Industrial robot real-time adaptive control method and system based on digital twinning

Abstract

The invention discloses an industrial robot real-time adaptive control method and system based on digital twinning, and relates to the technical field of industrial robots. The digital twin engine module runs a high-fidelity dynamics simulation model and an environment interaction model, performs real-time state estimation, abnormal working condition recognition and twin parameter dynamic updating, is seamlessly integrated with the control execution module, and provides decision support with high robustness and high adaptability for an industrial scene; the adaptive control module performs online rolling optimization on a control strategy based on a deep reinforcement learning algorithm, generates joint space trajectory correction, tail end precision compensation and dynamic load adaptability optimal instructions, and realizes parameter adaptive setting through fuzzy logic or a neural network; and the fault diagnosis module performs multi-scale time sequence analysis by using an LSTM and convolutional neural network fusion model, detects position offset, moment sudden change or temperature overrun and other abnormalities, and triggers emergency shutdown, sound-light alarm and an adaptive recovery strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial robots, in particular to an industrial robot real-time adaptive control method and system based on digital twinning. BACKGROUND

[0002] Industrial robots are automatic control, repeatable programming, multi-purpose, and can be programmed on 3 or more axes. Industrial robots are composed of three basic parts: main body, drive system and control system. Industrial robots can be fixed or mobile, and are used in industrial automation systems. A machine (mechanical device) whose mechanism is usually composed of a series of joints hinged or sliding with each other, has several degrees of freedom, and has similar action functions as human arms, to grab or carry objects (workpieces or tools) and other purposes. Industrial robots are widely used in electronics, logistics, chemical industry and other industrial fields.

[0003] The existing industrial robot adaptive control method has the following main disadvantages: 1. Poor adaptability, unable to quickly adapt to different environments, load changes or work path adjustments, resulting in low efficiency; 2. Weak robustness and anti-interference ability, in complex working conditions (such as multi-sensor fusion, dynamic environment change), the existing algorithm is easily affected by external interference, resulting in unstable welding quality, insufficient assembly precision and other problems; 3. Insufficient real-time, slow response speed when processing complex tasks, unable to meet the high efficiency production demand. Therefore, we propose an industrial robot real-time adaptive control method and system based on digital twinning. SUMMARY

[0004] The purpose of the present application is to solve the problems mentioned in the background art, and the present application provides an industrial robot real-time adaptive control method and system based on digital twinning.

[0005] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0006] The industrial robot real-time adaptive control method based on digital twinning comprises:

[0007] Real-time acquisition of industrial robot operation data, acquisition of multi-dimensional physical information based on multi-modal sensor fusion technology, and millisecond-level timestamp synchronization and data preprocessing through industrial Internet of Things edge computing nodes;

[0008] A digital twin model is constructed to simulate the physical system, and the preprocessed real-time operation data stream is transmitted to the cloud or locally deployed digital twin engine. The engine has built-in high-fidelity dynamic simulation model and environment interaction model, and uses adaptive Kalman filter and long short-term memory network for operation state estimation and abnormal condition identification, and dynamically updates the twin parameters to match the current state of the physical entity;

[0009] Based on the real-time updated digital twin, parallel simulation prediction is carried out, and the online rolling optimization control strategy is optimized using the deep reinforcement learning algorithm to generate the optimal control instruction set including joint space trajectory correction, end-point accuracy compensation and dynamic load adaptability;

[0010] The adaptive control algorithm is executed to output control instructions, and the optimized control instructions are sent to the robot's real-time controller through a low-latency deterministic network, driving the physical actuator to achieve closed-loop control. At the same time, the twin model and control strategy are continuously iterated and optimized based on the online learning mechanism.

[0011] Furthermore, the real-time collection of the industrial robot's operating data obtains position, speed and torque information based on a multi-sensor network, wherein the sensor network covers multiple types of encoders, accelerometers, and torque sensors to accurately obtain multi-dimensional physical information such as joint position, speed, torque, six-dimensional force / torque of the end effector, motor current and temperature.

[0012] Furthermore, the real-time collection of the operating data of the industrial robot includes the following steps:

[0013] Step 11: Deploy a multi-sensor network at each key joint of the industrial robot. This includes encoders for real-time monitoring of joint angle position, accelerometers for high-precision measurement of linear acceleration to infer velocity, and torque sensors for collecting torque data. Six-dimensional force / torque sensors are integrated into the end effector to obtain three-dimensional force and torque information. Furthermore, current sensors monitor motor current changes, and temperature sensors track the thermal state of key components to ensure comprehensive coverage of multi-dimensional physical information.

[0014] Step 12: Use the IIoT edge computing nodes to synchronize millisecond-level timestamps for all sensor data, using the IEEE 1588 precision time protocol to align clocks within the network, eliminating time drift errors and ensuring the timing consistency of data collection.

[0015] Step 13: Perform data preprocessing at the edge computing node, including applying a Kalman filter to smooth the position signal, using a median filter to remove accelerometer noise, normalizing the torque data, and implementing an outlier detection algorithm to remove invalid data points, thereby generating a high-quality, low-latency preprocessed dataset.

[0016] Step 14: Based on multimodal sensor fusion technology, integrate multi-dimensional information such as position, speed, torque, force / torque, current and temperature, and generate a unified state vector through a weighted fusion algorithm to provide real-time input for the subsequent digital twin model.

[0017] Furthermore, the digital twin model includes a physical dynamics model and a data-driven machine learning model. The physical dynamics model is constructed based on the Lagrange equation or Newton-Euler dynamics theory to accurately simulate joint torque, inertia effect and friction characteristics; the data-driven machine learning model integrates supervised learning and reinforcement learning frameworks, and is trained using historical operating data, including long and short-term memory networks for time series prediction and convolutional neural networks for processing visual sensor input, to enhance the model's generalization ability for unknown working conditions.

