Rate-Normalised Operating Signatures for Scalable Condition Monitoring of Commercial Refrigeration

Commercial refrigeration rarely fails without warning. A unit with a refrigerant leak or a faulty defrost heater can go on holding temperature for months while its compressor works towards failure and its energy use multiplies. Reactive maintenance only responds once the unit has broken down, and calendar-based maintenance will usually miss the problem altogether. Across an estate the condition is hard to detect, because healthy appliances run at different power levels, set points and cycle timings. This paper describes a rate-normalised operating signature that brings out the behaviour they have in common. Each minute's temperature change is paired with the electrical power drawn in that minute, and the time order is then discarded. Healthy appliances show a low-power warming mode and a higher-power cooling mode, and deterioration shifts the relative power of these modes, the cooling response, and the time spent in each. The method grew out of NoWatt's commercial monitoring and condition-directed maintenance work. Across three operators and 50 sites, emergency call-outs dropped from 24 to 1 over two years after maintenance began to follow analysts' assessments of which units were failing. Acting early saves months of wasted energy (one repaired freezer used a third of the electricity while running colder), causes less disruption to the business, and allows repairs to be planned and cheaper. The aim here is to move that assessment from the analysts to a network. Six walk-in appliances, with 2,625 labelled 12-hour periods between them, were used in a retrospective test in which each appliance was held out in turn. Given five features describing each period's operating modes, and trained only on the other appliances, the rate-normalised network detected 75.6% of unhealthy periods and flagged 16.7% of healthy ones. The direct-input network, which was given the power and temperature readings in time order and trained on the same appliances, detected 85.1% but flagged 37.9% of healthy periods, more than twice the false-alarm rate. With the threshold set to give one false alarm in ten healthy periods, the rate-normalised network detected 68% of unhealthy periods against 40% for the direct-input network. The features have a direct physical meaning, and a new appliance can be assessed knowing only its class, without any labelled history. Historical records, analysis workbooks, code and test outputs are archived alongside the paper.