A safety protection and control method and system for industrial robots

Abstract

This application provides a safety protection control method and system for industrial robots, relating to the field of industrial robot safety technology. By integrating multimodal sensor data and utilizing a pre-trained model, it accurately identifies intrusion intent and predicts motion trajectories, effectively overcoming the shortcomings of single sensors being susceptible to environmental interference leading to false alarms or missed alarms. Simultaneously, by combining real-time state parameters such as robot joint angles, speed, acceleration, and load mass, it dynamically generates instantaneous danger zones that expand and contract with changes in kinetic energy, solving the problem that traditional static fences cannot balance safety and production efficiency. More importantly, when a high-risk intrusion is detected, not only is the servo power cut off, but the mechanical locking mechanism at key joints is also immediately triggered, completely eliminating the displacement risk caused by the robot's inertial gliding due to power failure or gravity. Thus, a full-link safety protection system from intelligent perception and early warning to physical rigid locking is constructed, significantly improving the personnel safety protection capability in industrial sites.

Description

Technical Field

[0001] This application relates to the field of industrial robot safety technology, and in particular to a safety protection and control method and system for industrial robots. Background Technology

[0002] In current industrial production, industrial robot arms possess significant kinetic energy when operating at high speeds. If operators mistakenly enter the work area, serious safety accidents such as collisions or being run over can easily occur. Although traditional solutions employ physical safety fences, they have significant drawbacks: physical fences alone cannot detect unauthorized crossings; visual monitoring or manual emergency stops often result in delayed responses; and even after an accident, the robot may still cause injury due to inertial motion or gravitational displacement caused by power failure. Furthermore, reset mechanisms are complex. In addition, traditional static fences cannot dynamically adjust the protection range based on the robot's real-time speed and load, making it difficult to balance safety and efficiency. Single sensors are also susceptible to environmental interference, leading to false alarms or safety system failures due to malfunctions. Moreover, they lack multi-source data fusion verification and fault tolerance mechanisms. Summary of the Invention

[0003] This application provides a safety protection control method and system for industrial robots to solve one or more technical problems existing in the prior art, and at least provides a beneficial option or creates conditions that can detect personnel intrusion in real time through multi-source sensor data fusion and dynamic risk area modeling, and link the robot to perform graded deceleration or emergency braking, thereby effectively preventing secondary injuries caused by inertial motion or abnormal reset while eliminating the limitations of physical fences.

[0004] On the one hand, this application provides a safety protection and control method for industrial robots, including the following steps: Multi-dimensional sensing data is acquired in real time through a multi-modal sensor array, including at least two of the following: visual images, lidar point clouds, and infrared thermal imaging data. The multidimensional perception data is input into a pre-trained real-time intrusion intent recognition model, which outputs the authenticity probability value of the intrusion event and the motion trajectory prediction information of the intrusion object. The real-time motion state parameters of the industrial robot are obtained, and the motion state parameters include at least joint angles, end effector speed, acceleration and load mass; Based on the real-time motion state parameters, an adaptive dynamic safety boundary generation algorithm is used to calculate the instantaneous danger zone of the robot at the current moment. The range of the instantaneous danger zone dynamically expands and contracts with the change of the robot's kinetic energy. When the authenticity probability value exceeds a preset threshold, and it is determined that the predicted trajectory information of the intrusion object will enter the instantaneous danger zone, an emergency braking command is generated. In response to the emergency braking command, the power output of the robot's servo motor is cut off, and the mechanical locking actuators located at the robot's key joints are immediately triggered to lock the robot in its current position to eliminate the risk of gravitational displacement.

[0005] Furthermore, the construction and reasoning process of the real-time intrusion intent recognition model includes: The visual image is subjected to key point detection of human skeleton, the lidar point cloud is subjected to clustering and segmentation to obtain the obstacle volume, and the infrared thermal imaging data is subjected to temperature anomaly region extraction. The extracted feature vectors are spatiotemporally aligned and fused, then input into the intrusion intent recognition model, which outputs intrusion classification results and intrusion intent confidence; the intrusion intent recognition model uses a lightweight neural network. If a signal mutation caused by the fall of a non-living object or environmental interference is detected, but the confidence level of the intrusion intent is lower than a preset threshold, it is determined to be a false alarm and the robot continues to operate normally.

[0006] Furthermore, the method also includes a fault tolerance mechanism, including: The health status of each sensor in the multimodal sensor array is monitored in real time using an online anomaly detection algorithm; When missing or abnormal main sensor data is detected, the weight coefficients of the remaining available sensors are dynamically adjusted based on a Bayesian network, and the authenticity probability value is recalculated. If the remaining sensors are unable to support a safety assessment, the operating mode will automatically degrade, expanding the range of the instantaneous danger zone and reducing the robot's maximum permissible speed.

[0007] Furthermore, the adaptive dynamic security boundary generation algorithm specifically includes: Calculate the maximum braking distance and inertial swing range of the robot's end effector and each link at the current speed; The maximum braking distance and the range of inertial sway are extended outward by a safety margin to form the instantaneous danger zone; The safety margin is a dynamic variable. When the robot is in a high-speed, high-load state, the safety margin automatically increases; when the robot is in a low-speed or stationary state, the safety margin automatically decreases to the minimum physical contact threshold.

[0008] Furthermore, the scaling factor of the safety margin is obtained through optimization using a reinforcement learning strategy, specifically including: Construct a state space that includes at least the robot's real-time speed, payload mass, and relative distance to obstacles; Construct the action space, which is a scaling factor for the safety margin; Define a reward function consisting of a linearly weighted sum of a safety penalty term and an efficiency reward term; The safety penalty item is configured such that when a momentary danger zone collides with an intrusion object, it takes a value of negative infinity to forcibly avoid safety risks. The efficiency reward item is configured as follows: when no collision occurs and the instantaneous danger zone volume is smaller than a preset benchmark volume, the value is positive gain, and the smaller the instantaneous danger zone volume, the greater the positive gain; Training samples are constructed in a simulation environment using historical intrusion event data and robot obstacle avoidance cases. The optimal policy model is obtained by iteratively training the reinforcement learning agent until the policy network converges. In actual operation, the optimal strategy model is loaded, and the optimal safety margin scaling factor is output based on the current real-time status.

[0009] Furthermore, the mechanical locking actuator, which is triggered at a key joint of the robot, specifically includes: Within the first preset time window after receiving the emergency braking command, a free stop or rapid deceleration command is sent to the servo drive; After confirming that the motor torque has dropped to a safe threshold, or at the end of the first preset time window, immediately send an energizing signal to the mechanical locking actuator; The mechanical locking actuator adopts a power-off brake or hydraulic clamping structure, which generates mechanical friction torque within milliseconds after the action signal is triggered, forcibly locking the joint axis and preventing the robot arm from nodding or slipping under gravity due to power failure.

