The Data That Was There Before the Sensors
Predictive maintenance usually brings to mind vibration sensors, temperature probes, and IoT gateways. Yet most plants have a data source that predates any sensor by decades: maintenance logs and work order notes. Entries like "Line 3 conveyor belt tracking off, adjusted tension, OK" or "Hydraulic pump noise, coupling worn, replaced" have been piling up for ten or twenty years.
The reason this data goes unused is simple. It's free text, so it can't be aggregated, and because everyone writes differently, searching rarely surfaces the right case. Teams end up relying on memory: "Mr. Kim fixed this exact symptom a few years ago." The problem is that those veteran technicians are retiring. When they leave, the records stay, but the people who knew how to read them are gone.
Start With the Quality of Maintenance Text
Before maintenance logs can be analyzed, you have to face their quality problems head-on.
This is where LLMs earn their keep. Feed the model a shop-floor glossary and the equipment master alongside each log, and have it extract structured fields: symptom, probable cause, action taken, parts replaced, and equipment code. Crucially, instruct it to mark fields as "unknown" when the source text doesn't support them. A dataset with honest blanks is far safer than one filled with plausible guesses.
Building a Failure Classification Scheme
Once fields are extracted, you need a coding scheme. Design hierarchical codes for symptoms (noise, leak, overheating, stoppage), causes (wear, loosening, contamination, misconfiguration), and actions (adjust, replace, clean, reset). Borrowing from ISO 14224, the reliability data standard used in the oil and petrochemical industries, saves you from inventing failure modes and causes from scratch.
With the scheme fixed, reclassify the full history with the LLM, then always run human review sampling. For example, have maintenance staff check 300 random records out of 30,000, measure accuracy per code, and strengthen the definitions and examples for weak codes before reclassifying. After two or three rounds, you'll know which codes you can actually trust.
Similar-Failure Search and Fix Recommendations
With structured history in place, you can build features technicians will actually use. When someone types "Press #2 abnormal noise on slide downstroke," the system first shows past cases on the same machine, then cases from similar equipment. Combining embedding-based semantic search with equipment and code filters finds the same failure even when it was described in different words.
The non-negotiable part is showing the evidence. Every "likely cause" the LLM summarizes should link back to the original log entries it came from, and the system should decline to recommend anything when no supporting case exists. On the maintenance floor, a confident answer without a source is worse than no answer. Trust builds when technicians can read the original record and make the call themselves.
Integrating With CMMS and MES
A standalone tool will end up unused. Similar cases and recommended fixes need to appear at the moment a work order is created in the CMMS. When the job closes, capture feedback in one or two clicks: was the recommendation right, and what was actually done? That feedback steadily improves the next recommendation.
Combine this with equipment downtime history from the MES and you can generate repeat-failure reports. Automatically flagging machines where the same cause code occurred three or more times within 90 days gives you hard evidence for moving them off the repair-again cycle and into preventive maintenance or improvement plans.
Rollout Phases and Measuring Impact
A phased rollout works best:
Measure impact with MTTR (mean time to repair) and the repeat-failure rate. For example, if the pilot line's MTTR drops from 120 to 90 minutes, that's a 25% reduction. Capture a baseline over the three to six months before rollout so the comparison holds up.
POLYGLOTSOFT combines hands-on experience building MES and equipment management systems with LLM-based RAG to cover the full journey: cleaning maintenance history, designing failure code schemes, recommending fixes from similar cases, and integrating with your CMMS and MES. If you're ready to turn years of maintenance logs into a lasting maintenance knowledge asset, get in touch with us today.