, and automatically tune control parameters (such as Kp, Ti, Td) to improve system robustness and performance. S9. Data interaction between units is achieved through the communication bus unit. EtherCAT or CANopen protocol is used to ensure synchronous transmission with a cycle of ≤1ms. Real-time feedback loop is supported, and bandwidth is dynamically allocated to prioritize the processing of adaptive control commands. S10. Repeat steps two to nine to achieve continuous adaptive control. When a communication delay >2ms or a node is lost, trigger a degraded operation mode and report a fault code to ensure system reliability.

Images