An Order-Invariant Transformer-Based TSK Fault Diagnosis Model with Adaptive Neural Inference Engine

Existing Takagi–Sugeno–Kang (TSK)-based fault diagnosis (FD) methods widely suffer from insufficient feature interaction capability in the antecedent, vanishing fault-related activation degrees, and limited representation capability in the consequent. To address these issues, this paper proposes an order-invariant Transformer-based TSK with an adaptive neural inference engine (OIT-TSK-ANIE) model, aiming to improve FD performance. The model first employs an order-invariant Transformer encoder unit to learn temporal deep fault features while reducing the influence of feature ordering. Subsequently, the TSK with an adaptive neural inference engine unit—employing a neuralized rule-firing generation mechanism to enhance feature interactions and avoid the attenuation of fault-related rule activations, while using deep fault features as consequent inputs—produces the final FD result. A series of comprehensive experiments demonstrate that the proposed model achieves at least 11.49% and 0.76% improvements in fault recognition accuracy over state-of-the-art deep learning and TSK-based methods on the Tennessee Eastman dataset and the real-world three-phase flow dataset, respectively. The model represents a novel and generalizable TSK-based FD paradigm with significant practical value.