Intelligent industrial robot sorting system based on large model algorithm

Abstract

The invention belongs to the technical field of information, and particularly relates to an intelligent industrial robot sorting system based on a large model algorithm, and a high-precision material recognition algorithm based on multi-source information fusion can comprehensively utilize information of various sensors to realize high-precision recognition of various materials. According to the dynamic sorting path planning algorithm based on reinforcement learning, the industrial robot can dynamically plan the optimal sorting path according to the real-time working environment and task requirements, collision and detour are avoided, the sorting efficiency and the energy utilization rate are improved, the intelligent early warning and fault diagnosis functions are achieved, smooth proceeding of the sorting process is guaranteed, and the working efficiency is improved. The production management level is improved.

Description

Technical Field

[0001] The present invention belongs to the field of information technology, and in particular relates to an intelligent industrial robot sorting system based on a large model algorithm. Background Art

[0002] Material sorting, a critical component of industrial production, plays a vital role in numerous industries, including logistics and warehousing, electronics manufacturing, and food processing. With the continued expansion of industrial production and increasing demands for efficiency, traditional manual material sorting, due to its low efficiency and high error rates, is becoming increasingly difficult to adapt to modern production demands. Industrial robots, with their high speed, high precision, and strong repeatability, have become a revolutionary force in the material sorting field. However, their application still faces numerous technical bottlenecks, limiting their in-depth application in complex scenarios.

[0003] From the perspective of material identification, industrial scenarios present a rich variety of material types, with significant differences in shape, color, and material quality. Even subtle features of similar materials can be difficult to discern. Currently, most industrial robotic material sorting systems rely on a single visual recognition technology, which is significantly affected by environmental factors such as changes in light intensity, surface reflectivity, and occlusion, leading to incomplete feature extraction. In actual sorting operations, misclassification of similar materials and missed inspection of unusually shaped parts are common, reducing sorting efficiency and potentially causing production line downtime or product quality defects.

[0004] In terms of path planning, complex sorting environments often contain multiple target workstations, dynamic obstacles, and spatial constraints, placing higher demands on the real-time and safety of robot motion planning. Existing path planning algorithms are mostly based on static geometric modeling and preset rules, lacking comprehensive consideration of the robot's dynamic characteristics (such as joint torque limits and motion acceleration), material distribution density, and dynamic environmental changes. This makes the robot prone to problems such as circuitous paths and increased collision risks during the sorting process, resulting in a decrease in sorting output per unit time and increased energy loss.

[0005] In multi-robot collaborative scenarios, large-scale sorting tasks typically require multiple robots to work together, but existing systems lack dynamic task allocation and conflict resolution mechanisms. Communication delays and untimely global information sharing between robots can lead to overlapping work areas and repeated grabbing of target workpieces. Furthermore, each robot lacks a collaborative control strategy based on real-time working conditions, making it difficult to dynamically adjust its division of labor based on changes in order demand, equipment failures, and other emergencies. This results in an unsmooth overall operational process and difficulty improving system throughput. These technical pain points urgently need to be addressed through interdisciplinary technology integration and algorithm optimization to promote the development of industrial robot material sorting technology towards intelligence and flexibility. Summary of the Invention

[0006] The present invention provides an intelligent industrial robot sorting system based on a large model algorithm, comprising:

[0007] A high-precision material recognition algorithm module based on multi-source information fusion uses multiple sensors to collect and fuse multi-source material information in real time, and combines deep learning to achieve high-precision recognition;

[0008] A dynamic sorting path planning algorithm module based on reinforcement learning models the sorting process as a reinforcement learning environment and optimizes the path through a state-action-reward mechanism;

[0009] Based on the distributed collaborative optimization multi-robot task allocation and coordination algorithm module, a distributed objective function is constructed to achieve multi-robot task collaboration;

[0010] A robot motion stability improvement system module based on adaptive control, which monitors and adjusts motion parameters in real time through sensors;

[0011] The intelligent monitoring and management platform module for material sorting enables full-process data collection, analysis and early warning.

