Probabilistic Hierarchical Perception Networks for Reliable Intelligent Fault Diagnostics of Elevator Traction System

Abstract The rapid development of large language models (LLMs) has introduced a new paradigm for decision-making in mechanical system health management. However, due to the critical lack of interpretable data embeddings and uncertainty-aware, the direct application of LLMs in industrial settings remains ineffective. Therefore, the key to successful LLMs implementation lies in analyzing uncertainty in data embeddings to enhance reliability. This paper proposes a novel model called probabilistic hierarchical perception networks (PHPNs), which first leverages convolutional kernels to extract local features characterizing vibration semantics and then employs a Transformer backbone to capture their global semantic relationships. Furthermore, a probabilistic Bayesian approach is incorporated to enhance the model's perception of prediction uncertainty. In addition, a complete interpretable fault diagnosis framework is proposed to enable the model to automatically quantify and interpret sources of uncertainty when dealing with fault diagnosis tasks. By effectively separating aleatoric and epistemic uncertainty, the framework can identify the sources of unknown factors in the data, thereby improving the accuracy and confidence of the diagnosis. The experimental part is tested on an independently constructed structural fault dataset, which verifies the effectiveness and generalizability of the method proposed in this paper. Comparison results with existing methods show that PHPN can provide more comprehensive diagnostic results and better uncertainty separation performance. Meanwhile, the proposed method can also provide uncertainty analysis and risk assessment for LLMs in industrial health management.