[0018] Furthermore, the adaptive control algorithm in the execution of the adaptive control algorithm to output control instructions adopts fuzzy logic or neural network for parameter optimization, wherein the fuzzy logic controller constructs a fuzzy rule set based on the expert knowledge base, quantifies the input variables through the membership function, and infers the output control gain adjustment amount in real time to cope with the system nonlinearity and uncertainty; the neural network method adopts a deep feedforward network or a loop structure, uses historical operation data to train weight parameters online, optimizes the objective function through the back propagation algorithm, realizes adaptive parameter tuning, and ensures that the control instructions remain robust and stable in a dynamic environment.

[0019] Furthermore, the step of executing the adaptive control algorithm to output the control instructions and transmitting the control instructions to the robot actuator via the industrial bus protocol comprises the following steps:

[0020] Step 21: Format the optimized control instructions into a standard data frame that complies with the industrial bus protocol, including instruction type, joint position, speed, and torque parameters, and embed a millisecond-level timestamp to ensure timing consistency;

[0021] Step 22: Add a cyclic redundancy check code or hash check value to the data frame to perform real-time error detection and data integrity verification. If a transmission error is detected, the automatic retransmission mechanism is triggered to resend the instruction through the low-latency network;

[0022] Step 23: Broadcast the verified data frames to each actuator node of the robot via industrial Ethernet or fieldbus, using a deterministic scheduling algorithm to ensure transmission priority, avoid network congestion, and meet sub-millisecond real-time requirements;

[0023] Step 24: After receiving the data frame, the actuator node parses the frame header and load content, extracts the control instruction and converts it into a motor drive signal, and feeds back a confirmation signal to the digital twin engine to update the closed-loop control state.

[0024] Step 25: Dynamically adjust transmission parameters based on the feedback signal, such as bandwidth allocation or retransmission timeout threshold, and optimize bus protocol configuration in combination with online learning mechanism to improve system robustness and adaptability.

[0025] Furthermore, the method also includes a fault diagnosis module to detect anomalies and trigger a safety mechanism, specifically including:

[0026] Step 31: Based on the real-time state estimation data from the digital twin engine, a fusion model of long short-term memory network and convolutional neural network is applied to perform multi-scale time series analysis on the operating data to identify abnormal patterns such as position offset, torque mutation, or temperature overrun. The fault risk level is quantified by setting thresholds and probability distribution models.

[0027] Step 32: When a high-risk anomaly is detected, a multi-level safety mechanism is automatically triggered, including: a first-level response, which sends an emergency shutdown command to the robot controller via a low-latency network, executing joint braking and power cutoff; a second-level response, which activates the audible and visual alarms and pushes fault codes to the monitoring center; and a third-level response, which combines digital twin simulation results to generate adaptive recovery strategies, such as thermal management adjustments or load redistribution, to minimize production interruptions.

[0028] Step 33: Utilize the online learning mechanism to continuously collect fault event data, optimize the diagnostic model parameters through reinforcement learning, and improve the generalization ability for unknown faults and system robustness;

[0029] Step 34: After the safety mechanism is executed, the fault diagnosis results are fed back to the digital twin engine, and the interaction parameters between the twin model and the environment are dynamically updated to ensure the real-time adaptability of the subsequent control strategy.

[0030] Furthermore, the method integrates a cloud platform to achieve remote monitoring and historical data analysis, specifically including:

[0031] Step 41: Deploy a cloud server cluster and configure a distributed database to store the industrial robot's historical operation data, including position, speed, torque, temperature, and fault event records, supporting high-throughput data writing and long-term archiving.

[0032] Step 42: Develop a web-based remote monitoring interface, integrate a real-time data stream visualization module, synchronize the digital twin engine status via the WebSocket protocol, and implement dynamic graphical display of joint trajectories, end-point forces / torques, and energy consumption indicators. Also, embed an alarm notification system to automatically push notifications to mobile devices or PCs when an anomaly is detected.

[0033] Step 43: Build a historical data analysis module, apply time series analysis algorithms to mine operational data patterns, predict equipment remaining life, maintenance cycles, and performance degradation trends, and generate optimization recommendation reports;

[0034] Step 44: Feedback the analysis results to the local or cloud-based digital twin engine through the RESTful API interface to dynamically update the twin model parameters and control strategy to improve the accuracy and robustness of adaptive control;

[0035] Step 45. Implement end-to-end security mechanisms, including data encryption, access control policies, and audit logs, to ensure the confidentiality and integrity of remote monitoring and data transmission.

[0036] A real-time adaptive control system for an industrial robot based on digital twins, the system comprising:

[0037] The real-time data acquisition module integrates a multimodal sensor network and is deployed at each joint and end effector of the industrial robot. It includes encoders, accelerometers, torque sensors, six-dimensional force / torque sensors, current sensors, and temperature sensors. It is used to collect multi-dimensional physical information such as position, speed, torque, force / torque, current, and temperature with high precision.

[0038] The digital twin engine module is deployed in a cloud server cluster or a local high-performance computing node. It runs a high-fidelity dynamics simulation model and an environmental interaction model, and combines adaptive Kalman filtering with a long-short-term memory network for real-time state estimation, abnormal condition identification, and dynamic update of twin parameters.

[0039] The adaptive control module uses an online rolling optimization control strategy based on a deep reinforcement learning algorithm to generate optimal instructions for joint space trajectory correction, end-point accuracy compensation, and dynamic load adaptability, and implements parameter adaptive tuning through fuzzy logic or neural networks;

[0040] The fault diagnosis module uses a fusion model of LSTM and convolutional neural networks to perform multi-scale time series analysis to detect anomalies such as position offset, torque mutation, or temperature exceeding the limit, triggering emergency shutdown, audible and visual alarms, and adaptive recovery strategies;

[0041] Remote monitoring platform, deploying a cloud-based distributed database, supporting a web-based visualization interface, synchronizing data streams via the WebSocket protocol, and integrating a RESTful API interface to feedback analysis results;

[0042] The network transmission module uses industrial Ethernet or fieldbus to achieve low-latency deterministic data transmission, and ensures instruction integrity and sub-millisecond real-time performance through cyclic redundancy check code, hash check and deterministic scheduling algorithm.