[0010] On the other hand, this application provides a safety protection control system for an industrial robot, used to implement the aforementioned safety protection control method for an industrial robot, including: a data acquisition module, an intent recognition module, a status monitoring module, a master control decision module, and an execution drive module; The output signal of the data acquisition module is connected to the input of the intent recognition module; The output terminals of the status monitoring module and the intent recognition module are both signal-connected to the input terminal of the main control decision module; The control output signal of the main control decision module is connected to the input of the execution drive module; in: The data acquisition module, including a multimodal sensor array, is configured to: acquire multidimensional sensing data in real time, wherein the multidimensional sensing data includes at least two of the following: visual images, lidar point clouds, and infrared thermal imaging data; The intent recognition module has a built-in pre-trained real-time intrusion intent recognition model and is configured to: input the multi-dimensional perception data into the real-time intrusion intent recognition model and output the authenticity probability value of the intrusion event and the motion trajectory prediction information of the intrusion object. The status monitoring module is configured to acquire real-time motion status parameters of the industrial robot, the motion status parameters including at least joint angles, end effector speed, acceleration and load mass; The main control decision module is configured to: calculate the instantaneous danger zone of the robot at the current moment in real time based on the real-time motion state parameters using an adaptive dynamic safety boundary generation algorithm; the range of the instantaneous danger zone dynamically expands and contracts with the change of the robot's kinetic energy; when the authenticity probability value exceeds a preset threshold and it is determined that the motion trajectory prediction information of the intrusion object will enter the instantaneous danger zone, generate an emergency braking command. The execution drive module, which is connected to the robot controller and the mechanical locking actuator, is configured to: in response to the emergency braking command, cut off the power output of the robot servo motor and immediately trigger the mechanical locking actuator set at the key joints of the robot to lock the robot in the current position to eliminate the risk of gravity displacement.

[0011] Furthermore, the multimodal sensor array is integrated into the physical isolation components and the periphery of the working area; The physical isolation component is a rigid railing or a flexible light curtain bracket with sensor mounting slots. The multimodal sensor array is embedded inside the physical isolation component, so that the sensor's field of view and the physical barrier form a complementary blind zone coverage; When the physical isolation component is deformed by an external impact, a secondary alarm is triggered by the integrated strain sensor, and the robot operation is forcibly suspended.

[0012] Furthermore, the mechanical locking actuator includes at least one of an electromagnetic brake, a hydraulic locking clamp, or a mechanical pin device; The execution drive module includes redundant control circuitry, which can directly trigger the mechanical locking actuator via a hardware watchdog signal when the main control decision module fails.

[0013] Furthermore, the system also includes a human-machine interaction terminal, which is signal-connected to the main control decision module. The human-machine interaction terminal is used to visualize the real-time three-dimensional morphology of the instantaneous danger zone, intrusion warning information and sensor health status, and allows authorized users to adjust safety margin parameters and intent recognition thresholds.

[0014] The beneficial effects of this application are as follows: This application provides a safety protection and control method for industrial robots. By integrating multimodal sensor data such as visual images, LiDAR point clouds, and infrared thermal imaging, it utilizes a pre-trained model to accurately identify intrusion intentions and predict motion trajectories, effectively overcoming the shortcomings of single sensors being susceptible to environmental interference leading to false alarms or missed alarms. Simultaneously, by combining real-time state parameters such as robot joint angles, speed, acceleration, and load mass, it dynamically generates instantaneous danger zones that expand and contract with changes in kinetic energy, solving the problem that traditional static fences cannot simultaneously address safety and production efficiency. More importantly, upon determining a high-risk intrusion, it not only cuts off servo power but also immediately triggers the mechanical locking mechanism at key joints, completely eliminating the displacement risk caused by robot inertial gliding due to power failure or gravity. This constructs a full-link safety protection system from intelligent perception and early warning to physical rigid locking, significantly improving personnel safety in industrial settings. This application also provides a corresponding system; the beneficial effects of the system are similar to the method and will not be elaborated upon here.

[0015] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0016] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0017] Figure 1 This is a flowchart of the safety protection and control method for industrial robots provided in this application; Figure 2 This is a structural diagram of the safety protection control system for the industrial robot provided in this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] The present application will be further described below with reference to the accompanying drawings and specific embodiments. The described embodiments should not be considered as limitations on the present application, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present application.

[0020] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0022] In the current field of industrial automation, industrial robots, especially multi-jointed robotic arms, are widely used in high-risk operations such as welding, handling, and assembly. Because industrial robots carry enormous kinetic energy during high-speed operation, if an operator accidentally enters their work area or an abnormal human-robot interaction occurs, it can easily lead to serious collisions, crushing, or even crushing accidents, causing personal injury and equipment damage. To address this safety challenge, existing technical solutions mainly rely on a combination of physical isolation and basic electrical control.

[0023] Current mainstream protection technologies primarily involve setting up fixed physical safety fences to rigidly separate the robot's work area from the personnel activity area. However, this static fence has significant drawbacks: it cannot detect whether personnel have illegally climbed over or entered the danger zone through gaps in the conveyor belt, and once personnel enter the fenced area, there is a lack of effective dynamic early warning mechanisms. Secondly, some solutions use detection devices such as light curtain sensors or safety mats, triggering an emergency stop for the robot when an object is detected obstructing the view or a change in pressure. The limitations of this type of technology are that the detection range is fixed and singular, making it susceptible to false alarms caused by dust, changes in light, or vibration, leading to frequent unplanned shutdowns of the production line and severely impacting production efficiency. Furthermore, single-modal sensors lack the ability to determine the intent of the intruder, failing to distinguish between accidental personnel entry and normal material transport, and lacking intelligent identification capabilities.

[0024] A more critical flaw lies in the robot's braking and stopping mechanism. Traditional safety controls, upon triggering an emergency stop signal, typically only cut off the power supply to the servo motors, relying on internal electromagnetic brakes or mechanical friction to decelerate and stop the robot. This approach presents two major risks: First, due to the significant inertia of the robot and its load, it will continue to slide a distance after power is cut off, and this uncontrollable displacement can often cause injury. Second, for vertically mounted or heavily loaded joints, simply cutting off power may lead to delayed or failed brake response, causing the robot arm to fall or unexpectedly displace under gravity—the so-called gravitational displacement risk—which poses a significant threat to personnel below.

[0025] Furthermore, existing safety protection zones are typically pre-defined, fixed spaces that cannot dynamically adjust the protection range based on the robot's real-time speed, acceleration, and load. When the robot operates at low speeds, an excessively large fixed protection zone restricts human access and collaboration efficiency; conversely, when the robot operates at high speeds and under heavy loads, the fixed protection range may be insufficient to cover the actual braking distance, resulting in a mismatch between safety protection and actual risk. In summary, existing technologies lack intelligent sensing capabilities based on multi-source data fusion, dynamic boundary adjustment mechanisms based on real-time kinetic energy, and rigid locking methods capable of completely eliminating the risks of inertial and gravitational displacement, making it difficult to achieve efficient human-robot collaboration while ensuring absolute safety.

[0026] To address the aforementioned issues, this application proposes a safety protection control method and system for industrial robots. It constructs a full-link safety protection system integrating intelligent perception, dynamic evaluation, and rigid locking. Firstly, by fusing a multi-modal sensor array of visual images, LiDAR point clouds, and infrared thermal imaging, combined with a pre-trained real-time intrusion intent recognition model, it accurately outputs the probability of the authenticity of intrusion events and predicts motion trajectories, effectively solving the problems of single sensors being susceptible to interference and unable to recognize intrusion intent. Secondly, it innovatively introduces an adaptive dynamic safety boundary generation algorithm, calculating the instantaneous danger zone that dynamically expands and contracts with kinetic energy based on the robot's real-time joint angles, speed, acceleration, and load mass. This achieves real-time matching between the safety protection range and the actual risk level, balancing operational efficiency and safety. Finally, at the control execution level, when a high-risk intrusion is detected, not only is the servo power cut off, but the mechanical locking actuators located at key joints are immediately triggered, forcibly locking the robot in its current position, completely eliminating the risk of secondary displacement injury caused by inertial gliding or gravity under traditional emergency stop methods.

