Power inverters are critical equipment whose failure can drastically affect the operation of production lines and cause considerable financial losses. As a result, maintenance contracts are very strict, imposing very short recovery times on suppliers. These deadlines are all the more difficult to meet as the power inverters are spread all over the world. Because of this, the manufacturers try to avoid breakdowns at all costs, even if it means changing some parts ahead of time to prevent an outage.
The problem is therefore to predict when particular component could fail in order to schedule replacement at the most convenient time before any failure.
To do this, the probabilistic solution consists in following the evolution through time of the sensors present on the power inverters, and to compare this evolution to typical examples having led to failures, to infer the probabilities of failure. Here again, some useful information is missing. For example, the temperature of the accumulators is a key piece of information that cannot be measured directly but only estimated from indirect measurements. As in many of the other use-cases, the space of possibilities is huge and must be sampled.