[0012] Furthermore, in the high-precision material recognition algorithm module based on multi-source information fusion, the multiple sensors include visual sensors (accuracy ±0.1mm), lidars (ranging accuracy ±2mm), infrared sensors (temperature resolution ±0.5°C) and tactile sensors (pressure accuracy ±0.1N), and the deployment positions are determined through finite element simulation and actual testing.

[0013] Furthermore, multi-source information fusion adopts a fusion network based on the attention mechanism, and the fusion formula is:

[0014] Among them, xi is the data of each sensor, αi is the attention weight, which is dynamically calculated by a three-layer fully connected network with 128, 64, and 32 neurons respectively, and optimized by the back propagation algorithm.

[0015] Furthermore, in the dynamic sorting path planning algorithm module based on reinforcement learning, the state space includes the material position P, the obstacle position O, and the robot posture Q. The action space is the displacement Δx, rotation Δθ, and grasping action g in the Cartesian coordinate system. The reward function is defined as: Where d is the path length, c is the number of collisions, and λ is the collision penalty coefficient with a value between 0.5 and 1.5.

[0016] Furthermore, the proximal policy optimization (PPO) algorithm is used for training, with an experience replay buffer capacity of 100,000 entries, a learning rate of 1e-4, 10 epochs per iteration, and the target network updated every 2000 steps.

[0017] Furthermore, in the multi-robot task allocation and coordination algorithm module based on distributed collaborative optimization, task allocation adopts the Hungarian algorithm, collaborative control adopts the distributed consistency protocol, communication delay is ≤10ms, and task completion time variance is ≤5%.

[0018] Furthermore, in the adaptive control-based robot motion stability improvement system module, a model predictive control (MPC) algorithm is used, with a prediction time domain of 5 steps and a control time domain of 3 steps. The state vector includes the joint angle θ, angular velocity θ˙, and load m. The optimization goal is to minimize the tracking error:

[0019]

[0020] Furthermore, in the intelligent monitoring and management platform module, the data acquisition frequency is 100 Hz, and the long short-term memory network (LSTM) is used for fault prediction, with an input sequence length of 100 and 256 hidden layer neurons, and a prediction accuracy of ≥95%.

[0021] Furthermore, in the multi-robot collaborative sorting scenario, the task priority is determined by fuzzy logic, with the input being the material value V, urgency U and volume S, the output being the priority coefficient P, and the membership function adopting a triangular distribution.

[0022] Beneficial effects:

[0023] Improve material identification accuracy: The high-precision material identification algorithm based on multi-source information fusion can comprehensively utilize information from multiple sensors to achieve high-precision identification of various materials, improve the accuracy and stability of material identification, reduce misidentification, and improve sorting efficiency and product quality.

[0024] Optimized sorting path planning: The dynamic sorting path planning algorithm based on reinforcement learning enables industrial robots to dynamically plan the optimal sorting path according to the real-time working environment and task requirements, avoiding collisions and detours, and improving sorting efficiency and energy utilization.

[0025] Improve the collaborative sorting capabilities of multiple robots: The multi-robot task allocation and coordination algorithm based on distributed collaborative optimization can achieve efficient collaborative work among multiple robots, avoid work conflicts and duplication of work, improve overall sorting efficiency, and meet the needs of large-scale material sorting.

[0026] Enhance robot motion stability: The robot motion stability improvement system based on adaptive control can monitor the robot's motion state and force conditions in real time, automatically adjust the motion control parameters, improve the stability and reliability of the robot's motion, and reduce material dropping and sorting errors caused by unstable robot motion.

[0027] Convenient sorting process monitoring and management: The intelligent monitoring and management platform of the material sorting process provides managers with real-time sorting process monitoring and decision support, and has intelligent early warning and fault diagnosis functions to ensure the smooth progress of the sorting process and improve the level of production management. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Module flow diagram. DETAILED DESCRIPTION

[0029] Example 1:

[0030] The present invention provides an intelligent industrial robot sorting system based on a large model algorithm, comprising:

[0031] A high-precision material recognition algorithm module based on multi-source information fusion uses multiple sensors to collect and fuse multi-source material information in real time, and combines deep learning to achieve high-precision recognition;

[0032] A dynamic sorting path planning algorithm module based on reinforcement learning models the sorting process as a reinforcement learning environment and optimizes the path through a state-action-reward mechanism;