[0027] First, the safety protection and control method for industrial robots provided in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0028] Reference Figure 1 The implementation process of the safety protection and control method for industrial robots provided in this application includes, but is not limited to, the following steps.

[0029] Step S110: Real-time acquisition of multi-dimensional sensing data through a multi-modal sensor array.

[0030] The multidimensional sensing data includes at least two of the following: visual images, lidar point clouds, and infrared thermal imaging data.

[0031] In step S110, a multimodal sensor array including a visual camera, LiDAR, and infrared thermal imager is deployed to collect multidimensional perception data covering visible light texture, three-dimensional spatial depth, and temperature characteristics in real time. The complementary characteristics of different sensors are used to overcome the limitations of single detection methods. For example, in complex industrial environments such as strong light, dust, or darkness, visual images provide detailed textures, LiDAR point clouds construct accurate spatial contours, and infrared thermal imaging can effectively identify heat sources of living organisms. This ensures comprehensive and blind-spot-free perception of the working environment and provides a highly reliable and interference-resistant data foundation for subsequent safety decisions.

[0032] Step S120: Input the multidimensional perception data into the pre-trained real-time intrusion intent recognition model, and output the authenticity probability value of the intrusion event and the motion trajectory prediction information of the intrusion object.

[0033] In step S120, the multi-dimensional sensing data collected in the preceding steps is input into a real-time intrusion intent recognition model that has been pre-trained with a large number of samples. This upgrades the model from simple target detection to intelligent behavior analysis. The model can comprehensively calculate the probability value of whether the intrusion event is a real person mistakenly entering or a material interference by integrating information from multiple sources, effectively filtering false alarms. At the same time, based on time-series data analysis, it can predict the future movement trajectory of the intruder and determine in advance whether the person has a tendency to enter the dangerous area. This forward-looking intent recognition mechanism allows the system to reserve valuable reaction time before an accident occurs, avoiding the lag in traditional technologies that only react at the moment of contact.

[0034] Step S130: Obtain the real-time motion state parameters of the industrial robot. The motion state parameters include at least the joint angle, end effector speed, acceleration, and load mass.

[0035] In step S130, the current motion state parameters of the industrial robot are acquired in real time, specifically including the angular position of each joint, the linear and angular velocities of the end effector, acceleration and deceleration, and the current gripping load mass. This allows the safety control system to accurately grasp the robot's real-time kinetic energy reserves and inertial characteristics. Because the degree of danger of the robot is not constant, but directly related to its movement speed and load weight, only by quantifying these dynamic parameters can we accurately assess the sliding distance and impact energy that the robot may generate in an emergency, providing the necessary physical basis for formulating personalized safety strategies.

[0036] Step S140: Based on real-time motion state parameters, an adaptive dynamic safety boundary generation algorithm is used to calculate the instantaneous danger zone of the robot at the current moment. The range of the instantaneous danger zone dynamically expands and contracts with the change of the robot's kinetic energy.

[0037] In step S140, based on the acquired real-time motion state parameters, the adaptive dynamic safety boundary generation algorithm is used to calculate the instantaneous danger zone of the robot at the current moment. This breaks the rigid mode of traditional fixed safety fences or static protection zones. The algorithm can dynamically expand and contract the range of the danger zone according to the real-time kinetic energy of the robot. When the robot is running at high speed and under heavy load, it automatically expands the protection radius to cover a longer braking distance, while at low speed and light load or when stationary, it shrinks the protection range to allow personnel to cooperate and maintain at closer distances. This dynamic adjustment mechanism maximizes the utilization rate of production space and the flexibility of human-machine collaboration while ensuring absolute safety.

[0038] Step S150: When the authenticity probability value exceeds the preset threshold and it is determined that the motion trajectory prediction information of the intrusion object will enter the instantaneous danger zone, an emergency braking command is generated.

[0039] In step S150, when the probability value of the authenticity of the intrusion event exceeds the preset safety threshold, and the algorithm determines that the predicted movement trajectory of the intrusion object will overlap with the real-time calculated instantaneous danger area, an emergency braking command is immediately generated to realize a dual confirmation mechanism for risk assessment. This requires both a high degree of confidence in the identification result to eliminate false alarms and a substantial collision risk on the spatial trajectory, thereby ensuring that the highest level of intervention measures are triggered only in truly critical moments. This logical judgment effectively balances the sensitivity of security protection with the stability of production continuity, preventing a decline in production efficiency due to frequent false stops.

[0040] In step S160, in response to the emergency braking command, the power output of the robot's servo motor is cut off, and the mechanical locking actuator set at the key joints of the robot is immediately triggered to lock the robot in its current position to eliminate the risk of gravity displacement.

[0041] In step S160, after receiving the emergency braking command, a two-level physical blocking action is executed. First, the power output of the servo motor is cut off to stop the active drive. Then, the mechanical locking actuator preset at the key joints of the robot is immediately triggered to forcibly lock the robot arm in the current position. This solves the fatal flaw of the traditional emergency stop method, that is, after simply cutting off the power, the robot will continue to slide due to huge inertia or fall due to gravity. Through the rigid physical constraint provided by the mechanical locking mechanism, all residual kinetic energy and gravitational displacement risks can be eliminated instantly, ensuring that the robot can remain absolutely still in any posture, thereby maximizing the safety of the intruder.

[0042] In some embodiments of this application, step S120, the construction and reasoning process of the real-time intrusion intent recognition model, includes the following steps.

[0043] Step S210: Detect key points of the human skeleton in the visual image, perform clustering and segmentation on the lidar point cloud to obtain the obstacle volume, and extract temperature anomaly regions from the infrared thermal imaging data.

[0044] In step S210, key points of the human skeleton are detected for visual images to accurately locate human posture and joint positions. Clustering and segmentation algorithms are used to process the LiDAR point cloud to accurately calculate the three-dimensional volume and spatial contour of obstacles. Temperature anomaly regions are extracted from infrared thermal imaging data to identify heat source targets with life characteristics. The most discriminative high-level semantic features are extracted from the original messy sensor data. The complementary advantages of different modalities are used to construct a comprehensive feature description containing morphological, spatial and thermal information, providing structured and interference-resistant input data for subsequent deep analysis.

[0045] Step S220: The extracted feature vectors are spatiotemporally aligned and fused, and then input into the intrusion intent recognition model to output the intrusion classification result and the intrusion intent confidence score.

[0046] The intrusion intent recognition model employs a lightweight neural network. In step S220, the personnel skeleton features, obstacle volume information, and temperature anomaly features extracted in the previous steps are subjected to high-precision spatiotemporal alignment and multi-source fusion to eliminate the deviations of different sensors in timestamps and spatial coordinate systems. Then, the fused feature vector is input into a lightweight neural network optimized for edge computing. Through deep learning models, the deep correlation between multimodal features is mined, and clear intrusion classification results and quantified intrusion intent confidence are quickly output. This not only ensures millisecond-level real-time inference on resource-constrained industrial controllers, but also quantifies the reliability of the judgment through confidence index, providing a scientific decision-making basis for the system to distinguish between real threats and environmental noise.

[0047] Step S230: If a signal mutation caused by a non-living object falling or environmental interference is detected, but the confidence level of the intrusion intent is lower than a preset threshold, it is determined to be a false alarm and the robot continues to operate normally.

