PRISMA: Physically-Aware Reasoning and Intelligent Semantic Mining Architecture for IoT Process Discovery Using Deep Learning and Large Language Models

Internet of Things (IoT) environments generate large volumes of heterogeneous sensor data that remain difficult to exploit for process mining due to the absence of structured event logs with explicit case identifiers, activity labels, and semantic annotations. Bridging the gap between continuous sensor streams and process-level representations remains a major challenge for IoT-aware business process analysis. This paper introduces PRISMA (Physically-aware Reasoning and Intelligent Semantic Mining Architecture), an end-to-end framework for automatically discovering physically realistic business processes from raw IoT sensor data. PRISMA integrates three complementary components: (i) a neural transformation pipeline that converts heterogeneous sensor streams into structured event logs through adaptive parsing, temporal segmentation, activity recognition, and case identification; (ii) a semantic enrichment engine that augments events with IoT-aware business semantics using domain ontologies and neural language models; and (iii) a physics-aware discovery stage that enforces temporal, causal, and resource constraints during process model construction. PRISMA is evaluated on three real-world IoT datasets comprising 7.2 million events collected from 270 sensors across different application domains. Experimental results demonstrate that PRISMA achieves X% completeness, Y% semantic accuracy, and Z violations, outperforming state-of-the-art baselines. These results show that PRISMA provides apromisingunified approach for end-to-end IoT-aware process mining, withfuture work needed to address deployment challengessuch as streaming scalability and concept drift.