[0043] Furthermore, the real-time data acquisition module includes a data quality assessment unit, which performs real-time credibility analysis on sensor data based on preset thresholds, eliminates outliers and triggers a redundant sensor switching mechanism; the digital twin engine module integrates a hybrid simulation architecture, and the physical dynamics model and the data-driven model operate collaboratively through a weighted fusion algorithm, supporting online model parameter calibration and uncertainty quantification; the adaptive control module is embedded in a multi-objective optimization framework, combines energy consumption, accuracy and wear indicators, and generates a Pareto optimal solution set through deep reinforcement learning rolling optimization.

[0044] The beneficial effects of the present invention are as follows:

[0045] 1. The digital twin engine module in the present invention is deployed in a cloud server cluster or a local high-performance computing node, runs a high-fidelity dynamic simulation model and an environmental interaction model, combines adaptive Kalman filtering with a long short-term memory network for real-time state estimation, abnormal condition identification, and dynamic update of twin parameters. It achieves processing latency as low as 5ms through a GPU-accelerated computing framework and seamlessly integrates with the control execution module, providing highly robust and adaptable decision support for industrial scenarios.

[0046] 2. The adaptive control module in the present invention adopts an online rolling optimization control strategy based on a deep reinforcement learning algorithm to generate optimal instructions for joint space trajectory correction, terminal precision compensation, and dynamic load adaptability, and realizes parameter adaptive adjustment through fuzzy logic or neural networks. It works in conjunction with the state estimation and anomaly recognition modules to form a closed-loop adaptive control loop, supports a 100Hz update frequency, and has a processing delay of less than 5ms, significantly improving trajectory tracking accuracy and system anti-interference capability.

[0047] 3. The fault diagnosis module of the present invention uses the LSTM and convolutional neural network fusion model to perform multi-scale time series analysis to detect anomalies such as position offset, torque mutation or temperature exceeding the limit, triggering emergency shutdown, sound and light alarm and adaptive recovery strategy.

[0048] 4. The remote monitoring platform of the present invention utilizes the high concurrent processing capability of the distributed database to support thousands of data read and write operations per second, ensuring efficient storage and query of large-scale time series data. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of the method of the present invention;

[0050] Figure 2 This is a flow chart of the real-time collection of operating data of an industrial robot in the present invention;

[0051] Figure 3 It is a flow chart of executing the adaptive control algorithm to output control instructions in the present invention;

[0052] Figure 4 is a flow chart of the fault diagnosis module in the present application to detect abnormalities and trigger safety mechanisms;

[0053] Figure 5 is a flow chart of the cloud platform in the present application to realize remote monitoring and historical data analysis;

[0054] Figure 6 is a working block diagram of the system in the present application. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0056] Please refer to Figure 1 - Figure 5 The present application provides a real-time adaptive control method for industrial robots based on digital twinning, which comprises:

[0057] Real-time acquisition of operation data of the industrial robot, acquisition of multi-dimensional physical information based on multi-modal sensor fusion technology, and millisecond-level timestamp synchronization and data preprocessing through an Industrial Internet of Things (IIoT) edge computing node;

[0058] Construction of a digital twinning model to simulate the physical system, transmission of the pre-processed real-time operation data stream to a cloud or locally deployed digital twinning engine, the engine having a high-fidelity dynamics simulation model and an environment interaction model built-in, use of adaptive Kalman filtering and long short-term memory network (LSTM) for operation state estimation and abnormal working condition identification, and dynamic updating of the twin parameters to match the current state of the physical entity;

[0059] Parallel simulation prediction based on the real-time updated digital twin, online rolling optimization of the control strategy using deep reinforcement learning (DRL) algorithm, and generation of an optimal control instruction set containing joint space trajectory correction, end precision compensation and dynamic load adaptability;

[0060] Execution of the adaptive control algorithm to output control instructions, delivery of the optimized control instructions to the robot real-time controller through a low-latency deterministic network to drive the physical actuator to realize closed-loop control, and continuous iterative optimization of the twin model and control strategy based on the online learning mechanism.

[0061] In this embodiment, preferably, real-time acquisition of operation data of the industrial robot is based on a multi-sensor network to acquire position, speed and torque information, wherein the sensor network covers multiple types of encoders, accelerometers and torque sensors to accurately acquire multi-dimensional physical information such as joint position, speed, torque, six-dimensional force / torque of the end effector, motor current and temperature.

[0062] In this embodiment, preferably, real-time collection of the operating data of the industrial robot includes the following steps:

[0063] Step 11: Deploy a multi-sensor network at each key joint of the industrial robot. This includes encoders for real-time monitoring of joint angle position, accelerometers for high-precision measurement of linear acceleration to infer velocity, and torque sensors for collecting torque data. Six-dimensional force / torque sensors are integrated into the end effector to obtain three-dimensional force and torque information. Furthermore, current sensors monitor motor current changes, and temperature sensors track the thermal state of key components to ensure comprehensive coverage of multi-dimensional physical information.

[0064] Step 12: Use the Industrial Internet of Things (IIoT) edge computing node to synchronize all sensor data with millisecond-level timestamps. Use the IEEE 1588 Precision Time Protocol (PTP) to align clocks within the network, eliminate time drift errors, and ensure the timing consistency of data collection.

[0065] Step 13: Perform data preprocessing at the edge computing node, including applying a Kalman filter to smooth the position signal, using a median filter to remove accelerometer noise, normalizing the torque data, and implementing an outlier detection algorithm to remove invalid data points, thereby generating a high-quality, low-latency preprocessed dataset.

[0066] Step 14: Based on multimodal sensor fusion technology, integrate multi-dimensional information such as position, speed, torque, force / torque, current and temperature, and generate a unified state vector through a weighted fusion algorithm to provide real-time input for the subsequent digital twin model.