[0033] Based on the distributed collaborative optimization multi-robot task allocation and coordination algorithm module, a distributed objective function is constructed to achieve multi-robot task collaboration;

[0034] A robot motion stability improvement system module based on adaptive control, which monitors and adjusts motion parameters in real time through sensors;

[0035] The intelligent monitoring and management platform module for material sorting realizes data collection, analysis and early warning for the entire process. In the high-precision material identification algorithm module based on multi-source information fusion, multiple sensors include visual sensors (accuracy ±0.1mm), laser radars (ranging accuracy ±2mm), infrared sensors (temperature resolution ±0.5℃) and tactile sensors (pressure accuracy ±0.1N). The deployment positions are determined by finite element simulation and actual testing. Multi-source information fusion adopts a fusion network based on the attention mechanism, and the fusion formula is: Where xi represents the sensor data, and αi represents the attention weight, which is dynamically calculated by a three-layer fully connected network with 128, 64, and 32 neurons, respectively, and optimized by the backpropagation algorithm. In the dynamic sorting path planning algorithm module based on reinforcement learning, the state space includes the material position P, the obstacle position O, and the robot posture Q. The action space is the displacement Δx, rotation Δθ, and grasping action g in the Cartesian coordinate system. The reward function is defined as: Where d is the path length, c is the number of collisions, and λ is the collision penalty coefficient, which is 0.5-1.5. The proximal policy optimization (PPO) algorithm is used for training, the experience replay buffer capacity is 100,000, the learning rate is 1e-4, 10 epochs are updated each iteration, and the target network is updated every 2000 steps. In the multi-robot task allocation and coordination algorithm module based on distributed collaborative optimization, the Hungarian algorithm is used for task allocation, the distributed consistency protocol is used for collaborative control, the communication delay is ≤10ms, and the task completion time variance is ≤5%. In the robot motion stability improvement system module based on adaptive control, the model predictive control (MPC) algorithm is used, with a prediction time domain of 5 steps and a control time domain of 3 steps. The state vector includes the joint angle θ, the angular velocity θ˙ and the load m. The optimization goal is to minimize the tracking error:

[0036] In the intelligent monitoring and management platform module, the data acquisition frequency is 100Hz, and a long short-term memory network (LSTM) is used for fault prediction. The input sequence length is 100, the hidden layer neurons are 256, and the prediction accuracy is ≥95%. In the multi-robot collaborative sorting scenario, the task priority is determined by fuzzy logic. The input is the material value V, urgency U and volume S, and the output is the priority coefficient P. The membership function adopts a triangular distribution.

[0037] Perception layer: Industrial robots are equipped with various sensors, including visual sensors, lidar, infrared sensors, tactile sensors, accelerometers, gyroscopes, and force sensors, to obtain real-time information from multiple sources, including material information, the robot's motion status, and the forces acting on it. These sensors collect data in real time and transmit it to the data processing layer.

[0038] Data processing layer: This includes a multi-source information fusion module, a feature extraction module, a reinforcement learning model training module, and an adaptive control parameter adjustment module. The multi-source information fusion module fuses data collected by multiple sensors to eliminate conflicts and redundancies. The feature extraction module uses deep learning and data mining techniques to extract key information from the fused data that reflects material characteristics and the robot's motion state. The reinforcement learning model training module uses reinforcement learning algorithms to model and train the material sorting process and establish an optimal sorting path planning strategy. The adaptive control parameter adjustment module uses adaptive control algorithms to adjust the robot's motion control parameters based on the data collected by the sensors.

[0039] Decision-making layer: includes a high-precision material identification module, a dynamic sorting path planning module, a multi-robot task allocation and coordination module, a motion stability control module, and an intelligent monitoring and management module. The high-precision material identification module uses a high-precision material identification algorithm to achieve high-precision material identification based on the results of the multi-source information fusion module and the feature extraction module. The dynamic sorting path planning module dynamically plans the optimal sorting path based on the results of the reinforcement learning model training module, combined with real-time material distribution and environmental information. The multi-robot task allocation and coordination module uses a multi-robot task allocation and coordination algorithm to allocate and coordinate tasks based on global task requirements and the status of each robot. The motion stability control module controls the robot's motion stability based on the results of the adaptive control parameter adjustment module. The intelligent monitoring and management module is responsible for collecting, storing, analyzing, and displaying data from the sorting process, enabling intelligent monitoring and management of the sorting process.

