This paper proposes a multi-agent-based methodology for real-time fault diagnosis in industrial processes. This work aims to build a decision support tool that helps process operators identify and better manage abnormal situations. Supervised and semi-supervised machine learning methods are widely used to develop such tools. Despite their accuracy in classifying faults, supervised methods have a major limitation: they cannot diagnose novel faults. Semi-supervised methods can detect and isolate novel faults, but cannot disclose their root causes. The proposed methodology combines both supervised and semi-supervised methods in a parallel-serial structure, exploiting their respective strengths. Moreover, it provides the process expert with meaningful explanations of detected novel faults or otherwise. Two case studies are used in this paper to demonstrate the effectiveness of the proposed methodology. The first case is the Tennessee Eastman Process benchmark. The second one uses real data collected from a heat recovery system in a thermomechanical pulp mill.

Data-driven prognostics and health management solutions play an important role for condition monitoring of industrial systems. They offer promising perspectives for high reliability, availability, maintainability, and safety production, especially in Industry 4.0 where model-based approaches are not suitable for complex systems. To build an efficient diagnostic and prognostic model, the collected sensor measurements are injected into processing algorithms to extract features and construct health indicators. The performance of this procedure strictly depends on the data quality. However, the availability of reliable and exploitable data remains one of the most challenging issues in condition monitoring. This paper presents advanced data analytics to construct prognostic health indicator. This method aims at obtaining reliable and exploitable indicators that represent multiple degradations evolution of the system. It consists of an automated processing algorithm that detect and isolate trajectories representing the system health state from its nominal condition to its critical level. The performance of the proposed method is investigated through a real data of pulp and paper manufacturing where the degradation process is an evolution of fouling in a black liquor heat exchanger.