[0067] In this embodiment, preferably, the digital twin model includes a physical dynamics model and a data-driven machine learning model, wherein the physical dynamics model is constructed based on the Lagrange equation or Newton-Euler dynamics theory to accurately simulate joint torque, inertia effect and friction characteristics; the data-driven machine learning model integrates supervised learning and reinforcement learning frameworks, and is trained using historical operating data, including long short-term memory networks (LSTMs) for time series prediction and convolutional neural networks (CNNs) for processing visual sensor inputs, so as to enhance the model's generalization ability for unknown working conditions.

[0068] In this embodiment, preferably, the adaptive control algorithm in the adaptive control algorithm output control instructions adopts fuzzy logic or neural network for parameter optimization, wherein the fuzzy logic controller constructs a fuzzy rule set based on the expert knowledge base, quantifies the input variables (such as position deviation, speed error and external disturbance) through the membership function, and infers the output control gain adjustment amount in real time to cope with the system nonlinearity and uncertainty; the neural network method adopts a deep feedforward network or a loop structure, uses historical operation data to train weight parameters online, and optimizes the objective function (such as trajectory tracking accuracy or energy consumption index) through the back propagation algorithm to achieve adaptive parameter setting, ensuring that the control instructions remain robust and stable in a dynamic environment.

[0069] In this embodiment, preferably, executing the adaptive control algorithm to output the control instructions and transmitting the control instructions to the robot actuator via the industrial bus protocol includes the following steps:

[0070] Step 21: Format the optimized control instructions into a standard data frame that complies with an industrial bus protocol (such as EtherCAT or PROFIBUS), including instruction type, joint position, speed, and torque parameters, and embed a millisecond-level timestamp to ensure timing consistency.

[0071] Step 22: Add a cyclic redundancy check (CRC) code or hash check value to the data frame to perform real-time error detection and data integrity verification. If a transmission error is detected, the automatic retransmission mechanism is triggered to resend the instruction through the low-latency network.

[0072] Step 23: Broadcast the verified data frames to each actuator node of the robot via industrial Ethernet or fieldbus, using a deterministic scheduling algorithm to ensure transmission priority, avoid network congestion, and meet sub-millisecond real-time requirements;

[0073] Step 24: After receiving the data frame, the actuator node parses the frame header and load content, extracts the control instruction and converts it into a motor drive signal, and feeds back a confirmation signal to the digital twin engine to update the closed-loop control state.

[0074] Step 25: Dynamically adjust transmission parameters based on the feedback signal, such as bandwidth allocation or retransmission timeout threshold, and optimize bus protocol configuration in combination with online learning mechanism to improve system robustness and adaptability.

[0075] In this embodiment, preferably, the method further includes a fault diagnosis module to detect anomalies and trigger a safety mechanism, specifically including:

[0076] Step 31: Based on the real-time state estimation data from the digital twin engine, a long short-term memory network (LSTM) and convolutional neural network (CNN) fusion model is applied to perform multi-scale time series analysis on the operating data to identify abnormal patterns such as position offset, torque mutation, or temperature overrun. The fault risk level is quantified by setting thresholds and probability distribution models.

[0077] Step 32: When a high-risk anomaly is detected, a multi-level safety mechanism is automatically triggered, including: a first-level response, which sends an emergency shutdown command to the robot controller via a low-latency network, executing joint braking and power cutoff; a second-level response, which activates the audible and visual alarms and pushes fault codes to the monitoring center; and a third-level response, which combines digital twin simulation results to generate adaptive recovery strategies, such as thermal management adjustments or load redistribution, to minimize production interruptions.

[0078] Step 33: Utilize the online learning mechanism to continuously collect fault event data, optimize the diagnostic model parameters through reinforcement learning, and improve the generalization ability for unknown faults and system robustness;

[0079] Step 34: After the safety mechanism is executed, the fault diagnosis results are fed back to the digital twin engine, and the interaction parameters between the twin model and the environment are dynamically updated to ensure the real-time adaptability of the subsequent control strategy.

[0080] In this embodiment, preferably, the method integrates a cloud platform to realize remote monitoring and historical data analysis, specifically including:

[0081] Step 41: Deploy a cloud server cluster and configure a distributed database (such as a NoSQL or time series database) to store the industrial robot's historical operation data, including position, speed, torque, temperature, and fault event records, supporting high-throughput data writing and long-term archiving.

[0082] Step 42: Develop a web-based remote monitoring interface, integrate a real-time data stream visualization module, synchronize the digital twin engine status via the WebSocket protocol, and implement dynamic graphical display of joint trajectories, end-point forces / torques, and energy consumption indicators. Also, embed an alarm notification system to automatically push notifications to mobile devices or PCs when an anomaly is detected.

[0083] Step 43: Build a historical data analysis module, apply time series analysis algorithms (such as ARIMA or LSTM) to mine operating data patterns, predict equipment remaining life, maintenance cycle, and performance degradation trends, and generate optimization recommendation reports.

[0084] Step 44: Feedback the analysis results to the local or cloud-based digital twin engine through the RESTful API interface to dynamically update the twin model parameters and control strategy to improve the accuracy and robustness of adaptive control;

[0085] Step 45: Implement end-to-end security mechanisms, including data encryption (such as AES-256), access control policies (based on the RBAC model), and audit logs to ensure the confidentiality and integrity of remote monitoring and data transmission.

[0086] The working principle and use process of the present invention:

[0087] First, during the system initialization phase, users deploy Industrial Internet of Things (IIoT) edge computing nodes and sensor networks, and use configuration tools to set digital twin engine parameters, including dynamic model initial values, adaptive control gain thresholds, and fault diagnosis rule sets. At the same time, historical operating data is loaded into the cloud or local database to train LSTM and CNN models to improve state estimation accuracy.

