Failure of refrigerated cabinets costs millions annually to supermarkets, and a large market exists for systems which can predict such failures. Previous work, now moving towards deployment, has used neural networks to predict volumes of alarms from refrigeration system controllers, and also to predict likely refrigerant gas loss. Here, we use in-cabinet temperature data, aiming to predict faults from the pattern of temperature over time. We argue that artificial immune systems (AIS) are particularly appropriate for this, and report a series of preliminary experiments which investigate parameter and strategy choices. We also investigate a 'differential' encoding scheme designed to highlight essential elements of in-cabinet temperature patterns. The results prove feasibility for AIS in this application, with good self-detection rates, and a promising fault-detection rate. The best configuration of those examined seems to be that which uses the novel differential encoding with r-bits matching.
展开▼
