Abstract :
[en] Nowadays, industrial companies are engaging their global transition toward the fourth industrial revolution (the so-called Industry 4.0). The main objective is to increase the Overall Equipment Effectiveness (OEE), by collecting, storing and analyzing production data. Several challenges have to be tackled to propose a unified data-driven approach to rely on, from the low-layers data collection on the machine production lines using Operational Technologies (OT), to the monitoring and more importantly the analysis of the data using Information Technologies (IT). This is all the more important for companies having decades of existence – as Cebi Luxembourg S.A., our partner in a Research, Development and Innovation project subsidised by the ministry of the Economy in Luxembourg – to upgrade their on-site technologies and move towards new business models. Artificial Intelligence (AI) now knows a real interest from industrial actors and becomes a cornerstone technology for helping humans in decision-making and data-analysis tasks, thanks to the huge amount of (sensors-based) univariate time-series available in the production floor. However, such amount of data is not sufficient for AI to work properly and to make right decisions. This also requires a good data quality. Indeed, good theoretical performance and high accuracy can be obtained when trained and tested in isolation, but AI models may still provide degraded performance in real/industrial conditions.
In that context, the problem is twofold:
• Industrial production systems are vertically-oriented closed systems that make difficult their communication and their cooperation with each other, and intrinsically the data collection.
• Industrial companies used to implement deterministic processes. Introducing AI - that can be classified as stochastic - in the industry requires a full understanding of the potential deviation of the models in order to be aware of their domain of validity.
This dissertation proposes a unified strategy for digitizing an industrial system and methods for evaluating the performance and the robustness of AI models that can be used in such data-driven production plants.
In the first part of the dissertation, we propose a three-steps strategy to digitize an industrial system, called TRIDENT, that enables industrial actors to implement data collection on production lines, and in fine to monitor in real-time the production plant. Such strategy has been implemented and evaluated on a pilot case-study at Cebi Luxembourg S.A. Three protocols (OPC-UA, MQTT and O-MI/O-DF) are used for investigating their impact on the real-time performance. The results show that, even if these protocols have some disparity in terms of performance, they are suitable for an industrial deployment. This strategy has now been extended and implemented by our partner - Cebi Luxembourg S.A - in its production environment.
In the second part of the thesis dissertation, we aim at investigating the robustness of AI models in industrial settings. We then propose a systematic approach to evaluate the robustness under perturbations. Assuming that i) real perturbations - in particular on the data collection - cannot be recorded or generated in real industrial environment (that could lead to production stops) and ii) a model would not be implemented before evaluating its potential deviations, limits or weaknesses, our approach is based on artificial injections of perturbations into the data sets, and is evaluated on state-of-the-art classifiers (both Machine-Learning and Deep-Learning) and data sets (in particular, public sensors-based univariate time series). First, we propose a coarse-grained study, with two artificial perturbations - called swapping effect and dropping effect - in which simple random algorithms are used. This already highlights a great disparity of the models’ robustness under such perturbations that industrial actors need to be aware of. Second, we propose a fine-grained study where instead of testing randomly some parameters' values, we used Genetic Algorithms to look for the models' limits. To do so, we define our multi-objectives optimisation problem with a fitness function as: maximising the impact of the perturbations (i.e. decreasing the most the model's accuracy), while minimising the changes in the time-series (with regards to our two parameters). This can be seen as an adversarial case, where the goal is not to exploit these weaknesses in a malicious way but to be aware of. Based on such a study, methods for making more robust the model and/or for observing such behaviour on the infrastructure could be investigated and implemented if needed. The tool developed in this latter study is therefore ready for being used in a real industrial case, where data sets and perturbations can now be fitted to the scenario.