[0048] In step S230, an intelligent filtering and fault-tolerance mechanism is constructed, specifically designed to handle scenarios such as accidental drops of non-living objects, drastic changes in light, or electromagnetic interference that cause sudden changes in sensor signals. When the system detects such abnormal signals but the calculated confidence level of the intrusion intent is lower than the preset safety threshold, the logic determines it as a false alarm event and instructs the robot to maintain normal operation. This effectively solves the pain point of traditional safety systems frequently triggering unnecessary emergency stops due to environmental interference, significantly reduces the unplanned downtime rate of the production line, and ensures that industrial robots can maintain high sensitivity to real personnel intrusion in complex and ever-changing production environments while also possessing intelligent immunity to false alarms, thereby significantly improving the continuity of overall operations and production efficiency.

[0049] In some embodiments of this application, in order to solve the problems of time asynchrony and spatial parallax of multi-source heterogeneous data, the system adopts hardware-triggered synchronization combined with software interpolation algorithm.

[0050] In the spatial coordinate registration stage, feature points in the visual camera coordinate system are... Transform to LiDAR coordinate system The process involves using a pre-calibrated extrinsic rotation matrix. Translation vector The implementation is shown in equation (1) below: (1); For time synchronization, low-frequency visual features should be included. Mapped to the time axis of the high-frequency lidar, assuming the lidar sampling time is... And adjacent visual images and The timestamps are respectively and (satisfy Then, the linear interpolation algorithm is used to calculate... Visual feature state corresponding to time As shown in equation (2): (2); This process eliminates motion blur and installation errors, constructs a unified 3D fusion feature vector, and provides high-confidence input for intent recognition.

[0051] In some embodiments of this application, the intrusion intent recognition model employs a lightweight neural network; specifically, it uses a depthwise separable convolutional network to reduce computational requirements, and its convolutional operations are decomposed into channel-wise convolution and pointwise convolution. ;in, For the input feature map, For channel-wise convolution kernels, For channel-wise convolution output, For pointwise convolution kernels, This is the final output feature map.

[0052] To adapt to edge deployment, model weights With activation value Eight-bit integer quantization is performed, and the mapping is shown in equation (3) below: (3); In equation (3), round represents the rounding function. This is the scaling factor, used to control the quantization granularity; This is the zero-point offset, used to align zero values; This is the quantized integer value (8-bit). The floating-point value to be quantized (weight or activation value). This is the approximate floating-point value recovered after dequantization. The training dataset contains samples of normal operation, personnel intrusion, and industrial interference (sparks, dust) in a 6:3:1 ratio. The system supports incremental learning; when changing scenarios, only the parameters of the terminal fully connected layer need to be fine-tuned to adapt to the new scenario, without the need for full retraining.

[0053] In some embodiments of this application, the method further includes a fault-tolerant mechanism, comprising the following steps.

[0054] Step S310: Use an online anomaly detection algorithm to monitor the health status of each sensor in the multimodal sensor array in real time.

[0055] In step S310, by deploying an online anomaly detection algorithm, the operational health status of each independent sensor in the multimodal sensor array is continuously monitored and diagnosed in real time. This enables the immediate detection of potential problems such as hardware failures, data drift, signal interruptions, or noise surges, preventing the entire safety system from making incorrect judgments due to the failure of a single sensor. This proactive health monitoring mechanism ensures the reliability and integrity of the perception layer data, providing a solid foundation for subsequent safety decisions and preventing the system from operating with defects without the system's knowledge, thus avoiding potential safety hazards.

[0056] Step S320: When missing or abnormal main sensor data is detected, the weight coefficients of the remaining available sensors are dynamically adjusted based on the Bayesian network, and the authenticity probability value is recalculated.

[0057] In step S320, when a data loss or serious anomaly is detected in the main sensor, a dynamic reconstruction mechanism based on a Bayesian network is immediately activated. The weight coefficients of the remaining available sensors are automatically re-evaluated and adjusted. The system uses probabilistic reasoning to fuse the remaining multi-source information to recalculate the probability value of the intrusion event. This gives the system strong robustness and self-healing ability, ensuring that the security protection function will not be completely paralyzed due to the damage of a single key component. The system can intelligently use redundant information to continue to maintain high-precision risk identification, ensuring that accurate security judgments can still be made under extreme conditions of partial hardware failure, thus maximizing the continuity of the production process.

[0058] In step S330, if the remaining sensors cannot support the safety determination, the operating mode is automatically downgraded, the range of the instantaneous danger zone is expanded and the maximum permissible speed of the robot is reduced.

[0059] In step S330, when the remaining number or quality of sensors after weight adjustment is insufficient to support a high-confidence safety determination, the system automatically triggers a degraded operation mode, actively expanding the coverage of the instantaneous danger zone and forcibly reducing the robot's maximum permissible speed. Adopting a conservative strategy in uncertain environments with limited perception capabilities, sacrificing some operational efficiency for absolute safety, the expanded danger zone provides a larger buffer for potential undetected risks, while the reduced speed directly decreases the robot's kinetic energy and braking distance. This ensures that even in the worst-case scenario of large-scale sensor failure, potential accidental injuries can be minimized, reflecting the human-centered safety design philosophy of this application.

[0060] In some embodiments of this application, the system constructs a Bayesian network to evaluate sensor health, with parent nodes representing the state of each sensor and child nodes representing the probability of intrusion authenticity. (Target event). Given observational evidence (Data from multiple sensors) The true posterior probability of intrusion is calculated as shown in equation (4): (4); In equation (4), To observe evidence under real-world intrusion conditions The likelihood; This represents the true prior probability of intrusion. For the first Possible hypothetical states (such as intrusion detection under different sensor combinations).

[0061] Based on this probability, the first... One sensor Fusion weights The calculation is shown in equation (5): (5); In equation (5), For given evidence Below, sensor The probability of being in a healthy state. This represents the total number of sensors.

[0062] When a sensor malfunctions (e.g., vision is obstructed), When a sensor fails, its weight is automatically reset to zero, and the weights of the remaining normal sensors are automatically normalized and recombined. This logic ensures that even under extreme conditions where some sensors fail, the system can still use redundant information to output highly reliable safety decisions, achieving inherently safe fault-tolerant control.

[0063] In some embodiments of this application, step S140 of the adaptive dynamic security boundary generation algorithm specifically includes the following steps.

[0064] Step S410: Calculate the maximum braking distance and inertial swing range of the robot end effector and each link at the current speed.

[0065] In step S410, by real-time calculation of the robot's kinematic model and dynamic parameters, the maximum braking distance required for the robot's end effector and each link to come to a complete stop from the current state at the current instantaneous speed is accurately calculated. At the same time, the range of inertial sway that may be caused by the elastic deformation of the mechanical structure and control delay is predicted. The abstract motion state parameters are transformed into specific physical space risk quantification indicators, accurately defining the maximum spatial envelope that the robot may actually occupy during emergency braking. This provides an indispensable physical basis for the subsequent construction of accurate safety protection boundaries, ensuring that the system can fully predict and cover potential displacement risks caused by inertia and mechanical characteristics.

[0066] Step S420: Extend the maximum braking distance and inertial swing range outward by a safety margin to form an instantaneous danger zone.

[0067] Among them, the safety margin is a dynamic variable. When the robot is in a high-speed, high-load state, the safety margin automatically increases; when the robot is in a low-speed or stationary state, the safety margin automatically decreases to the minimum physical contact threshold.

