Files

44: Industry in the Future: UTC an academic partner for enterprise

The ongoing digital transformation of industry is a major societal challenge. For UTC, accompanying a growing number of companies during the changes, the phenomenon represents an increasingly strategic field for studies. This Dossier zooms in on the university’s main activities and on the specific nature of its approach to the industries of the future.

44: Industry in the Future: UTC an academic partner for enterprise

Data Analytics (DA) in the quest for industrial excellence

UTC can rely on a precious advantage when it comes to engaging research into industrial applications of automated data analysis: ‘pluridisciplinarity’. Among current projects: designing aids to decision tools to optimize manufacturing process control and product quality assurance (QA).

Production lines today are becoming increasingly fitted with sensors that record large quantities of parameters about the process in hand and the products coming off the line. With the ‘Internet of Things’ (IoT) or object-oriented Internet, these self-same products will be transmitting increasing amounts of data – about use, level of wear and tear, etc. Thanks to increased data storage capacities and associate data processing possibilities, the industrialists hope to draw new knowledge and significant added-values from the data flows: improvement in product design using the information incoming from real life utilization, reduction of the number of production rejects ... this are the challenge and stakes of automated data analysis, a theme where UTC can provide further, complementary skills.

 

Making the data speak

Heudiasyc, UTC’s laboratory for computer sciences and applications (ICTs) already has amassed a lot of field experience in data analysis techniques, a mix of statistical methods, machine-learning (automat learning algorithms) and artificial intelligence (AI). “These techniques aim at identifying regular patterns in a data flow – firstly from a purely descriptive stance”, explains Sébastien Destercke, a research scientist at UTC-Heudiasyc.

“For example, in order to try to explain a manufacturing faulty part in products, an a posteriori analysis can be conducted on the variables as they evolve in use. Such regularly noted patterns can be used to correlate parameters that the company’s professional experts had already suspected or to reveal connexions that they had not identified hitherto, especially when the faults resulted from combinations of numerous input variables. From there on, the idea was to build predictive models – applying the hypothesis that the past will resemble the future, quite probable in a relatively stable manufacturing process, in which case the model will be able to predict a plausible output from a given set of observed inputs. The idea, notably, is to use these predictive algorithms during the production phases, to anticipate in real time, risks of faults occurring in product parts. The models can also be ‘prescriptive’ and suggest actions to correct a faulty or improve a product. In certain cases, we can even imagine taking this still further: automating the decision implementation itself, something we are currently trying to do with driverless cars. Nonetheless, it still proves very difficult to replace human expertise in cutting-edge industrial sectors, even using the most recent methods, the objective here often consisting of providing the process operators with aids to their decision-taking, thereby helping them carry out complex tasks”.

A new QA method

UTC’s Roberval Laboratory, combining expertise capabilities in process control and applied mathematics, is engaged in developing tools to optimize the control of manufacturing processes and control of quality assessment (QA). For example, Roberval scientist have designed and built a unique method combining multi-variable Statistical Process Control (SPC) and process diagnosis. “SPC is used to detect abnormal shifts in the critical characteristics that define product conformity with design specs, and this may be used to prevent production of more faulty parts”, explains Nassim Boudaoud, a Roberval research scientist. “It is a method currently used in industrial sectors, but generally speaking only in its simplest format, i.e., with the capacity to track the characteristics of a product, one by one. But there are often correlations between these characteristics. To illustrate: the control process for the assembly tolerances for a car door may point to a defect without compromising the overall integrity of the assembled door unit, inasmuch as there are also built-in geometric compensations.” 

Notwithstanding, it is practically impossible to establish the decision rules applicable to monitoring several characteristics simultaneously. Thanks to Data Analysis (DA), the tools on which the Roberval Lab is working can achieve this. Moreover, it is a truly novel innovation, combining, as it does, product and process data.” What we offer is a hybrid approach that aims at showing a better detection of ‘spec drifts’ and a forecast of evolution”, underscores Nassim Boudaoud. “In concrete terms, thanks to a historic analysis of data, we can infer connections between product characteristics and various observable states in the process and tis allows us to say, at a given time T, whether the process is operating perfectly, according to plan, or not. In the latter event, we can predict evolution to anticipate a future process state that will results in faulty product parts”.

The immediate objective of UTC is to push these investigations further, testing these tools Roberval have designed and developed with real data flows and thereby proving their capacity to improve operation performance statistics.