[0040] Execution layer: Industrial robots perform tasks such as material identification, sorting path planning, and material grabbing and placement according to control instructions generated by the decision layer. At the same time, feedback from the execution process is transmitted back to the data processing layer and decision layer to adjust and optimize the system.

[0041] Application layer: It is mainly an intelligent monitoring and management platform for the material sorting process, providing managers with a friendly interactive interface to realize real-time monitoring of the sorting process, data analysis, decision support and intelligent early warning functions.

[0042] Implementation of high-precision material recognition algorithm based on multi-source information fusion

[0043] Sensor Deployment and Data Collection: Based on the needs and characteristics of material sorting, multi-source sensors such as visual sensors, lidar, infrared sensors, and tactile sensors are deployed to ensure comprehensive and accurate acquisition of multi-source information about materials. Sensors collect data in real time at a set frequency and transmit it via the network to the multi-source information fusion module in the data processing layer.

[0044] Multi-source information fusion algorithm implementation: The multi-source information fusion module uses a deep learning-based multi-source information fusion algorithm, such as an attention-based fusion network, to fuse different types of sensor data. This algorithm automatically learns the relationships between different information sources, highlighting key information and improving the accuracy and reliability of information fusion.

[0045] Material Feature Extraction and Identification: The feature extraction module uses deep learning algorithms, such as convolutional neural networks (CNNs), to analyze and process the fused data to extract comprehensive material features. The high-precision material identification module leverages these comprehensive features, combined with pre-trained recognition models, to achieve high-precision identification of various materials.

[0046] Implementation of dynamic sorting path planning algorithm based on reinforcement learning

[0047] Definition of state space, action space, and reward function: The state space includes information such as the location of the material, the location of the obstacle, the location and posture of the robot; the action space includes various movements of the robot, such as moving, rotating, grasping, placing, etc.; the reward function is designed based on factors such as the robot's sorting efficiency, energy consumption, and whether there is a collision. When the robot can complete the sorting task efficiently and safely, a higher reward is given, otherwise a lower reward is given.

[0048] Reinforcement Learning Model Training: The reinforcement learning model training module uses reinforcement learning algorithms, such as the Deep Q-Network (DQN) or the Proximal Policy Optimization (PPO) algorithm, to model and train the material sorting process. During training, the robot continuously interacts with the environment, selecting actions based on its current state. After executing these actions, it observes the reward signals and new state reflected by the environment. The optimization algorithm then updates the policy network to maximize the cumulative reward. After repeated training, the robot learns the optimal sorting path planning strategy.

[0049] Dynamic sorting path planning and execution: During the actual sorting process, the dynamic sorting path planning module dynamically plans the optimal sorting path based on the real-time material distribution, obstacle location, and robot motion status, combined with the trained reinforcement learning model. The path information is then sent to the industrial robot at the execution layer to guide the robot to complete the sorting task.

[0050] Implementation of multi-robot task allocation and coordination algorithm based on distributed collaborative optimization

[0051] Task Analysis and Modeling: Based on the material sorting task requirements and production plan, we conduct detailed task analysis and modeling to determine task priority, time requirements, resource requirements, etc. At the same time, we evaluate and model each industrial robot's working capabilities, working scope, and current task status.

[0052] Distributed collaborative optimization algorithm implementation: Using a distributed optimization algorithm, such as the distributed particle swarm optimization algorithm (DPSO) or distributed genetic algorithm (DGA), we construct distributed objective functions and constraints. Objective functions can include maximizing sorting efficiency, minimizing sorting time, and ensuring material sorting accuracy. Constraints include robot capacity limitations, task sequence constraints, and resource constraints. This algorithm collaboratively optimizes the task allocation, work paths, and operation sequence of multiple robots to achieve the optimal task allocation solution and collaborative work strategy.