[0068] In step S420, based on the previously calculated maximum braking distance and inertial swing range, a dynamically changing safety margin is extended outward to ultimately form an instantaneous danger zone. This safety margin is not a fixed constant but a variable that is intelligently adjusted according to the working conditions. When the robot is in a high-speed running or high-energy state carrying a heavy load, the system will automatically increase the safety margin to cope with greater uncertainty errors and more complex braking environments. When the robot is in a low-speed crawling or completely stationary state, the safety margin will automatically shrink to a very small range that only includes the minimum physical contact threshold. This mechanism realizes the adaptive optimization of the safety protection strategy, which provides sufficient buffer space to prevent accidental collisions in high-risk working conditions, and maximizes the compression of the restricted area in low-risk working conditions, allowing operators to cooperate and maintain more closely, perfectly balancing the extreme safety requirements and high-efficiency production requirements in industrial scenarios.

[0069] In some embodiments of this application, the scaling factor of the aforementioned safety margin is obtained through optimization using a reinforcement learning strategy, and the specific process includes the following steps.

[0070] Step S510: Construct a state space that includes at least the robot's real-time speed, load mass, and relative distance to obstacles.

[0071] In step S510, by constructing a state space that includes multi-dimensional variables such as the robot's real-time speed, current load mass, and relative distance to surrounding obstacles, a comprehensive and real-time environmental perception input is provided for the reinforcement learning agent. This enables the agent to accurately quantify the current motion risk level and operational scenario characteristics, ensuring that the subsequently generated safety strategy can dynamically adapt to various complex working conditions, from high-speed heavy load to low-speed light load, and avoid decision-making bias due to missing state information.

[0072] Step S520: Construct the action space, which is a scaling factor for the safety margin.

[0073] In step S520, the scaling factor of the safety margin is defined as the action space. This means that the system no longer relies on human experience to set a fixed safety boundary expansion ratio, but instead gives the algorithm the ability to autonomously adjust the length of the safety buffer. This transforms the adjustment of the safety boundary into an optimizable continuous decision-making process, enabling the system to flexibly expand or shrink the protection range based on real-time risk assessment results. This achieves a fundamental shift from rigid static protection to intelligent dynamic adaptive protection.

[0074] Step S530: Define a reward function, which consists of a linearly weighted sum of a safety penalty term and an efficiency reward term.

[0075] The safety penalty item is configured as follows: when the instantaneous danger zone collides with the intrusion object, the value is negative infinity to force the avoidance of safety risks; the efficiency reward item is configured as follows: when no collision occurs and the volume of the instantaneous danger zone is less than the preset benchmark volume, the value is positive gain, and the smaller the volume of the instantaneous danger zone, the greater the positive gain.

[0076] In step S530, a composite reward function consisting of a linearly weighted safety penalty term and an efficiency reward term is designed. The safety penalty term is given an extreme value of negative infinity when a collision occurs, which forces the policy network to completely avoid any safety accidents and establishes the absolute principle of safety first. The efficiency reward term, under the premise of no collision, gives a positive gain according to the degree to which the instantaneous danger zone volume is smaller than the preset benchmark volume, and the smaller the zone, the greater the gain. This mechanism guides the active compression of unnecessary protection space on the basis of ensuring zero accidents, thereby achieving the optimal balance between safety and production efficiency at the mathematical level.

[0077] Step S540: Using historical intrusion event data and robot obstacle avoidance cases, training samples are constructed in the simulation environment. The reinforcement learning agent is iteratively trained until the policy network converges to obtain the optimal policy model.

[0078] In step S540, training samples are constructed in a highly realistic simulation environment using historical intrusion event data and robot obstacle avoidance cases. Through reinforcement learning agents, millions of iterations of trial and error and policy updates are performed until the policy network converges. This allows the system to experience various extreme dangerous scenarios and edge cases in the virtual environment in advance. It can learn the optimal defense strategy without bearing the safety risks on a real production line. The final optimal policy model has a generalization ability and decision-making intelligence that surpasses the setting of manual rules and can handle complex dynamic interaction situations that have never been seen before.

[0079] In step S550, during actual operation, the optimal strategy model is loaded, and the optimal safety margin scaling factor is output according to the current real-time state, so as to minimize the volume of the instantaneous danger zone while ensuring safety, thereby maximizing the robot's effective workspace.

[0080] In step S550, a fully trained optimal strategy model is deployed and loaded in the actual industrial site, enabling it to output the optimal safety margin scaling factor in milliseconds based on the robot's current real-time state. The ultimate goal of this step is to transform theoretical intelligent algorithms into actual productivity, dynamically compressing the volume of instantaneous dangerous areas to the minimum limit while ensuring absolute safety, thereby maximizing the release of the robot's effective workspace and allowing personnel and machines to collaborate efficiently in a more compact space, significantly improving the space utilization and overall operational efficiency of the production line.

[0081] In some embodiments of this application, the optimization of the safety margin scaling factor employs a deep learning algorithm based on proximal policy optimization (PPO). The neural network architecture comprises two fully connected hidden layers (128 neurons per layer, ReLU activation), and the output layer is mapped to the scaling factor range via a Tanh function. To eliminate dimensional differences, the state-space variable (velocity) is... ,load ,distance Normalization is performed before input: ,in These are the original sensor values. This represents the theoretical extreme value. Reward function. Designed as a collision penalty mechanism, when the distance... Below the threshold Apply a maximum negative value (set to) ), as shown in equation (6): (6); In equation (6), The reward value under safe conditions is not a fixed constant, but rather uses a dynamic calculation mechanism based on relative distance to guide the robot to actively increase the safe distance from the obstacle. Its calculation formula is defined as follows (7): (7); In equation (7), This is the reward gain coefficient (preferred to be 0.5 to 1.0), used to adjust the overall strength of the safety reward; The distance is monitored in real time, that is, the Euclidean distance between the robot's end effector or body and the nearest obstacle at the current moment; This represents the maximum effective detection range of the sensor, indicating the farthest range at which the system can reliably perceive obstacles. It is used to normalize the distance difference.

[0082] Therefore, Equation (7) ensures that the robot receives a lower reward when it is in a safe area but close to the boundary, while the reward value increases linearly with the distance, thereby driving the robot to move to a more open safe area during policy gradient updates.

[0083] The convergence criterion for the model is the average rate of change of return over N = 500 consecutive rounds. ,when Parameters are frozen at any time. Furthermore, domain randomization is introduced to randomly perturb the friction coefficient during simulation. and noise level, The standard value for the friction coefficient is used to ensure that the model has the ability to generalize and transfer from simulation to real environment.

[0084] In some embodiments of this application, step S160 involves triggering a mechanical locking actuator located at a key joint of the robot, specifically including the following steps.

[0085] Step S610: Within the first preset time window after receiving the emergency braking command, send a free stop or rapid deceleration command to the servo driver.

[0086] In step S610, within the first preset time window after receiving the emergency braking command, a free stop or rapid deceleration command is quickly sent to the servo driver. The electromagnetic braking characteristics of the motor itself are used to start consuming the robot's kinetic energy and suppressing inertial impulse within milliseconds. The joint speed and output torque are reduced in advance through electrical control means, creating a low-impact and low-load safety condition for the subsequent intervention of the mechanical locking mechanism, avoiding damage to transmission components or severe vibration that may be caused by direct mechanical hard locking under high speed and high torque conditions.

[0087] Step S620: After confirming that the motor torque has dropped to a safe threshold, or at the end of the first preset time window, immediately send an energizing signal to the mechanical locking actuator.

[0088] In step S620, the system monitors the motor torque status in real time. Once it is confirmed that the torque has dropped to the preset safety threshold or the first preset time window has ended, it immediately sends an energizing signal to the mechanical locking actuator. This ensures the optimal switching time from dynamic deceleration to static locking, preventing excessive mechanical stress caused by forced locking when the motor still has a large residual torque. It also ensures that the physical backup mechanism can be activated in time if the electric braking efficiency is insufficient or exceeds the time limit. This achieves a seamless and smooth transition between the two braking methods, ensuring the continuity and reliability of the braking process.

