Predictive Maintenance for Supermarket Refrigeration Systems Using Only Case Temperature Data

We present a machine-learning based approach for early detection of issues emerging in refrigeration and cold-storage systems that has the following desirable features: 1) Minimal sensor dependencies: only requires temperature readings and defrost state from the refrigeration cases 2) high precision, and 3) high generalizability of the learnt model. We achieve this by casting the time-series prediction problem as a classification problem, wherein we craft a set of features that capture key time-series characteristics specific to defrost and operating regimes. Our feature extraction employs seasonality-trend decomposition and pattern learning using dynamic time warping and clustering. The extracted features are used to learn a random forest-based binary classifier that can indicate the presence or absence of an issue in any given refrigeration case at any given time. We validate our approach on real data from 2265 refrigeration cases from several large supermarkets. The approach achieves a precision of 89%, lead time of approximately seven days, and a recall of 46% when evaluated on unseen cases.