State space neural networks and model-decomposition methods for fault diagnosis of complex industrial systems

Abstract

Reliable and timely fault detection and isolation are necessary tasks to guarantee continuous performance in complex industrial systems, avoiding failure propagation in the system and helping to minimize downtime. Model-based diagnosis fulfils those requirements, and has the additional advantage of using reusable models. However, reusing existing complex non-linear models for diagnosis in large industrial systems is not straightforward. Most of the times, the models have been created for other purposes different from diagnosis, and many times the required analytical redundancy is small. The approach proposed in this work combines techniques from two different research communities within Artificial Intelligence: Model-based Reasoning and Neural Networks. In particular, in this work we propose to use Possible Conflicts, which is a model decomposition technique from the Artificial Intelligence community to provide the structure (equations, inputs, outputs, and state variables) of minimal models able to perform fault detection and isolation. Such structural information is then used to design a grey box model by means of state space neural networks. In this work we prove that the structure of the Minimal Evaluable Model for a Possible Conflict can be used in real-world industrial systems to guide the design of the state space model of the neural network, reducing its complexity and avoiding the process of multiple unknown parameter estimation in the first principles models. We demonstrate the feasibility of the approach in an evaporator for a beet sugar factory using real data.