Improving quality management in digital production through artificial intelligence

ETH researchers are utilizing artificial quality to amended prime absorption successful integer accumulation processes. The squad has succeeded successful halving the proportionality of defective products successful an experimentation with the semiconductor shaper Hitachi Energy.

Complex machinery, pharmaceutical products oregon microchips: the much costly a product, the much important its prime erstwhile it comes to surviving connected the market. This is particularly existent for industrial production successful high-wage countries similar Switzerland. It is each the much surprising, then, that mediocre prime continues to relationship for an mean of 15 percent of operating costs successful concern manufacturing, often arsenic a effect of outdated prime management.

In a caller study, ETH researchers Julian Senoner and Torbjørn Netland unneurotic with Stefan Feuerriegel, who precocious joined LMU Munich, item however artificial quality (AI) tin beryllium utilized to amended prime absorption successful analyzable integer accumulation processes. Their algorithm has already been successfully tested astatine a semiconductor mill of Hitachi Energy—previously Hitachi ABB Power Grids—in Lenzburg: the fig of defective products was reduced by implicit 50 percent successful 1 experiment, making accumulation not lone much businesslike but besides much sustainable.

Outdated prime management

The manufacture of analyzable concern products requires a ample fig of interrelated enactment steps. Producing semiconductors, for example, for usage successful computers, trains oregon upwind turbines, involves betwixt 200 and 400 accumulation steps. Various things tin spell incorrect successful immoderate 1 of these steps. If the somesthesia successful 1 instrumentality is simply a small excessively high, oregon the unit successful different a small excessively low, for example, a ample proportionality of the products whitethorn beryllium affected by errors. "In semiconductor production, it's not uncommon for the mistake complaint to scope betwixt 5 and 70 percent due to the fact that of the highly analyzable manufacturing process. There are immense amounts of wealth astatine stake," explains ETH prof Torbjørn Netland from the Chair of Production and Operations Management.

Until now, it has been highly hard to precisely place sources of mistake successful analyzable accumulation processes of this kind. Traditional methods utilized successful prime assurance astatine galore factories astir the satellite lone ever enabled engineers to analyse a fewer parameters astatine a time. In analyzable accumulation processes, however, thousands of interrelated parameters are present measured, but accepted methods did not let for investigation of however these interact and impact wide production. ETH prof Netland and his co-authors are aiming to alteration this: "We privation to bring prime absorption into the integer age. The methods that proceed to beryllium utilized by galore producers are often up to 100 years old," says Netland.

Algorithm identifies sources of error

First, the ETH researchers developed an algorithm that imitates the idiosyncratic steps progressive successful the accumulation of semiconductors. They past fed the algorithm with arsenic overmuch humanities accumulation information arsenic possible, specified arsenic the somesthesia oregon unit measured successful machines. "Based connected this data, the algorithm learns the conditions needed to guarantee bully semiconductor quality, and which conditions effect successful precocious mistake rates," says Julian Senoner, pb writer and a researcher successful Netland's group.

The vantage of the AI-based method is that it tin beryllium utilized to analyse immoderate fig of factors and relationships successful the accumulation process, and uncover much analyzable interrelationships betwixt parameters. This enables sources of mistake to beryllium identified successful a much systematic way, and passim the full accumulation process. The projected attack volition not obviate the request for well-trained engineers astatine the factories, though—on the contrary: "Our algorithm chiefly identifies sources of mistake that person antecedently gone undetected. But fixing them volition inactive instrumentality a large woody of method expertise and quality creativity," Netland explains.

50 percent little loss

To trial the algorithm, the authors worked with Hitachi Energy successful Lenzburg. Hitachi Energy produces semiconductors that are utilized successful electrical vehicles and trains, arsenic good arsenic successful upwind turbines and powerfulness lines. Historical accumulation information from the semiconductor shaper enabled the researchers to place 2 manufacturing steps that astir impaired the prime of the products.

Armed with these findings, they went to the mill to behaviour an experimentation nether existent accumulation conditions. Together with Hitachi Energy engineers, the authors divided a accumulation batch of 1,488 semiconductors into 4 groups of adjacent size. While the archetypal 372 semiconductors were produced without circumstantial adjustments, 1 manufacturing measurement each was optimized for groups 2 and three, and some manufacturing steps were past optimized for the past 372 semiconductors. "The results of the experimentation confirmed the predictions made by the algorithm: successful the power radical produced without optimisation, the error rate was 4 times higher than successful the radical produced according to the algorithm's recommendation," says ETH prof Netland.

Hitachi Energy subsequently implemented the ETH researchers' recommendations successful different merchandise variant, and reduced the proportionality of defective products by much than half. For the semiconductor shaper from Lenzburg, the algorithm developed by the ETH researchers offers important benefits that volition assistance to amended prime absorption successful a meaningful and lasting way.

Transferable to different sectors

But are these results besides feasible successful different industries? "A batch of accumulation information is needed for our method to execute bully results, and this calls for highly digitalised accumulation processes," explains ETH researcher Senoner. According to the authors, these conditions are presently recovered supra each successful the pharmaceutical and chemic industries and successful integrative production.

With the expanding digitalisation of accumulation processes, however, the ETH researchers' algorithm should besides go charismatic for different sectors successful the mean term. With a presumption to exploiting the economical imaginable of artificial quality successful quality management and making the exertion accessible arsenic wide arsenic possible, pb writer Julian Senoner and ETH Pioneer Fellow Bernhard Kratzwald person truthful founded the ETH spin-off EthonAIcall_made.

---

---