ReDBN: An Interpretable Deep Belief Network for Fan Fault Diagnosis in Iron and Steel Production Lines
Xiaoqiang Liao, Dong Wang, Siqi Qiu, Xinguo Ming
Abstract
Fan fault diagnosis in steelmaking production lines is critical for safety production and environmental protection. Deep neural networks (DNNs) achieve slight success for fan fault diagnosis. These models are unable to provide explanations for fan diagnostic decisions due to DNN's opaque structure. From the perspective of neural-symbolic integration, researchers gradually pay attention to how to extract relational knowledge from DNNs to provide an explainable representation of DNN's features learning and reasoning. This study introduces a neural–symbolic model, reverse deep belief network (ReDBN), where interpretable logic representations (CR-rules) are derived based on the integration of confidence and rough rules so as to tackle fan uncertain diagnosis decision-making. To extract confidence rules, a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i>-logic restricted Boltzmann machine (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">${\bm{k}}$</tex-math></inline-formula>-LRBM) is deployed and evaluated. In <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">${\bm{k}}$</tex-math></inline-formula>-LRBMs, confidence rules can be extracted by considering the effect of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i> different literal clusters on neuron's activation. Besides, this article introduces a symbolic language, termed rough rules, which can solve uncertain reasoning during fan fault diagnosis. Rough rules, assigning a belief value for attribute variables, can represent the probability of how likely the sample belongs to specific fault labels. Verified on two fan datasets from a fan testbed and a real production site in Shanghai, ReDBN can achieve better performance than other typical models.