| Diagnosing abnormal behavior of different severity and convenience effects in a real-time manner is of paramount importance for energy-intensive building appliances. Both industrial and residential sectors suffer from post-incident maintenance where undetected faults occur for several days until the total breakdown of the equipment. To generate the necessary data set, a simulative test bed from Energym initiative was considered, exploiting an already validated residential environment. In this work, a Convolutional Neural Network (CNN) model was considered for classifying non-intrusive, low-cost temperature sensor embeddings in 3 categories with different abnormal heat pump severity levels. The features considered available derived from indoor zones temperatures and the outdoor / ambient temperature of the building; omitting intentionally readings from more elaborate sensors e.g., power analyzers or energy meters. The trained CNN model was eventually able to achieve very high accuracy i.e., around 95%; ensuring its high operational reliability by consuming real-time 15 min sequential temperature embeddings. |
*** Title, author list and abstract as seen in the Camera-Ready version of the paper that was provided to Conference Committee. Small changes that may have occurred during processing by Springer may not appear in this window.