With the gradual maturity of a new generation of technologies such as artificial intelligence, the wave of the fourth industrial revolution has swept the world, and intelligent manufacturing has also become an important strategy for the mainland. As one of the key areas of the industry, the intelligent transformation of the automobile industry has also received the attention of the upstream and downstream of the industrial chain.
Recently, in the final assembly workshop of the "lighthouse factory" and "intelligent manufacturing demonstration factory" of the automotive industry, the Foton Cummins Engine Production Plant, there has been a shutter sound. The sound of "clicking" is like the pulse of the automobile industry in the new era, showing the strong vitality of the collision between artificial intelligence technology and the automobile industry. Supporting this vitality is the Deep Spring Industrial Quality Inspection and Training Platform (hereinafter referred to as the Deep Spring Platform) based on the integration of optical, electromechanical and soft computing built by SenseCore AI large device. Relying on this platform, Foton Cummins not only realizes the automatic detection of surface defects and assembly defects of key engine components, greatly improves the efficiency of quality inspection, but also enhances the ability of workers, liberates workers from boring and tedious quality inspection work to engage in higher level work such as technology, and ultimately enhances the overall competitiveness of the enterprise and takes a key step towards the transformation and upgrading of intelligent manufacturing.
Senior Project Manager of Foton Cummins said: "The combination of SenseTime's Deep Spring platform and our production practice has achieved the rapid launch of AI quality inspection capabilities in multiple scenarios, improved the efficiency of quality inspection, helped us improve the level of intelligent manufacturing, and laid the foundation for the agile innovation and application of AI quality inspection in the field of automotive industry manufacturing." ”
Deep Spring Platform Landed, Empowering Industrial Quality Inspection "Quality", "Sensitive" and "Soft"
In today's car factories, robot applications are already everywhere. However, in the process of quality inspection, there is still a huge demand and challenge for manual visual inspection. To solve this problem, it is necessary to interpret the three-step song of "quality, sensitivity and softness" to make the automobile production action as accurate, brisk and flexible as the musical melody.
"Quality", that is, to be accurate enough. There are generally two indicators to measure accuracy, the rate of missed detection and the rate of false detection. The missed inspection rate affects the yield rate of production, and the false detection rate affects the production capacity, only the missed detection rate and the false detection rate are low enough, which means that the quality inspection product can be practically applied.
"Min", that is, the whole process of landing should be efficient enough, so that all departments of the factory can efficiently land AI intelligent applications, make full use of every inch of space on the production line, and let workers' hands and eyes and intelligent applications work together and cooperate perfectly, and "man-machine collaboration" can achieve efficient production.
"Soft", that is, defect detection should have sufficient flexibility, with the development of the times, the definition of large-scale production has changed from a single large batch to a multi-piece small batch. It is required to be able to support the production of a variety of parts in a limited production line space.
It can be said that the deep spring platform has performed well in the three aspects of "quality", "sensitivity" and "softness", which also ensures that the deep spring platform can form a real landing practice in many stations of Foton Cummins.
"Qualitative" improvement: AI precision exceeds the industrial red line
Industrial scenarios often face the "three more" problems of many types of parts, models, and defects. First, there are many types of parts, according to statistics, there are as many as tens of thousands of parts on a car; second, there are many parts and components, even if it is the same kind of parts, the models used in different cars are often very different; the third is that there are many defects in parts, and different production processes will produce different defects. It becomes extremely difficult to meet the "quality" requirements for each defect detection scenario.
In order to solve this problem, the Deep Spring platform provides solutions from three angles: multi-optical solution support, multi-component form support, and multiple quality inspection support. In terms of multi-optical solution support, the platform supports more than ten different image processing schemes, from bright field to dark field, from coaxial light to dome diffuser light, which can well complete image quality inspection, data preprocessing, data enhancement, etc. Industrial imaging is characterized by a large range of pixels over the entire image, but the defects are usually small. In order to detect such defects, the Deep Spring platform provides high-precision image segmentation capabilities, which can accurately detect defects with only 3-5 pixels in extreme scenes with hundreds of millions of pixel resolutions, so that the defect location can be seen at a glance.
In terms of multi-part form support, the platform supports a variety of zero forms such as concave parts, arched parts, and polyhedra. When shooting complex parts, there will also be problems in judging whether the defect standard is not clear, and it is difficult to avoid errors in labeling. In response to this situation, the Deep Spring platform provides technologies such as Auto Denoise to ensure that the training dataset can converge to the optimal point in the presence of noise.
In terms of multiple quality inspection support, for assembly defects and production process defects in industrial production, a complete model training system of unsupervised, semi-supervised and strong supervision is provided to support it. It is used in conjunction with inference to ensure the lowest degree of missed tests and the highest quality detections.
The practice of "sensitivity": process improvement enters a rapid iteration mode
Each AI quality inspection project is a set of systematic engineering, involving multiple systems of opto-mechanical-electrical soft calculation, and will also involve the update and reorganization of the original quality inspection process, how to perfectly integrate intelligent technology and production line to improve the efficiency of the production line, the deep spring platform provides a perfect solution, and the iteration of the process is changed from "month" to "week" unit. Truly practice "min".
