Self-Powered IIoT Steam-Trap Condition Monitoring Using Thermoelectric Energy Harvesting, Multimodal Sensing, and Machine Learning

Industrial steam-trap monitoring requires informative sensor measurements and reliable interpretation under practical power and communication constraints. This study examines a window-level machine learning workflow built on a previously reported thermoelectrically powered Industrial Internet of Things (IIoT) platform combining temperature and ultrasonic sensing with LoRaWAN communication. The dataset comprises 6354 windows from 2 monitoring-device series at 1 pharmaceutical facility, with 117 candidate engineered features. Six sensor-rule-informed pseudo-label categories are considered: normal, failed-open, failed-closed, degraded, excessive cycling, and short cycling. Random forest, extra trees, and histogram-based gradient boosting are compared using feature-exclusion configurations and a reported time-block evaluation. Histogram-based gradient boosting achieves reported test accuracy of 0.9853 and macro-F1 of 0.8822 with 108 retained features, and test macro-F1 decreases to 0.4999 under stricter exclusion retaining 81 features. These scores measure agreement with engineering labels rather than independently verified physical fault conditions. Normal and failed-closed are reported absent from the validation and test blocks, preventing complete six-class evaluation, and the recovered partitioning procedure does not establish global chronological or device-disjoint separation. A runtime prototype demonstrates registry-based eligibility control and visualisation of predictions and sensor context. The findings support a restricted pseudo-label classification workflow. Independent fault validation, cross-device evaluation, and measurement of the complete sensing and communication energy budget remain necessary to establish diagnostic validity and sustained energy-autonomous operation.