SLDAE: An interpretable stacked Denoising Auto-Encoders for fan fault diagnosis on steelmaking workshops

Abstract

Fault diagnosis of the fan in steelmaking shops has significant practical value in guaranteeing smooth operation and achieving quality control of steel coils. The existing data-driven systems, mainly deep neural network-based diagnostic models, have achieved some success in recognizing fan faults. However, these models face a significant challenge in reaching reliable diagnostic conclusions with uncertainty due to the absence of interpretable representations between fault labels and features. To address this issue, this paper develops and evaluates an interpretable model, namely Stacked Logic Denoising Auto-Encoders (SLDAE). SLDAE is a flexible neural-symbolic system that extracts confidence rules and Binary Decision Logic (BDL) rules to explain how SDAE conducts feature learning, and conducts uncertain reasoning of diagnostic decision-making. To extract confidence rules, a logic grouping DAE is designed to consider the impact of different literals on neuron activation so as to reduce information loss. To extract BDL rules, we design a novel network structure, Discrete Logic Networks (DLNs), to facilitate extracting implicit relationships between fault features and labels while learning and representing the belief of BDL rules. Experiments verified on two fan datasets indicate that SLDAE can perform quantitative reasoning of uncertain diagnostic logic and exhibits a notable performance in fault recognition.