Automated Fault Tree Learning from Continuous-valued Sensor Data

Many industrial sectors have been collecting big sensor data. With recent technologies for processing data, companies can exploit this automatic failure detection and prevention. We propose the first completely automated method analysis, machine-learning fault trees from raw observational data with continuous variables. Our scales well is tested on a real-world, five-year dataset of domestic heater operations in The Netherlands, 31 million unique heater-day readings, each containing 27 11 builds two previous procedures: C4.5 decision-tree learning algorithm, LIFT tree algorithm Boolean pre-processes variable: it learns an optimal numerical threshold which distinguishes between faulty normal operation top-level system. These thresholds discretise variables, thus allowing to learn model root mechanisms system are explainable. obtain evaluate them ways: quantitatively, significance score, qualitatively, domain specialists. Some learnt almost maximum (above 0.95), while others medium-to-low (around 0.30), reflecting difficulty big, noisy, real-world specialists confirm that meaningful relationships among