Explainable Deep Fuzzy Systems Applied to Sulfur Recovery Unit

The adoption of Machine LearningMachine learning (ML) in the Smart IndustrySmart industry concept is desirable to meet the requirements of reliability and explainabilityExplainability, guaranteeing safer models for the ever-growing number of complex industrial process systems. The development of ML models, particularly those based on Deep LearningDeep learning (DL) techniques, has made comprehensible communication with human operators increasingly challenging. The literature on eXplainable Artificial Intelligence (XAI) attempts to increase the degree of understanding by presenting techniques that aim to explain the results achieved by pre-trained models based on post-hoc strategies. To achieve explanations that are intrinsic to the model’s structure, this work explores the use of fuzzy systems that exploit the universe of discourse of the data, organising it into multiple parts associated with fuzzy rules (information granulesInformation granules). Specifically, this chapter discusses the hybridisation of multiple univariate additive zero-order Takagi-Sugeno (T-S) fuzzy systems with Long Short-Term Memory (LSTM) and Autoencoder, namely NFN-LSTM and NFN-AE, respectively, seeking an effective balance between interpretabilityInterpretability and accuracy. Evaluation is performed on representative datasets collected from the sulfur recovery unit, a real-world industrial system, to assess the quality of estimation and to shed light on the interpretabilityInterpretability of the input variables and their impact on the system. The experimental results demonstrate superior accuracy of the NFN-LSTM model over the selected methodologies for comparison. On the other hand, the NFN-AE model exhibits a favourable trade-off between accuracy and interpretabilityInterpretability compared to the NFN-LSTM model, requiring fewer parameters and computational load.