Making maintenance notes useful for repeat-event reports

Similar events under different descriptions

A maintenance manager wants to see which problems keep coming back. In the first work order a technician wrote: „Sensor dirty, works correctly after cleaning”. In the second, on the same machine: „Dirt on the sensor, cleaned”. It is the same observation, so the model puts both work orders in one group and keeps the original notes with them.

The third work order says only „Sensor replaced, machine running”. This record describes the work done but gives no cause. The model does not add dirt as a guess: the work order goes into the group of actions with no established cause. The manager sees where an observation recurs and where the documentation describes only the repair.

Ready maintenance systems split event information into codes. IBM Maximo describes a failure class hierarchy with the problem, the cause and the remedy on the work order. If technicians already choose such codes, a report is enough.

What a simpler change gives, and what training adds

In this count, a ready model with a list of groups gives most of the effect for a fraction of the cost. Training makes sense at a larger scale or when the remaining errors are expensive.

The set of training examples has to be maintained: when the process, the naming or the base model changes, the evaluation is repeated on the same test set. We put that cost in the quote together with the training itself.

What would the model learn?

First check a ready model with a clearly described split of events. If it sorts the notes correctly, training is unnecessary. Fine-tuning on examples with correct assignments helps it recognise plant abbreviations and equipment names when the model regularly misses them. Google describes supervised tuning of Gemini models on such examples. A change in how work is described, or new event groups, need a new review, so someone has to look after what the labels mean.

Start with one maintenance report

We start with the report that today needs the notes read by hand before a meeting. We adapt models as part of our private LLMs, and we build the app where an employee sees the result next to its source as part of our AI apps. For the free analysis, a few examples of an error that keeps coming back and a note of who corrects it today are enough.