Level-based unsupervised anomaly detection for industrial thermocouple sensor data in stainless steel furnaces

In stainless steel manufacturing, monitoring accurate temperature is critical for maintaining product quality and process efficiency. Thermocouples which are placed across furnace zones generate sensor data that must be continuously analyzed for early detection of process deviations. This thesis presents a level-based, unsupervised anomaly detection system developed in collaboration with Outokumpu, the largest producer of stainless steel in Europe and the second largest producer in the Americas, which aims to identify anomalies in thermocouple and flow sensor data without having labeled training examples. The proposed system operates across multiple levels: detecting anomalies at the overall furnace level, narrowing down to specific furnace zones, and finally identifying unusual patterns at the individual thermocouple level. This layered approach allows for both broad anomaly detection and detailed diagnostics, which helps to identify the location and potential root cause of abnormal behavior. This system provides enhanced operational awareness in industrial furnace operations by enabling more targeted investigation of deviations and faster response times. This capability helps with proactive maintenance strategies. The methodology is scalable and data-driven, which improves reliability and efficiency in continuous steel processing.