Planning of artificial intelligence big model in industrial robots

Author: Peng Xiantao, Wang Peng Zhejiang Hengyi Group Co., Ltd. Zhejiang University-Hengyi Global Future Advanced Technology Research Institute

Introduction

With the development of manufacturing, AI has become an indispensable part of the scene and plays a huge role in the production site. With the deepening and expansion of application scenarios, more requirements have been put forward for its application. By configuring robots, robots can be made more intelligent and able to adapt to more complex scene requirements; at the same time, the development of AI has also provided strong support for the development of industrial robots. After traditional artificial intelligence is deployed on site, it can be executed efficiently. Due to the complexity and changeability of industrial scenes, untrained abnormal data will appear, resulting in unrecognizable output. The emergence of artificial intelligence large model technology can enable industrial robots to handle complex industrial scenes more flexibly, further improve accuracy and efficiency, assist enterprises in digital and intelligent construction, and promote the transformation and upgrading of the manufacturing industry.

01

introduction

Industrial robots have become one of the common and indispensable important equipment in industrial production, bringing great convenience to the development of industry. Due to the large number of industrial scenarios and diverse needs, in order to meet the needs of actual business scenarios, the form of industrial robots on site has gradually evolved from traditional arms to customized AGVs, gantry truss robots, AGVs equipped with mechanical arms and other forms. With the empowerment of new generation technologies, industrial robots equipped with industrial cameras (area array cameras, line scan cameras, 3D cameras, etc.), intelligent temperature sensors, high-sensitivity microphones and other sensors can solve more complex needs, thereby greatly improving the utilization rate of industrial robots and generating more practical value. The main application of artificial intelligence is to deploy the trained reasoning/detection model, process the real-time transmitted images and other data, and execute the next action according to the output results; this method has an absolute advantage in fixed scenarios. If other abnormal data is collected during the movement of industrial robots, the abnormality will not be identified if the model is not trained, which poses certain risks. The emergence of artificial intelligence large models has given industrial robots more capabilities while improving the accuracy of detection, which has greatly promoted the development of the manufacturing industry towards intelligent manufacturing.

02

Planning of artificial intelligence big model in industrial robots

Companies that use a lot of industrial robots already have a good digital foundation and, based on actual needs, carry out overall design planning for the application of large artificial intelligence models in the field of industrial robots.

2.1 Industrial Robot Layer

Industrial robots deployed on site currently have various forms, such as multi-joint robotic arms, multi-degree-of-freedom truss robots, AGVs equipped with light-load robotic arms, etc., to meet various operational needs. Equipping industrial robots with high-definition cameras to achieve functions such as precise positioning, deviation correction, defect detection and dimension measurement is a mature and common solution; thermal imagers, spectrometers, gas analyzers and microphones can also be installed on industrial robots to collect data in different scenarios and give full play to the capabilities of industrial robots. Industrial robots are widely used in my country, covering multiple scenarios in process and discrete manufacturing industries. Non-standard customized industrial robots have obvious advantages in localization, with high customization and flexibility, but there are also certain challenges, such as: overall long-term stable operation, MTBF (Mean me Between Failure) improvement; domestic replacement of core components such as, and; localization of industrial software for large-scale scheduling systems; breakthroughs in the localization of multi-joint high-end robots, etc.

2.2 Layer

It is the basic demand of industrial sites, and its stability is the first priority. The application design level of this layer is quite different, and it is necessary to uniformly plan the industrial and office networks, establish corresponding security measures, and ensure the stability of the network. The common solution with good application effect is: the industrial network uses industrial-grade (such as, Phoenix and other industrial network management switches) components of the optical fiber ring network, and the office network uses commercial switches (such as H3C, Huawei and other brands) components of the IT network, and firewall protection is used between the two. There are many communication protocols for sensors such as industrial robots and cameras. This layer needs to ensure the accuracy and timeliness of data transmission. It can consider using the transmission of pictures and other large-size data, and other protocols to interact with detection results, instructions and other data. For the process industry, the use of RTU, 5G and other technologies is a more practical and safe solution. The stability of the communication interface must take into account high concurrency, scalability, compatibility, etc. When it comes to data transmission across companies and platforms, it is possible to consider building an industrial Internet platform and adding technologies such as, quantum communication to enhance the security of communication.

2.3 Model Layer

With the development of artificial intelligence, a relatively complete theory has been provided for the model layer, and scholars have also done a lot of theoretical and applied research. The key steps of model layer application are data collection, data processing, model training and model deployment. The core is to solve the "unchanged" needs in the production process through trained models (because the production process is relatively stable and repetitive, the detection needs for equipment, etc. are relatively fixed), such as various defect detection models, measurement models, prediction models, etc. The final applied model is the embodiment of the ability to solve complex problems and generate value. Several commonly used models are briefly introduced as follows.

(1) CV model On-site detection/recognition involving images and videos are all computer vision recognition, and the use of CV models is a reliable and mature solution. In actual industrial applications, it is necessary to comprehensively consider detection accuracy, rhythm, stability and cost-effectiveness to ensure that AI can be successfully implemented and stably operated through engineering projects to generate value. For example, industrial products with large sizes and complex structures need to be inspected for appearance defects. The ideal solution is to capture the product in the air for all-round, no-dead-angle photography and inspection. There are two major risk points in actual operation: slow rhythm will affect production; dynamic equipment has high loss, high failure rate and high operating cost. According to experience, the optimal engineering solution is to use static equipment design to evaluate defects in parts of the product that cannot be photographed. If the probability of defects is low and the impact is small, they can be ignored. Otherwise, improvements need to be made to the process, equipment, management, etc. to reduce the frequency of defects. For scenarios where only defects are detected, traditional methods such as grayscale recognition, SVM (Suppt Vector Machine), SIFT (Scale-Invariant Feature Transform), HOG (Histog of Oriend Gradient), ORB (Oriented FAST and Rotated BRIEF), and LBP (Local Binary Patterns) can be used to save computing power and investment. For complex tasks that require marking defect locations, defect types, and distinguishing levels according to defect severity, the use of CNN algorithms is the best solution, such as YOLO, VGG, ResNet, AlexNet, and RevNet. First, collect samples, and then train the processed samples (marked defects, plus) through the algorithm model. After the model reaches the preset detection accuracy, the model is deployed for use. When the framework uses the relevant algorithm model for training, the training process is currently still a black box (that is, difficult to explain), and the quality and quantity of the samples have a great impact on the final detection model. In engineering applications, increasing the number of actual samples (derived from actual production, not artificially created samples; the quality of annotations should be high) can significantly improve the training results of the model. In order to make the algorithm output the inference model according to the expected effect, scholars have done a lot of work, among which constructing a loss function is more common and easier to implement in engineering. For example, Berkan Demirel et al. proposed a new meta-tuning loss function, which significantly improved the detection results. In engineering applications, designing a loss function based on actual conditions will achieve better results.

(2) The robot prediction model collects and obtains the operating data of the industrial robot, such as action duration, load, voltage, operation trajectory, battery power and charging and discharging status, etc., and uses Transformer or GNN training prediction models to predict the maintenance status and faults of the industrial robot, realize the full life cycle management of the industrial robot, rationally plan the use of industrial robots and their spare parts, avoid operation with defects, and improve the service life and efficiency of the industrial robot.