[0089] In step S630, the mechanical locking actuator adopts a power-off brake or hydraulic clamping structure, which generates mechanical friction torque within milliseconds after the action signal is triggered, forcibly locking the joint axis and preventing the robot arm from nodding or slipping under gravity due to power failure.

[0090] In step S630, by employing a mechanical locking actuator with a power-off brake or hydraulic clamping structure, a huge mechanical friction torque is generated within milliseconds after receiving the action signal, forcibly locking the robot joint axis in the current position. This provides a purely physical safety barrier independent of the electronic control system, completely eliminating the risk of uncontrollable nodding, slipping, or swinging of the robot arm under gravity due to sudden power outages, controller failures, or servo malfunctions. This ensures that even in extreme catastrophic scenarios where the power supply is completely lost, the robot can still maintain a stable posture, thereby effectively protecting the safety of the workers below and the integrity of surrounding precision equipment and products.

[0091] In some embodiments of this application, the mechanical locking actuator is coaxially mounted at the rear end of the joint motor, and its triggering timing is calculated in real time by the dynamic equations. First preset time window. Defined from the current angular velocity The theoretical time required to brake to zero is shown in equation (8) below: (8); In equation (8), For rotational inertia, which varies with joint angle change; and These are the frictional torque and the gravitational torque, respectively. This is the maximum available braking torque.

[0092] Furthermore, the safe torque threshold The dynamic setting is a safety multiple of the gravity load component, used to determine whether mechanical locking is triggered, and satisfies the following formula (9): (9); In equation (9), For safety factors, the value range is [1.1, 1.2]. For load quality; It is the acceleration due to gravity; The distance from the load's center of mass to the axis of rotation varies with the joint angle. The mechanism ensures seamless switching between pneumatic braking and mechanical locking, preventing both high-speed impacts and gravity slippage.

[0093] Secondly, refer to Figure 2 This application provides a safety protection control system for an industrial robot, used to implement the aforementioned safety protection control method for an industrial robot, including: a data acquisition module, an intent recognition module, a status monitoring module, a master control decision module, and an execution drive module.

[0094] Specifically, the output signal of the data acquisition module is connected to the input signal of the intent recognition module; the output signals of both the status monitoring module and the intent recognition module are connected to the input signal of the main control decision module; and the control output signal of the main control decision module is connected to the input signal of the execution drive module.

[0095] The data acquisition module includes a multimodal sensor array configured to acquire multidimensional sensing data in real time, including at least two of the following: visual images, lidar point clouds, and infrared thermal imaging data.

[0096] Specifically, the data acquisition module integrates a multimodal sensor array that includes visual images, lidar point clouds, and infrared thermal imaging data. It can collect multidimensional environmental information covering the robot's workspace in real time, around the clock. This breaks through the perception limitations of a single sensor and leverages the complementary advantages of different sensing technologies to overcome challenges in complex working conditions such as changes in lighting, dust interference, or difficulties in detecting transparent objects. It provides the upper-level decision-making system with highly redundant, high-precision, and comprehensive raw environmental data, ensuring that the system can keenly capture any subtle potential risk signals. This is a prerequisite for building a highly reliable human-machine collaborative safety defense line.

[0097] The intent recognition module has a built-in pre-trained real-time intrusion intent recognition model and is configured to: input multi-dimensional perception data into the real-time intrusion intent recognition model and output the authenticity probability value of the intrusion event and the motion trajectory prediction information of the intrusion object.

[0098] Specifically, the intent recognition module is responsible for deeply fusing and extracting features from the multi-source heterogeneous sensing data transmitted by the data acquisition module, and then outputting a quantified probability value of the authenticity of the intrusion event and a prediction of the future movement trajectory of the intruder. This transforms the original physical signals into risk intelligence with semantic information, enabling the system to distinguish between unintentional approach and malicious intrusion, and to predict the movement trend of people or objects in advance. Thus, the risk level assessment is completed before a collision occurs, providing a scientific quantitative basis for subsequent graded response strategies, greatly reducing the false alarm rate and improving the foresight of the early warning.

[0099] The status monitoring module is configured to acquire real-time motion status parameters of the industrial robot, including at least joint angles, end effector speed, acceleration, and load mass.

[0100] Specifically, the state monitoring module acquires key motion state parameters in real time, including joint angles, end effector speed, acceleration, and load mass. This allows the safety control system to constantly monitor the robot's current kinetic energy level and inertial characteristics. These data are not only the basic inputs for calculating the robot's braking distance and swing range, but also the core indicators for assessing the degree of danger of the current working condition. This ensures that the system can dynamically adjust safety strategies based on whether the robot is in a high-speed, heavy-load or low-speed, light-load state, achieving comprehensive coverage from simple environmental perception to the interaction between the robot's own state and the environment.

[0101] The main control decision module is configured to: calculate the instantaneous danger zone of the robot at the current moment in real time based on real-time motion state parameters and using an adaptive dynamic safety boundary generation algorithm; the range of the instantaneous danger zone dynamically expands and contracts with the change of the robot's kinetic energy; when the authenticity probability value exceeds a preset threshold and it is determined that the motion trajectory prediction information of the intrusion object will enter the instantaneous danger zone, an emergency braking command is generated.

[0102] Specifically, based on the real-time motion parameters provided by the state monitoring module, the main control decision module uses an adaptive dynamic safety boundary generation algorithm to calculate the robot's instantaneous danger zone at the current moment in milliseconds. The range of this zone is not fixed but intelligently expands and contracts with changes in the robot's kinetic energy. Its core value lies in breaking the rigid mode of traditional fixed safety fences. When the robot is moving at high speed, it automatically expands the protection range to reserve sufficient buffer, and automatically contracts when moving at low speed or stationary to release more collaborative space. Thus, it maximizes production efficiency and the flexibility of human-machine collaboration while ensuring absolute safety, and realizes refined and intelligent control of safety boundaries.

[0103] Next, the main control decision module comprehensively receives risk probability and trajectory prediction information from the intent recognition module, as well as the calculated instantaneous danger zone data. Through logical analysis, when the probability of intrusion authenticity exceeds a preset threshold and the predicted trajectory shows that the intruder is about to enter the danger zone, it decisively generates an emergency braking command. Its significance lies in transforming scattered perception information and status data into clear action commands, realizing a closed-loop decision-making process from risk identification to emergency response. This ensures that the system can react as quickly as possible when facing real threats, avoiding security incidents caused by decision delays or logical confusion. It is the final arbiter for ensuring safe human-machine interaction.

[0104] The execution drive module, which is connected to the robot controller and the mechanical locking actuator respectively, is configured to: in response to an emergency braking command, cut off the power output of the robot's servo motors and immediately trigger the mechanical locking actuators set at the key joints of the robot to lock the robot in its current position to eliminate the risk of gravity displacement.

[0105] Specifically, the execution drive module is directly connected to the robot controller and the mechanical locking actuator at key joints. Once it receives an emergency braking command from the main control decision module, it immediately cuts off the power output of the servo motor and simultaneously triggers the dual action of the mechanical locking mechanism. Its fundamental function is to not only stop the robot's active movement instantly through the dual insurance mechanism of electrical power failure and mechanical brake, but also completely eliminate the physical risk of the robot arm nodding, slipping, or swinging due to inertia under the action of gravity caused by power failure or malfunction. This ensures that the robot can be firmly locked in its current position, building the last indestructible physical safety barrier for on-site personnel and truly achieving intrinsic safety.

