🤖 AI Summary
In industrial settings (e.g., mining, manufacturing), the absence of structured event logs and the semantic ambiguity of raw sensor data hinder process mining. To address this, we propose an AI-driven four-stage pipeline: (1) preprocessing and unsupervised clustering of raw IoT sensor data; (2) domain-aware large language model (LLM) integration with structured prompting to automatically generate semantically rich, business-interpretable activity labels; (3) construction of XES-compliant high-level event logs. To objectively evaluate label quality, we introduce a similarity-weighted accuracy metric. Experiments on real-world LHD mining equipment sensor data demonstrate significant improvements in label consistency and log usability, effectively bridging the semantic gap between low-level sensor data and process mining requirements. Our approach enables robust, interpretable, and standards-compliant event log generation without manual annotation or domain-specific rule engineering.
📝 Abstract
This paper presents IoT Miner, a novel framework for automatically creating high-level event logs from raw industrial sensor data to support process mining. In many real-world settings, such as mining or manufacturing, standard event logs are unavailable, and sensor data lacks the structure and semantics needed for analysis. IoT Miner addresses this gap using a four-stage pipeline: data preprocessing, unsupervised clustering, large language model (LLM)-based labeling, and event log construction. A key innovation is the use of LLMs to generate meaningful activity labels from cluster statistics, guided by domain-specific prompts. We evaluate the approach on sensor data from a Load-Haul-Dump (LHD) mining machine and introduce a new metric, Similarity-Weighted Accuracy, to assess labeling quality. Results show that richer prompts lead to more accurate and consistent labels. By combining AI with domain-aware data processing, IoT Miner offers a scalable and interpretable method for generating event logs from IoT data, enabling process mining in settings where traditional logs are missing.