[0088] Secondly, during operation, the industrial robot starts task execution, collects multi-dimensional physical information (such as joint position, speed, torque and temperature) in real time, and performs millisecond-level synchronization and preprocessing through edge nodes; the preprocessed data stream is input into the digital twin engine in real time, which uses adaptive Kalman filtering to update the twin parameters, simulates the current state of the physical entity, and combines the deep reinforcement learning (DRL) algorithm for rolling optimization prediction to generate the optimal control instruction set.

[0089] Next, the optimized control instructions are sent to the robot's real-time controller via a low-latency deterministic network (such as EtherCAT), driving the actuator to achieve closed-loop tracking. At the same time, the system continuously monitors the operating status, applies fault diagnosis modules to detect abnormal operating conditions (such as sudden torque changes or temperature exceeding the limit), and triggers multi-level safety mechanisms (such as emergency shutdown or adaptive recovery strategies) to ensure operational safety.

[0090] Finally, during the maintenance and optimization phase, the cloud platform integrates historical data analysis capabilities, predicting equipment life and performance degradation through time series algorithms (such as ARIMA) and generating maintenance recommendations. The remote monitoring interface dynamically displays real-time indicators (such as trajectory error and energy consumption), and users can adjust control strategies or model parameters using web-based tools. The entire process supports online learning mechanisms, continuously iterating and optimizing the twin model and control algorithm to improve the system's robustness and adaptability in dynamic environments.

[0091] The typical usage process of this invention includes: task planning (setting target trajectory and constraints through the GUI), real-time execution (automatic data collection and adaptive control), and post-analysis (exporting operation reports and optimization suggestions). It is suitable for industrial scenarios such as automobile manufacturing and precision assembly, and can achieve high-precision, low-latency robot control.

[0092] See also Figure 6The present invention also provides a real-time adaptive control system for industrial robots based on digital twins, the system comprising:

[0093] The real-time data acquisition module integrates a multimodal sensor network deployed at each joint and end effector of the industrial robot. These sensors include encoders, accelerometers, torque sensors, six-dimensional force / torque sensors, current sensors, and temperature sensors. These sensors are used to accurately collect multidimensional physical information such as position, velocity, torque, force / torque, current, and temperature. This data is continuously captured at millisecond sampling rates (e.g., 1kHz) and transmitted to the central processing unit in real time via the Industrial Ethernet protocol, ensuring low-latency performance in the control loop. During the acquisition process, built-in preprocessing algorithms (such as Kalman filtering and range normalization) automatically eliminate noise interference, improve data reliability, and provide a precise input source for subsequent adaptive control.

[0094] The digital twin engine module, deployed on a cloud server cluster or local high-performance computing node, runs a high-fidelity dynamics simulation model and an environmental interaction model, combining adaptive Kalman filtering with a long short-term memory (LSTM) network for real-time state estimation, abnormal condition identification, and dynamic update of twin parameters. The state estimation process uses an adaptive Kalman filter algorithm to integrate multi-dimensional physical information (such as position, velocity, and torque) from the real-time data acquisition module to generate high-precision predictions of the robot's joint angles, end-effector pose, and environmental forces, with an error control range of ±0.1mm. Abnormal condition identification uses an LSTM network to analyze historical and real-time data streams, automatically detecting typical failure modes such as collisions, excessive vibration, and sudden load changes, triggering immediate alarms and recording event logs. The dynamic update of twin parameters is based on abnormality identification results and environmental feedback, optimizing simulation model parameters (such as dynamic coefficients and friction model constants) online to ensure consistency between the digital twin and the physical robot state. The update frequency can reach 100Hz, supporting real-time adjustment of adaptive control strategies. This module achieves processing latency as low as 5ms through a GPU-accelerated computing framework and is seamlessly integrated with the control execution module, providing highly robust and adaptable decision support for industrial scenarios.

[0095] The adaptive control module uses a deep reinforcement learning (DRL) algorithm to implement an online rolling optimization control strategy, generating optimal instructions for joint space trajectory correction, end-point accuracy compensation, and dynamic load adaptability. It also implements parameter adaptive tuning through fuzzy logic or neural networks. This module employs the Deep Deterministic Policy Gradient (DDPG) framework, combined with a real-time reward mechanism to evaluate action effectiveness, dynamically generating joint torque adjustment and end-point posture compensation instructions to effectively address load fluctuations or external interference. The fuzzy logic system fuzzifies input variables (such as position deviation and velocity error) based on error feedback and outputs optimal proportional-integral-derivative (PID) gain parameters through a preset rule base, achieving millisecond-level response. The neural network model utilizes historical control data for online training and adaptively adjusts the dynamic model coefficients to ensure the robustness of the control strategy in changing industrial environments. This module works in conjunction with the state estimation and anomaly recognition modules to form a closed-loop adaptive control loop, supporting a 100Hz update frequency and processing latency of less than 5ms, significantly improving trajectory tracking accuracy and the system's anti-interference capability.

[0096] The fault diagnosis module applies a fusion of LSTM and convolutional neural network (CNN) models to perform multi-scale time series analysis, detecting anomalies such as position offset, sudden torque changes, or temperature overruns, triggering emergency shutdowns, audible and visual alarms, and adaptive recovery strategies. This module supports real-time monitoring of joint encoder signals, motor current, and temperature sensor data streams, constructing multidimensional time series inputs and utilizing a sliding window mechanism for online prediction. When an anomaly is detected, the system dynamically triggers a graded response based on confidence thresholds: an emergency shutdown command is issued to the drive via the CAN bus, audible and visual alarms are activated to alert the operator via the HMI interface, and an adaptive recovery strategy invokes a pre-trained strategy network to generate joint position compensation or load redistribution strategies, ensuring steady-state operation is restored within 500ms. The module integrates an online learning mechanism, utilizing historical fault data to fine-tune model parameters, improving the false alarm rejection rate to 99.5%. The module shares a data bus with the state estimation module, forming a closed-loop diagnosis-control collaborative framework.