[0106] In some embodiments of this application, the multimodal sensor array in the data acquisition module is integrated into the physical isolation components and the periphery of the working area.

[0107] The physical isolation components are rigid railings or flexible light curtain supports with sensor mounting slots, creating clear physical boundaries to prevent unintentional entry into high-risk areas. They provide a stable and accurate integrated carrier for multimodal sensor arrays. This integrated design upgrades traditional passive isolation facilities into active sensing nodes, retaining the intuitive warning and physical barrier functions of mechanical barriers while solving the problems of sensor external damage and viewing angle deviation through structural design. This ensures the long-term stability and detection consistency of sensing equipment in complex industrial environments.

[0108] The multimodal sensor array is embedded inside the physical isolation component, enabling the sensor's detection field of view to form a complementary blind zone coverage with the physical barrier. This eliminates the detection blind spots caused by bracket obstruction or installation angle limitations in traditional separate deployments. By utilizing the geometry of the rigid railing or light curtain bracket itself, the sensor beams or field of view are guided to interweave, constructing a comprehensive three-dimensional perception network. This allows any intrusion attempt to cross, climb, or bypass the physical barrier to be captured instantly, greatly improving the system's ability to identify covert intrusions and ensuring all-round, blind-spot-free monitoring of the protected area.

[0109] When the physical isolation components are deformed by external impact, the integrated strain sensor triggers a secondary alarm and forces the robot to stop operating. This expands the triggering conditions for safety protection from simple personnel intrusion to damage to the integrity of the infrastructure itself, preventing secondary safety accidents caused by the collapse, displacement or failure of the physical barrier, and ensuring that the system can enter the highest alert state and stop dangerous actions the moment the protective boundary itself is damaged.

[0110] In some embodiments of this application, the mechanical locking actuator includes at least one of an electromagnetic brake, a hydraulic locking clamp, or a mechanical pin device, providing a purely mechanical locking capability independent of the software control logic. When the system determines that an emergency stop is required, the mechanism can generate a huge frictional resistance torque or mechanical interference force within milliseconds by using electromagnetic attraction to release spring pressure, hydraulic high-pressure clamping, or hard insertion of a metal pin, thereby forcibly fixing the robot's key joint axis in its current position. This completely eliminates the risk of the robotic arm accidentally slipping, nodding, or swinging due to servo power failure, controller crash, or gravity imbalance, ensuring that the robot body maintains absolute static stability under extreme fault conditions, thus building an insurmountable physical safety barrier for on-site personnel and equipment.

[0111] The execution drive module includes redundant control circuitry. In the event of a failure in the main control decision module, it can directly trigger the mechanical locking actuator via a hardware watchdog signal, establishing an independent triggering channel independent of the main control decision module's software logic. If the main control module fails due to program crashes, communication interruptions, or processor malfunctions, the built-in hardware watchdog signal can bypass upper-level software checks and send action commands to the mechanically locked actuator via hardwired connection. This design effectively mitigates the risk of system-wide paralysis due to single-point failures, ensuring that even in catastrophic scenarios involving complete control system collapse, the safety protection mechanism can reliably activate and execute locking actions. This truly achieves comprehensive coverage from software fault tolerance to hardware failure safety, significantly enhancing the robustness and reliability of the entire industrial robot safety protection system.

[0112] In some embodiments of this application, the system further includes a human-machine interaction terminal, which is signal-connected to the main control decision module and is used to visualize the real-time three-dimensional shape of the instantaneous danger area, intrusion warning information and sensor health status, and allows authorized users to adjust safety margin parameters and intent recognition thresholds.

[0113] Specifically, the human-machine interface terminal, through real-time signal connection with the main control decision module, transforms the originally abstract algorithm data into intuitive and easy-to-understand three-dimensional dynamic images. This allows on-site personnel to clearly see the real-time changes in the three-dimensional shape of the instantaneous danger zone as the robot moves, specific intrusion warning details, and the health status of each sensor. This transparent information display not only greatly enhances the operator's perception of the current safety situation but also allows authorized users to flexibly adjust safety margin parameters and intent recognition thresholds according to actual production needs. This optimizes the production cycle while ensuring absolute safety, realizing a shift from passively receiving alarms to actively managing safety strategies and enhancing the system's adaptability and user-friendly interactive experience under different working conditions.

[0114] In summary, the safety protection and control method and system for industrial robots provided in this application have the following technical effects.

[0115] This solution achieves complementary blind spot coverage between the detection field of view and the physical barrier through the deep integration of a multimodal sensor array and physical isolation components, completely eliminating detection blind spots in traditional protection. Furthermore, strain sensors endow the isolation facility with self-sensing capabilities, instantly triggering a secondary alarm and forcibly shutting down the system when structural damage occurs, effectively preventing secondary safety incidents. The system's built-in real-time intrusion intent recognition model, combined with an adaptive dynamic safety boundary generation algorithm, can intelligently expand and contract the instantaneous danger zone based on changes in robot kinetic energy, accurately distinguishing between unintentional approach and genuine intrusion and predicting movement trajectories. This maximizes human-machine collaboration space and production efficiency while ensuring absolute safety.

[0116] The main control decision module generates emergency braking commands by integrating multi-dimensional data. This, combined with an execution drive module equipped with redundant control circuitry, creates a dual hardware and software safety mechanism. Even in the event of main control failure, a hardware watchdog can directly trigger a mechanical lock on the actuator, securing the robot firmly using electromagnetic, hydraulic, or mechanical pin methods to completely eliminate the risk of gravity displacement. Furthermore, the human-machine interface terminal provides 3D visualization and adjustable parameters, further enhancing the transparency of the safety situation and the flexibility of system management. The overall solution significantly improves the inherent safety level, system robustness, and intelligent human-machine collaboration of industrial robots under complex working conditions.

[0117] It should be noted that in all specific embodiments of this application, all data processing activities related to user identity or personal characteristics, such as user information, user behavior data, historical data, and location information, will be conducted in accordance with the principles of legality, legitimacy, and necessity. All data collection, use, storage, and processing will be subject to compliance with applicable national and regional laws, regulations, and industry standards, and informed consent from users will be obtained in a clear and explicit manner before processing. For the processing of sensitive personal information, separate consent from users will be obtained through prominent means such as pop-up prompts and independent confirmation pages. If any processing conflicts with laws and regulations, the laws and regulations will prevail, and necessary data processing will only be carried out within the scope permitted by laws and regulations, ensuring that all data-based applications, analyses, and technical implementations are conducted within the scope permitted by laws and regulations.

[0118] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0119] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of ordinary skill of an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary skill. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0120] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several programs to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0121] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequential list of executable programs for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, a program execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can retrieve and execute a program from or in conjunction with such a program execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit a program for use by or in conjunction with a program execution system, apparatus, or device.

[0122] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Additionally, computer-readable media can even be paper or other suitable media on which programs can be printed, for example, by optically scanning the paper or other media, then editing, interpreting, or, if necessary, processing it in a suitable manner to obtain the program electronically, and then storing it in computer memory.