The Deep Spring platform provides multi-faceted solutions for pre-production - lightweight production line, production - software and hardware integration efficient reasoning, and post-production - rapid iteration of the process:
Lightweight production line, the deep spring platform provides two kinds of reasoning product forms of cloudification and lightweight edge side, helping the production line to be clouded, lightweight and wireless. Reduce the variety of equipment on the production line. Let workers have a clear production vision when producing. Create a smart "production pod" to lay the foundation for faster production beats. It also lays a solid foundation for various changes such as future changes in the production line.
The combination of soft and hardware integrated efficient reasoning, combined with the company's self-developed AI chip, forms a software-hardware integrated efficient reasoning scheme, ensuring that SenseTime's academic and industrial practice results can maximize the use of AI computing power and achieve optimal algorithm accuracy and speed. The deep spring platform also opens up the end equipment on the production line such as industrial cameras, PLCs, and robotic arms, realizing the collaborative work of multiple devices, so as to ensure that AI quality inspection can meet the limit beat of hundreds of milliseconds or even tens of milliseconds on the production line.
Rapid process iteration, through real-time data analysis of process production defects, the number, category, grade and other information of defects are provided to the production managers of enterprises in the form of reports, so that they can see the production quality at a glance, realize the quantitative data to guide the production process improvement, and greatly improve the production quality management efficiency.
At Foton Cummins' station, with the accurate identification of defects and the summary analysis of defects, the factory can distinguish between critical defects and general defects faster. Process iterations for key defects rely on data and are already being optimized on a weekly basis. Truly interpreted the "sensitivity" of the effect of the deep spring platform.
The pursuit of "soft": low code supports flexible character inspection
The traditional automobile industry production mode is "rigid" production, that is, to achieve the mass production of a single product. However, with the development of production concepts and technologies, industrial enterprises have begun to turn to "flexible" production, emphasizing the production of small batches of products with high quality, which puts forward higher requirements for both manufacturing and quality testing.
In order to solve this problem, the Deep Spring platform provides industrial model training components, inference workflow scheduling components, and report configuration components. The industrial model training component can make full use of the capabilities of SenseCore AI large devices of SenseTime, like a pipeline factory, to achieve the underlying abstraction of algorithm models for different scenarios. The Deep Spring platform not only integrates the SenseCore AI device, but also creates a zero-code industrial model production platform on top of it, using the "teach people to fish" method to fully empower the ability of AI model production to industrial developers. The Deep Spring platform also integrates AutoML technology, which is different from conventional AutoML, which is specially designed for industrial small data sets to achieve the best balance of computing power and accuracy in subdivision scenarios such as industrial quality inspection, so as to achieve a real parameter. Practice shows that in the multiple stations of Foton Cummins, the AI model obtained by super-parameter automatic search has a significantly better missing kill rate and manslaughter rate than the ordinary model.
The inference workflow scheduling component can flexibly adjust the quality inspection process after the AI model is freely customized. The quality inspection process relies on the close cooperation of various parts of the opto-mechanical-electrical soft calculation, and to achieve effective adjustment, it is necessary to consider it as a whole. The Deep Spring platform integrates industrial cameras from mainstream manufacturers in the industry, whether it is megapixel or 10 million pixels, whether it is a area scan camera or a line scan camera, it can be plugged and played. At the same time, the Deep Spring platform also supports a variety of PLC mainstream protocols, which can achieve interconnection with dozens of PLCs. When the opto-mechanical and electrical equipment is fully connected, as the hub of the quality inspection production line, the deep spring platform will flexibly design the quality inspection process through the low-code platform, so as to realize the control of various mechanical automation equipment such as assembly lines and robotic arms, and ensure the efficient and collaborative work of multiple equipment. In this low-code platform, industrial cameras, PLCs, robotic arms, etc. are abstracted into nodes, and users only need to visualize the configuration to make different arrangements and combinations of nodes, so as to easily redefine the quality inspection process. With a few minutes of reconfiguration, it is possible to quickly switch the quality inspection process and ensure the support of "soft".
In the report configuration components, the deep spring platform provides flexible report configuration, using the design of horizontal and vertical tables, to ensure that the production line produces different types of equipment, the system can use plug-in loading, the report will automatically display the quality of production parts, and truly achieve flexible production in all aspects.
Taking The quality inspection scene of Foton Cummins as an example, in the bearing inspection station, the defect detection of the surface of the two bearings needs to be completed at a time, and because the bearing shape is arched, each bearing needs two cameras to shoot, and one station will contain 4 industrial cameras. The working mode, camera sequence and result integration method of these four industrial cameras will affect the final quality inspection result given to the PLC, and ultimately affect the quality inspection process. The Deep Spring platform abstracts various devices and AI models into nodes, and users can define the work order and process of nodes by themselves, so as to customize the quality inspection business process. When the customer's scene is changed from using a pad detection station with 4 cameras to a flywheel shell gluing station using 1 camera, the user only needs to re-drag the node and design a quality inspection business process that conforms to the flywheel shell gluing station, and the whole process can take effect.
At present, the deep spring platform of optical,mechatronic and soft computing integration built by SenseCore AI large device of SenseTime has landed in multiple stations at Foton Cummins. Let AI quality inspection find its own space for shining in the century-old automobile industry. With being rated as a "lighthouse factory" and an "intelligent manufacturing demonstration factory", Foton Cummins will continue to promote the transformation and upgrading of engine production to intelligent manufacturing, and SenseTime will continue to help the development of the automotive industry through a powerful deep spring platform, so that the production of the automotive industry is as beautiful and brisk as waltz, and help the automotive industry continue to move forward in the wave of Industry 4.0.