Artificial intelligence has been used in a EUREKA-backed project to develop a quality control system that greatly reduces the risk of human error on assembly lines.
The four teams who worked on Project E!3450-QSPAI have achieved a non-contact activation program that commands and monitors laser-scanning for precise panel measurement, and triangulation methods for positioning components. Although existing technologies, notably in computing and lasers, have been used, it is their integration that makes this control system unique.
The project was instigated and led by Trimo d.d., a specialist engineer and producer of prefabricated steel buildings and components, based at Trebnje, some 50 km south-west of Ljubljana. The other partners were two faculties of the University of Ljubljana – Computer and Information Science, and Electrical Engineering – and the Institut fuer Sandwichtechnik, of Mainz, Germany.
Trimo wanted greater quality control during its manufacture of Trimoterm lightweight, fireproof sandwich-panels as the process was prone to delays and other glitches, including human actions, which impact on product quality.
Operators were unable to monitor continuously each of the many production steps; neither could they predict all the indirect consequences of actions performed on the line. And manual inspection could miss such faults as measurement errors, and colour deviations between batches.
The main concern was the long reaction time in correcting errors. As destructive and discrete analysis of sample panels was practical only a few times each day, faulty panels could go unnoticed until arrival at the construction site, or, worse, after application.
The project task teams have created a system that achieves control of disparate parameters, ranging from the type and quality of input materials to the settings and current state of the assembly line. The unified system governs both the speed of production, and, even more importantly, the individual processes that take place on the line.
One of the first development tasks was to write a program for artificial intelligence (AI) – advanced data processing – that could “learn” the manufacturing process by “mining” the records of assembly line parameters. AI proved its value in detecting errors, discovering correlations between parameters, and indicating areas where the process could be improved.
Initial monitoring of the process identified numerous reasons for delays. These reasons fell into three basic categories: organizational demands, processing errors, and inappropriate quality of material. Organisational delays could occur when equipment was re-set for different types of product, during the changeover to other components, and even in the scheduling of workers' rest breaks. Production delays included breakdowns of mechanical equipment, poor line control, and process errors. The human factor proved especially difficult to determine as actions could have indirect influences.
Results and outlook
A prototype system – which was installed without disrupting the factory’s production schedule – is running successfully, but without the AI program. Although AI was central to the initial phase of development, the reliability of the learning algorithms (instruction sequences) needs to be improved, especially concerning the measurement of input material, and the speed of gathering information.
The present system, however, is providing a high degree of control, resulting in a significant increase in productivity with fewer rejects. Viktor Zaletelj, the QSPAI Project Manager at Trimo, says that feedback from continuous monitoring of the entire process enables operators to correct faults almost as soon as they develop, and even to spot potential problems.
“These results have encouraged the participants to continue developing the AI program so that it can be interfaced with the control system’s measurement and data processing capabilities. It is feasible that we can fulfil our original intent to build a system that mostly relies on ‘machine learning’ to maintain quality.”