[0053] Multi-robot Collaborative Work: The Multi-robot Task Allocation and Coordination Module transmits optimized task allocation plans and collaborative work strategies to each industrial robot. During actual work, each robot exchanges information via a real-time communication network, dynamically adjusting its work strategy based on global task requirements and its own status. In the event of anomalies such as task changes or equipment failures, the Multi-robot Task Allocation and Coordination Module promptly reassigns and re-coordinates tasks, ensuring the efficiency and stability of multi-robot collaborative work.

[0054] Implementation of robot motion stability improvement system based on adaptive control

[0055] Sensor Deployment and Data Collection: Accelerometers, gyroscopes, force sensors, and other sensors are installed in key areas of industrial robots, such as joints and end effectors, to monitor the robot's motion and force in real time. These sensors collect data in real time at a set frequency and transmit it over the network to the adaptive control parameter adjustment module in the data processing layer.

[0056] Adaptive Control Algorithm Implementation: The adaptive control parameter adjustment module uses adaptive control algorithms, such as model predictive control (MPC) or fuzzy control algorithms, to calculate real-time adjustments to the robot's motion control parameters based on sensor data. The motion stability control module automatically adjusts the robot's motion control parameters, such as joint torque, velocity, and acceleration, based on these adjustments to adapt to varying workloads and operating environments.

[0057] Improved execution of robot motion stability: The industrial robot at the execution layer adjusts its own motion state according to the control instructions sent by the motion stability control module, improves the stability and reliability of the motion, and reduces material falling and sorting errors caused by unstable motion.

[0058] Implementation of intelligent monitoring and management platform for material sorting process

[0059] Data Collection and Storage: The intelligent monitoring and management module, through a data interface with the industrial robot material sorting system, collects various data from the sorting process in real time, such as material identification results, sorting paths, robot motion status, and task completion status. The collected data is formatted and stored in a database, creating a digital archive of the sorting process.

[0060] Data Analysis and Visualization: Utilizing big data analytics technology, we analyze and mine the data stored in the database to extract valuable information, such as sorting efficiency analysis, material identification accuracy statistics, and equipment failure warnings. Through visualization technology, we present analysis results in the form of charts and reports on the intelligent monitoring and management platform interface, providing managers with intuitive and clear sorting process monitoring and decision support.

[0061] Intelligent early warning and fault diagnosis: The intelligent monitoring and management platform features intelligent early warning and fault diagnosis capabilities. By setting thresholds and analyzing models, it monitors data during the sorting process in real time. When data exceeds set thresholds or an abnormality occurs, the platform automatically issues an early warning signal and performs fault diagnosis, prompting management personnel to take appropriate measures to ensure the smooth progress of the sorting process.

[0062] Example 2

[0063] Example of improving etching precision

[0064] In the etching process of a 14nm chip production line, two sets of etching equipment of the same model were selected for comparison:

[0065] Traditional etching systems use a single ion beam sensor (accuracy ±0.5eV) and fixed parameter control. Post-etch inspection of 100 wafers revealed an average critical dimension deviation of ±8nm, resulting in a 12% scrap rate and requiring frequent manual adjustments.

[0066] Multi-field sensor fusion system: Deploys a five-dimensional sensor matrix (ion beam energy ±0.1eV, angle ±0.05°, beam current density ±0.5μA / cm²; electric field strength ±0.1V / m; gas flow ±0.1sccm). This system uses an attention-based fusion network (four-layer fully connected architecture with 128-64-32-16 neurons) to adjust etching parameters in real time. Under the same conditions, the mean deviation of critical dimensions is controlled to ±1.2nm, the scrap rate is reduced to 2.3%, and production efficiency is increased by 35% without manual intervention.

[0067] Etching uniformity optimization embodiment

[0068] In the 300mm wafer etching scenario, the comparison between the traditional process and the new algorithm is as follows:

[0069] Traditional process group: Using a fixed etching path, the depth difference between the edge and center of the wafer after etching reached ±65nm, and the yield rate was 78%.