[0097] The remote monitoring platform leverages the high concurrent processing capabilities of the distributed database to support thousands of data read and write operations per second, ensuring efficient storage and query of large-scale time series data. The web-based visualization interface provides dynamic chart display, historical trend analysis, and real-time alarm log viewing functions. Operators can access it remotely through any standard browser to monitor the robot status in real time. The WebSocket protocol implements a full-duplex communication mechanism, reducing network latency to milliseconds and ensuring that abnormal events are pushed immediately. The RESTful API interface allows external systems (such as MES or ERP platforms) to obtain diagnostic reports, control recommendations, and adaptive strategy feedback through standardized HTTP request calls. At the same time, the platform integrates the OAuth2.0 authentication mechanism and performs AES-256 encryption on all data transmission to ensure the security and integrity of cloud communications.

[0098] The network transmission module uses industrial Ethernet or fieldbus (such as EtherCAT or PROFIBUS) to achieve low-latency deterministic data transmission, and ensures instruction integrity and sub-millisecond real-time performance through cyclic redundancy check (CRC) code, hash check and deterministic scheduling algorithm.

[0099] In this embodiment, preferably, the real-time data acquisition module includes a data quality assessment unit, which performs real-time credibility analysis on sensor data based on preset thresholds, eliminates outliers and triggers a redundant sensor switching mechanism; the digital twin engine module integrates a hybrid simulation architecture, and the physical dynamics model and the data-driven model run collaboratively through a weighted fusion algorithm, supporting online model parameter calibration and uncertainty quantification; the adaptive control module is embedded in a multi-objective optimization framework, combines energy consumption, accuracy and wear indicators, and generates a Pareto optimal solution set through deep reinforcement learning (DRL) rolling optimization.

[0100] The working principle and use process of the present invention:

[0101] When the system starts, the real-time data acquisition module first uses a multimodal sensor network to collect multi-dimensional physical information of each joint and end effector of the industrial robot with high precision, including position, speed, torque, force / torque, current, and temperature data. The data quality assessment unit analyzes the sensor credibility in real time, eliminates outliers, and activates the redundant switching mechanism to ensure the reliability of the input data. The collected data is transmitted to the digital twin engine module in a low-latency, high-integrity manner through the network transmission module. This module runs a high-fidelity dynamic simulation model and an environmental interaction model on the cloud or local node, using adaptive Kalman filtering and LSTM networks for real-time state estimation, abnormal working condition identification, and dynamic update of twin parameters. It supports online parameter calibration and uncertainty quantification through a hybrid simulation architecture that weightedly fuses physical dynamics models and data-driven models. The updated twin data is fed into the adaptive control module, which uses a deep reinforcement learning algorithm to optimize the control strategy online. This module generates optimal commands for joint trajectory correction, end-point precision compensation, and dynamic load adaptability. Furthermore, a multi-objective optimization framework combining energy consumption, precision, and wear metrics generates a Pareto-optimal solution set. Parameters are adaptively tuned using fuzzy logic or neural networks, and control commands are output to the industrial robot's actuators. The fault diagnosis module continuously monitors the data stream and applies a fusion of LSTM and CNN models for multi-scale time series analysis to detect abnormal events such as position offset, sudden torque changes, or temperature excursions. Upon identifying an anomaly, it immediately triggers an emergency shutdown, audible and visual alarms, and adaptive recovery strategies to ensure system safety. The remote monitoring platform displays the data stream and analysis results in real time through a web-based visualization interface. It utilizes a cloud-based distributed database to store historical information, synchronizes updates via the WebSocket protocol, and integrates a RESTful API to support external system calls and feedback. This entire workflow is executed cyclically, enabling real-time adaptive control of the industrial robot and improving operational precision and reliability.

[0102] Example 1 (precision assembly scenario):

[0103] During the precision assembly of automotive engine cylinder blocks, the system uses a digital twin engine to compare the contact forces predicted by the simulation model with the six-dimensional force sensor data from the end-user's physical robot in real time. When a hole-shaft mismatch deviation exceeding ±5μm is detected, an adaptive impedance control strategy based on reinforcement learning is implemented. The cloud-based training module uses historical assembly data to construct a deep deterministic policy gradient (DDPG) network, dynamically generating an optimal stiffness / damping parameter matrix. This matrix is ​​then transmitted to the on-site PLC via the OPC UA protocol within 50ms. This approach addresses the problem of jamming caused by workpiece tolerances in high-precision, compliant assembly. Testing with the ABB IRB 6700 has increased the assembly success rate from 82% to 98%.

[0104] In this embodiment, the system uses distributed edge computing nodes to process high-frequency sensor data, while offloading model update tasks to cloud servers, leveraging cloud elastic computing resources to accelerate large-scale parameter optimization.

[0105] Example 2 (arc welding scenario):

[0106] For multi-layer and multi-pass welding of thick ship plates, the system integrates data from infrared thermal imagers and laser displacement sensors to construct a three-dimensional temperature field-deformation field coupling model in the digital twin.

[0107] The long short-term memory network (LSTM) is used to predict the thermal deformation in the next 10 seconds, and the model predictive controller (MPC) is linked to achieve third-order compensation:

[0108] 1) Real-time offset compensation of welding gun trajectory (±0.2mm accuracy);

[0109] 2) Adaptive adjustment of welding current pulse frequency (±15A dynamic range);

[0110] 3) Cooling gas flow rate graded control.

[0111] This technology breaks through the industry's difficult problem of uncontrollable heat accumulation and deformation in traditional welding. When applied to the KUKA KR 500 welding machine, the weld porosity is reduced by 60%.

[0112] In this embodiment, the digital twin engine runs in real time based on a locally deployed GPU cluster, achieves millisecond-level simulation prediction through FPGA hardware acceleration, and supports multi-robot collaborative control scenarios.

