When a machine stops, can your newest technician find what your best one knows?

A machine stops on the night shift, and the technician on duty can either page through a thick manual or phone the senior colleague who fixed that fault before. Store that colleague's knowledge as fault records your engineers have checked, and the technician gets the checked answer and its source on the tablet in front of them.

What Rockwell built in Singapore

Since October 2025, technicians at Rockwell Automation's Singapore site have used an in-house maintenance copilot. It searches a database built by veteran engineers with 20 to 30 years each on the shop floor, organised by symptom, cause and reaction, alongside manufacturing-software data, manuals, colleagues' notes and occurrences on other lines. Plant director Li Wang describes the target as “the loss of knowledge, what we call the tribal knowledge”, plus one consistent troubleshooting approach whoever is on shift.

Wang reports that machine downtime fell 33 percent; Rockwell's internal tracking shows servicing and spare-parts costs down by about 25 percent. Chief supply chain officer Bob Buttermore estimates that learning to troubleshoot now takes three months instead of nine. These are Rockwell's own figures from a showcase plant that also runs other AI systems, so read them as a direction, not a forecast.

Why checked fault knowledge is the asset to build

The lasting value sits in the checked records of symptom, cause and fix; the chat interface on top can change. A telecoms benchmark accepted at KDD 2026 found that eight leading models reached 90% accuracy on recognising intent and extracting details, but about 30% on procedural tasks such as generating the fix. It did not test factory maintenance, but our reading for your plant is clear: let the model interpret the technician's question, and let checked records supply the fix.

An ontology connects those records. It is a shared model of your business: the things you run and how they relate. For maintenance, it links machine models, fault codes, symptoms, causes, checked fixes, manual sections, production lines and the engineer who checked each record.

What a connected fault model gives you

Newer technicians spend less time looking for someone who knows. Rockwell engineering assistant Mangleswaran Mahalingam needed more than a year before he felt confident on night shifts, when fewer colleagues are on duty. Connected records put the senior engineer's checked answer on that shift.

Your existing software earns more. Tickets stay in your maintenance system and manuals stay as documents; with a connector to each, the ontology links them so one question draws on both. Because each machine and fault is defined once, later automation, such as predictive alerts or an agent that drafts work orders, reuses the same records.

FAIS builds this layer. We model your fault knowledge in your company ontology, connect it to your maintenance records and manuals, and build an assistant that cites the record or manual section behind each step and marks anything not yet checked.

What one fault record should hold

We propose extending Rockwell's symptom, cause and reaction structure into eight fields: machine family and model; fault code as shown on the controller; symptom; likely causes, ranked; the checks that tell those causes apart; the fix that worked; the engineer who checked it, with the date; and links to the manual section, line and past tickets.

Consider a Taiwan manufacturer running one family of injection-moulding machines on two lines. Before: a night-shift technician sees a recurring fault, pages through the manual, then messages a senior engineer who half-remembers a fix from the other line. After: the assistant shows the ranked causes, the check that separates a sensor fault from a mechanical one, the fix recorded on the other line, who checked it and the manual page. The technician still decides and does the repair.

Keep safety steps with the approved procedure

The assistant should link to lockout and energised-work procedures, not write them. In Taiwan, the amended Article 57 of the Occupational Safety and Health Facilities Rules takes effect on 1 January 2027. Where repair or adjustment could endanger workers, employers must stop the machine and its feed, use locks and signs against accidental start-up, and work to a hazard-prevention plan that includes a work-permit procedure. Link each fault record to that plan.

Start with one machine family and test on past tickets

Pull the most frequent or disruptive fault codes for one machine family from your maintenance history. Interview experienced technicians about real closed tickets rather than faults recalled from memory; IBM describes recording senior technicians as they narrate complex repairs, which gives these interviews good material. Write each code into the record and link it to machines, lines and manuals.

Before anyone relies on the assistant, replay past tickets through it and have a technician review every answer: did it reach the fix that worked, cite the right source and flag unchecked material?

A realistic goal for this quarter is one machine family, its main fault codes recorded, checked, linked and tested. That knowledge stays yours for any assistant you adopt later. If your best troubleshooting still lives in a few senior engineers' heads, that is worth discussing with FAIS: which machine family to model first, and how to connect it to the systems you already run.