[0070] Multimodal data system: Integrates a laser interferometer (accuracy ±0.01μm), an atomic force microscope (resolution ±0.1nm), and a fuzzy control algorithm. By dynamically adjusting the gas flow distribution (formula ΔF = g(Q), where Q represents fused data), the etch depth difference is reduced to ±12nm, the yield rate is increased to 96%, and the etch rate in edge areas is increased by 22%.

[0071] Rate-quality collaborative optimization implementation example

[0072] For the 5nm FinFET etching process:

[0073] Traditional parameter settings: When the etching rate is 25nm / min, the defective rate is 15%, and when the rate is increased to 35nm / min, the defective rate soars to 32%.

[0074] Multi-objective optimization system: Using the PPO reinforcement learning algorithm (clipping coefficient 0.2, value function coefficient 0.5), while ensuring that the key dimension deviation is ≤±2nm and the uniformity deviation is ≤±15nm, the etching rate is stabilized at 42nm / min, the defective rate is controlled at 4.1%, and the overall production capacity is increased by 58%.

Claims

1. An intelligent industrial robot sorting system based on a large model algorithm, characterized in that: include: A high-precision material recognition algorithm module based on multi-source information fusion uses multiple sensors to collect and fuse multi-source material information in real time, and combines deep learning to achieve high-precision recognition; A dynamic sorting path planning algorithm module based on reinforcement learning models the sorting process as a reinforcement learning environment and optimizes the path through a state-action-reward mechanism; Based on the distributed collaborative optimization multi-robot task allocation and coordination algorithm module, a distributed objective function is constructed to achieve multi-robot task collaboration; A robot motion stability improvement system module based on adaptive control, which monitors and adjusts motion parameters in real time through sensors; The intelligent monitoring and management platform module for material sorting enables full-process data collection, analysis and early warning.

2. The system according to claim 1, wherein: In the high-precision material recognition algorithm module based on multi-source information fusion, multiple sensors include visual sensors (accuracy ±0.1mm), lidar (ranging accuracy ±2mm), infrared sensors (temperature resolution ±0.5℃) and tactile sensors (pressure accuracy ±0.1N). The deployment positions are determined through finite element simulation and actual testing.

3. The algorithm module according to claim 2, characterized in that: Multi-source information fusion adopts a fusion network based on the attention mechanism, and the fusion formula is: Among them, xi is the data of each sensor, αi is the attention weight, which is dynamically calculated by a three-layer fully connected network with 128, 64, and 32 neurons respectively, and optimized by the back propagation algorithm.

4. The system according to claim 1, wherein: In the reinforcement learning-based dynamic sorting path planning algorithm module, the state space includes the material position P, the obstacle position O, and the robot posture Q. The action space is the displacement Δx, rotation Δθ, and grasping action g in the Cartesian coordinate system. The reward function is defined as: Where d is the path length, c is the number of collisions, and λ is the collision penalty coefficient with a value between 0.5 and 1.

5.

5. The algorithm module according to claim 4, characterized in that: The Proximal Policy Optimization (PPO) algorithm is used for training, with an experience replay buffer capacity of 100,000 entries, a learning rate of 1e-4, 10 epochs per iteration, and the target network updated every 2000 steps.

6. The system according to claim 1, wherein: In the multi-robot task allocation and coordination algorithm module based on distributed collaborative optimization, task allocation adopts the Hungarian algorithm, collaborative control adopts the distributed consistency protocol, communication delay is ≤10ms, and task completion time variance is ≤5%.

7. The system according to claim 1, wherein: In the adaptive control-based robot motion stability improvement system module, the model predictive control (MPC) algorithm is used, with a prediction time domain of 5 steps and a control time domain of 3 steps. The state vector includes the joint angle θ, angular velocity θ˙, and load m. The optimization goal is to minimize the tracking error:

8. The system according to claim 1, wherein: In the intelligent monitoring and management platform module, the data acquisition frequency is 100 Hz, and the long short-term memory network (LSTM) is used for fault prediction. The input sequence length is 100, the hidden layer neurons are 256, and the prediction accuracy is ≥95%.

9. The system according to claim 1, wherein: In the multi-robot collaborative sorting scenario, task priority is determined by fuzzy logic. The input is the material value V, urgency U and volume S. The output is the priority coefficient P. The membership function adopts a triangular distribution.

Images