[0113] Example 3 (cross-industry logistics sorting scenario):

[0114] In the dual-scenario verification of 3C electronic component sorting and pharmaceutical logistics, the system built a universal digital twin framework:

[0115] Physical layer: Deploy an inertial measurement unit (IMU) and an RGB-D camera to achieve 0.1° attitude perception;

[0116] Twin layer: uses lightweight rigid body dynamics model (Featherstone algorithm) for parallel calculation;

[0117] Control layer: Develop a distributed collaborative optimizer.

[0118] When handling an iPhone camera module (weighing less than 5g), microsecond-level servo gain fine-tuning is triggered based on vibration spectrum analysis; when sorting medical reagent bottles (height > 20cm), an anti-tip prediction algorithm is activated. This architecture, validated on the FANUC M-20iB / 35, uses a pluggable modular design to seamlessly switch between sorting fragile components and high-center-of-gravity objects within the same control system.

[0119] In this embodiment, the control instructions are transmitted in a 5G URLLC network, ensuring an end-to-end delay lower than 1 ms, and integrating blockchain technology to guarantee data integrity and security audit.

[0120] The foregoing description of the disclosed embodiments enables a person skilled in the art to carry out or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Therefore, the present application is not intended 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 real-time adaptive control method for industrial robots based on digital twins, characterized by: include: Collect the operating data of industrial robots in real time, obtain multi-dimensional physical information based on multimodal sensor fusion technology, and perform millisecond-level timestamp synchronization and data preprocessing through industrial Internet of Things edge computing nodes; Build a digital twin model to simulate the physical system and transmit the pre-processed real-time operation data stream to the cloud or a locally deployed digital twin engine. This engine has built-in high-fidelity dynamic simulation models and environmental interaction models, and uses adaptive Kalman filtering and long-short-term memory networks to estimate operating states and identify abnormal operating conditions. It dynamically updates the twin parameters to match the current state of the physical entity. Based on the real-time updated digital twin, parallel simulation prediction is carried out, and the online rolling optimization control strategy is optimized using the deep reinforcement learning algorithm to generate the optimal control instruction set including joint space trajectory correction, end-point accuracy compensation and dynamic load adaptability; The adaptive control algorithm is executed to output control instructions, and the optimized control instructions are sent to the robot's real-time controller through a low-latency deterministic network, driving the physical actuator to achieve closed-loop control. At the same time, the twin model and control strategy are continuously iterated and optimized based on the online learning mechanism.

2. The real-time adaptive control method for industrial robots based on digital twins according to claim 1 is characterized in that: The real-time collection of the industrial robot's operating data is based on a multi-sensor network to obtain position, speed and torque information, wherein the sensor network covers multiple types of encoders, accelerometers, and torque sensors to accurately obtain joint position, speed, torque, six-dimensional force / torque of the end effector, motor current and temperature multi-dimensional physical information.

3. The real-time adaptive control method for industrial robots based on digital twins according to claim 1 is characterized in that: The real-time collection of the operating data of the industrial robot includes the following steps: Step 11: Deploy a multi-sensor network at each key joint of the industrial robot, including encoders for real-time monitoring of joint angle position, accelerometers for high-precision measurement of linear acceleration to infer velocity, and torque sensors for collecting torque data. In addition, a six-dimensional force / torque sensor is integrated into the end effector to obtain three-dimensional force and torque information. In addition, current sensors are used to monitor motor current changes, and temperature sensors are used to track the thermal status of key components, ensuring comprehensive coverage of multi-dimensional physical information; Step 12: Use the IIoT edge computing nodes to synchronize millisecond-level timestamps for all sensor data, using the IEEE 1588 precision time protocol to align clocks within the network, eliminating time drift errors and ensuring the timing consistency of data collection. Step 13: Perform data preprocessing at the edge computing node, including applying a Kalman filter to smooth the position signal, using a median filter to remove accelerometer noise, normalizing the torque data, and implementing an outlier detection algorithm to remove invalid data points, thereby generating a high-quality, low-latency preprocessed dataset. Step 14: Based on multimodal sensor fusion technology, integrate multi-dimensional information such as position, speed, torque, force / torque, current and temperature, and generate a unified state vector through a weighted fusion algorithm to provide real-time input for the subsequent digital twin model.

4. The real-time adaptive control method for industrial robots based on digital twins according to claim 1, characterized in that: The digital twin model includes a physical dynamics model and a data-driven machine learning model. The physical dynamics model is constructed based on the Lagrange equation or Newton-Euler dynamics theory to accurately simulate joint torque, inertia effects and friction characteristics; the data-driven machine learning model integrates supervised learning and reinforcement learning frameworks, and is trained using historical operating data, including long-short-term memory networks for time series prediction and convolutional neural networks for processing visual sensor inputs, to enhance the model's generalization ability for unknown working conditions.

5. The real-time adaptive control method for industrial robots based on digital twins according to claim 1, characterized in that: The adaptive control algorithm in the execution of the adaptive control algorithm to output control instructions adopts fuzzy logic or neural network for parameter optimization, wherein the fuzzy logic controller constructs a fuzzy rule set based on an expert knowledge base, quantifies input variables through membership functions, and infers the output control gain adjustment amount in real time to cope with system nonlinearity and uncertainty; the neural network method adopts a deep feedforward network or a loop structure, uses historical operation data to train weight parameters online, optimizes the objective function through the back propagation algorithm, realizes adaptive parameter tuning, and ensures that the control instructions remain robust and stable in a dynamic environment.