[0123] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable program execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0124] In the foregoing description of this specification, the reference to terms such as "one embodiment / implementation," "another embodiment / implementation," or "certain embodiments / implementations," etc., indicates that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in an embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0125] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0126] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A safety protection and control method for an industrial robot, characterized in that, Includes the following steps: Multi-dimensional sensing data is acquired in real time through a multi-modal sensor array, including at least two of the following: visual images, lidar point clouds, and infrared thermal imaging data. The multidimensional perception data is input into a pre-trained real-time intrusion intent recognition model, which outputs the authenticity probability value of the intrusion event and the motion trajectory prediction information of the intrusion object. The real-time motion state parameters of the industrial robot are obtained, including at least joint angles, end effector speed, acceleration, and load mass. Based on the real-time motion state parameters, an adaptive dynamic safety boundary generation algorithm is used to calculate the instantaneous danger zone of the robot at the current moment. The range of the instantaneous danger zone dynamically expands and contracts with the change of the robot's kinetic energy. When the authenticity probability value exceeds a preset threshold, and it is determined that the predicted trajectory information of the intrusion object will enter the instantaneous danger zone, an emergency braking command is generated. In response to the emergency braking command, the power output of the robot's servo motor is cut off, and the mechanical locking actuators located at the robot's key joints are immediately triggered to lock the robot in its current position to eliminate the risk of gravitational displacement.

2. The safety protection and control method for industrial robots according to claim 1, characterized in that, The construction and reasoning process of the real-time intrusion intent recognition model includes: The visual image is subjected to key point detection of human skeleton, the lidar point cloud is subjected to clustering and segmentation to obtain the obstacle volume, and the infrared thermal imaging data is subjected to temperature anomaly region extraction. The extracted feature vectors are spatiotemporally aligned and fused, then input into the intrusion intent recognition model, which outputs intrusion classification results and intrusion intent confidence; the intrusion intent recognition model uses a lightweight neural network. If a signal mutation caused by the fall of a non-living object or environmental interference is detected, but the confidence level of the intrusion intent is lower than a preset threshold, it is determined to be a false alarm and the robot continues to operate normally.

3. The safety protection and control method for industrial robots according to claim 1, characterized in that, The method also includes a fault tolerance mechanism, including: The health status of each sensor in the multimodal sensor array is monitored in real time using an online anomaly detection algorithm; When missing or abnormal main sensor data is detected, the weight coefficients of the remaining available sensors are dynamically adjusted based on a Bayesian network, and the authenticity probability value is recalculated. If the remaining sensors are unable to support a safety assessment, the operating mode will automatically degrade, expanding the range of the instantaneous danger zone and reducing the robot's maximum permissible speed.

4. The safety protection and control method for industrial robots according to claim 1, characterized in that, The adaptive dynamic security boundary generation algorithm specifically includes: Calculate the maximum braking distance and inertial swing range of the robot's end effector and each link at the current speed; The maximum braking distance and the range of inertial sway are extended outward by a safety margin to form the instantaneous danger zone; The safety margin is a dynamic variable. When the robot is in a high-speed, high-load state, the safety margin automatically increases; when the robot is in a low-speed or stationary state, the safety margin automatically decreases to the minimum physical contact threshold.

5. The safety protection and control method for industrial robots according to claim 4, characterized in that, The scaling factor of the safety margin is obtained through optimization using a reinforcement learning strategy, the specific process of which includes: Construct a state space that includes at least the robot's real-time speed, payload mass, and relative distance to obstacles; Construct the action space, which is a scaling factor for the safety margin; Define a reward function consisting of a linearly weighted sum of a safety penalty term and an efficiency reward term; The safety penalty item is configured such that when a momentary danger zone collides with an intrusion object, it takes a value of negative infinity to forcibly avoid safety risks. The efficiency reward item is configured as follows: when no collision occurs and the instantaneous danger zone volume is smaller than a preset benchmark volume, the value is positive gain, and the smaller the instantaneous danger zone volume, the greater the positive gain; Training samples are constructed in a simulation environment using historical intrusion event data and robot obstacle avoidance cases. The optimal policy model is obtained by iteratively training the reinforcement learning agent until the policy network converges. In actual operation, the optimal strategy model is loaded, and the optimal safety margin scaling factor is output based on the current real-time status.

6. The safety protection and control method for industrial robots according to claim 1, characterized in that, The mechanical locking actuator, which is triggered at a key joint of the robot, specifically includes: Within the first preset time window after receiving the emergency braking command, a free stop or rapid deceleration command is sent to the servo drive; After confirming that the motor torque has dropped to a safe threshold, or at the end of the first preset time window, immediately send an energizing signal to the mechanical locking actuator; The mechanical locking actuator adopts a power-off brake or hydraulic clamping structure, which generates mechanical friction torque within milliseconds after the action signal is triggered, forcibly locking the joint axis and preventing the robot arm from nodding or slipping under gravity due to power failure.

7. A safety protection control system for an industrial robot, used to implement the safety protection control method for an industrial robot as described in any one of claims 1 to 6, characterized in that, include: The module includes a data acquisition module, an intent recognition module, a status monitoring module, a master control decision-making module, and an execution drive module. The output signal of the data acquisition module is connected to the input of the intent recognition module; The output terminals of the status monitoring module and the intent recognition module are both signal-connected to the input terminal of the main control decision module; The control output signal of the main control decision module is connected to the input of the execution drive module; in: The data acquisition module, including a multimodal sensor array, is configured to: acquire multidimensional sensing data in real time, wherein the multidimensional sensing data includes at least two of the following: visual images, lidar point clouds, and infrared thermal imaging data; The intent recognition module has a built-in pre-trained real-time intrusion intent recognition model and is configured to: input the multi-dimensional perception data into the real-time intrusion intent recognition model and output the authenticity probability value of the intrusion event and the motion trajectory prediction information of the intrusion object. The status monitoring module is configured to acquire real-time motion status parameters of the industrial robot, the motion status parameters including at least joint angles, end effector speed, acceleration and load mass; The main control decision module is configured to: calculate the instantaneous danger zone of the robot at the current moment in real time based on the real-time motion state parameters using an adaptive dynamic safety boundary generation algorithm; the range of the instantaneous danger zone dynamically expands and contracts with the change of the robot's kinetic energy; when the authenticity probability value exceeds a preset threshold and it is determined that the motion trajectory prediction information of the intrusion object will enter the instantaneous danger zone, generate an emergency braking command. The execution drive module, which is connected to the robot controller and the mechanical locking actuator, is configured to: in response to the emergency braking command, cut off the power output of the robot servo motor and immediately trigger the mechanical locking actuator set at the key joints of the robot to lock the robot in the current position to eliminate the risk of gravity displacement.

8. The safety protection control system for industrial robots according to claim 7, characterized in that, The multimodal sensor array is integrated into the physical isolation components and the periphery of the working area; The physical isolation component is a rigid railing or a flexible light curtain bracket with sensor mounting slots. The multimodal sensor array is embedded inside the physical isolation component, so that the sensor's field of view and the physical barrier form a complementary blind zone coverage; When the physical isolation component is deformed by an external impact, a secondary alarm is triggered by the integrated strain sensor, and the robot operation is forcibly suspended.

9. The safety protection control system for industrial robots according to claim 7, characterized in that, The mechanical locking actuator includes at least one of an electromagnetic brake, a hydraulic locking clamp, or a mechanical pin device. The execution drive module includes redundant control circuitry, which can directly trigger the mechanical locking actuator via a hardware watchdog signal when the main control decision module fails.

10. The safety protection control system for industrial robots according to claim 7, characterized in that, The system also includes a human-machine interface terminal, which is signal-connected to the main control decision module. The human-machine interface terminal is used to visualize the real-time three-dimensional shape of the instantaneous danger zone, intrusion warning information and sensor health status, and allows authorized users to adjust safety margin parameters and intent recognition thresholds.

Images