6. The real-time adaptive control method for industrial robots based on digital twins according to claim 1, characterized in that: The step of executing the adaptive control algorithm to output the control instructions and transmitting the control instructions to the robot actuator via the industrial bus protocol comprises the following steps: Step 21: Format the optimized control instructions into a standard data frame that complies with the industrial bus protocol, including instruction type, joint position, speed, and torque parameters, and embed a millisecond-level timestamp to ensure timing consistency; Step 22: Add a cyclic redundancy check code or hash check value to the data frame to perform real-time error detection and data integrity verification. If a transmission error is detected, the automatic retransmission mechanism is triggered to resend the instruction through the low-latency network; Step 23: Broadcast the verified data frames to each actuator node of the robot via industrial Ethernet or fieldbus, using a deterministic scheduling algorithm to ensure transmission priority, avoid network congestion, and meet sub-millisecond real-time requirements; Step 24: After receiving the data frame, the actuator node parses the frame header and load content, extracts the control instruction and converts it into a motor drive signal, and feeds back a confirmation signal to the digital twin engine to update the closed-loop control state. Step 25: Dynamically adjust transmission parameters based on the feedback signal, such as bandwidth allocation or retransmission timeout threshold, and optimize bus protocol configuration in combination with online learning mechanism to improve system robustness and adaptability.

7. The real-time adaptive control method for industrial robots based on digital twins according to claim 1, characterized in that: The method also includes a fault diagnosis module to detect anomalies and trigger safety mechanisms, specifically including: Step 31: Based on the real-time state estimation data from the digital twin engine, a fusion model of long short-term memory network and convolutional neural network is applied to perform multi-scale time series analysis on the operating data to identify abnormal patterns such as position offset, torque mutation, or temperature overrun. The fault risk level is quantified by setting thresholds and probability distribution models. Step 32: When a high-risk anomaly is detected, a multi-level safety mechanism is automatically triggered, including: a first-level response, which sends an emergency shutdown command to the robot controller via a low-latency network, executing joint braking and power cutoff; a second-level response, which activates the audible and visual alarms and pushes fault codes to the monitoring center; and a third-level response, which combines digital twin simulation results to generate adaptive recovery strategies, such as thermal management adjustments or load redistribution, to minimize production interruptions. Step 33: Utilize the online learning mechanism to continuously collect fault event data, optimize the diagnostic model parameters through reinforcement learning, and improve the generalization ability for unknown faults and system robustness; Step 34: After the safety mechanism is executed, the fault diagnosis results are fed back to the digital twin engine, and the interaction parameters between the twin model and the environment are dynamically updated to ensure the real-time adaptability of the subsequent control strategy.

8. The real-time adaptive control method for industrial robots based on digital twins according to claim 1, characterized in that: The method integrates a cloud platform to achieve remote monitoring and historical data analysis, specifically including: Step 41: Deploy a cloud server cluster and configure a distributed database to store the industrial robot's historical operation data, including position, speed, torque, temperature, and fault event records, supporting high-throughput data writing and long-term archiving. Step 42: Develop a web-based remote monitoring interface, integrate a real-time data stream visualization module, synchronize the digital twin engine status via the WebSocket protocol, and implement dynamic graphical display of joint trajectories, end-point forces / torques, and energy consumption indicators. Also, embed an alarm notification system to automatically push notifications to mobile devices or PCs when an anomaly is detected. Step 43: Build a historical data analysis module, apply time series analysis algorithms to mine operational data patterns, predict equipment remaining life, maintenance cycles, and performance degradation trends, and generate optimization recommendation reports; Step 44: Feedback the analysis results to the local or cloud-based digital twin engine through the RESTful API interface to dynamically update the twin model parameters and control strategy to improve the accuracy and robustness of adaptive control; Step 45. Implement end-to-end security mechanisms, including data encryption, access control policies, and audit logs, to ensure the confidentiality and integrity of remote monitoring and data transmission.

9. Real-time adaptive control system of industrial robots based on digital twins, characterized by: The system comprises: The real-time data acquisition module integrates a multimodal sensor network and is deployed at each joint and end effector of the industrial robot. It includes encoders, accelerometers, torque sensors, six-dimensional force / torque sensors, current sensors, and temperature sensors. It is used to collect multi-dimensional physical information such as position, speed, torque, force / torque, current, and temperature with high precision. The digital twin engine module is deployed in a cloud server cluster or a local high-performance computing node. It runs a high-fidelity dynamics simulation model and an environmental interaction model, and combines adaptive Kalman filtering with a long short-term memory (LSTM) network for real-time state estimation, abnormal condition identification, and dynamic update of twin parameters. The adaptive control module uses the online rolling optimization control strategy of the deep reinforcement learning (DRL) algorithm to generate optimal instructions for joint space trajectory correction, end-point accuracy compensation, and dynamic load adaptability, and implements parameter adaptive tuning through fuzzy logic or neural networks; The fault diagnosis module uses a fusion model of LSTM and convolutional neural network (CNN) to perform multi-scale time series analysis to detect anomalies such as position offset, torque mutation, or temperature exceeding the limit, triggering emergency shutdown, audible and visual alarms, and adaptive recovery strategies; Remote monitoring platform, deploying a cloud-based distributed database (such as NoSQL or time series database), supporting a web-based visualization interface, synchronizing data streams via the WebSocket protocol, and integrating a RESTful API interface to feedback analysis results; The network transmission module uses industrial Ethernet or fieldbus (such as EtherCAT or PROFIBUS) to achieve low-latency deterministic data transmission, and ensures instruction integrity and sub-millisecond real-time performance through cyclic redundancy check (CRC) code, hash check and deterministic scheduling algorithm.

10. The real-time adaptive control system for industrial robots based on digital twins according to claim 9, characterized in that: The real-time data acquisition module includes a data quality assessment unit, which performs real-time credibility analysis on sensor data based on preset thresholds, eliminates outliers and triggers a redundant sensor switching mechanism; the digital twin engine module integrates a hybrid simulation architecture, and the physical dynamics model and the data-driven model operate collaboratively through a weighted fusion algorithm, supporting online model parameter calibration and uncertainty quantification; the adaptive control module is embedded in a multi-objective optimization framework, combining energy consumption, accuracy and wear indicators, and generating a Pareto optimal solution set through deep reinforcement learning (DRL) rolling optimization.

Images