All Member Open Forum

Contextual Failure Analysis: A Text-based Search Engine for Failure Analysis

Respectfully submitted,
Jeff Wahl

• For those data elements which had an object as the target (asset and maintainable item) it is a Text Link Analysis library lookup which assigns scoring based on some rules and then suggests the most likely answer​. These models were also subsequently adapted to evaluate free-text material requests to identify which item masters should / could have been used, as well as requirements for new item master creation.
• For the other data elements we developed a hybrid model consisting of rules and decision trees. These models incorporate a number of structured data elements like lead craft in the learning mode which refines the accuracy of the models, however we found that the model improved best in iterations by adding more specific rules.

• For those data elements which had an object as the target (asset and maintainable item) it is a Text Link Analysis library lookup which assigns scoring based on some rules and then suggests the most likely answer​. These models were also subsequently adapted to evaluate free-text material requests to identify which item masters should / could have been used, as well as requirements for new item master creation.
• For the other data elements we developed a hybrid model consisting of rules and decision trees. These models incorporate a number of structured data elements like lead craft in the learning mode which refines the accuracy of the models, however we found that the model improved best in iterations by adding more specific rules.

• For those data elements which had an object as the target (asset and maintainable item) it is a Text Link Analysis library lookup which assigns scoring based on some rules and then suggests the most likely answer​. These models were also subsequently adapted to evaluate free-text material requests to identify which item masters should / could have been used, as well as requirements for new item master creation.
• For the other data elements we developed a hybrid model consisting of rules and decision trees. These models incorporate a number of structured data elements like lead craft in the learning mode which refines the accuracy of the models, however we found that the model improved best in iterations by adding more specific rules.

1. Equipment = Pump (e.g. P-1234)
2. Observed Problem = Low Output (Table B5 LOO)
3. Failed Maintainable Item = Impeller (Table A21)
4. Failure Mechanism = Wear (Table B2 2.4)

• Number 1 and 2 must be recorded when an operator requests work to be done.
• Number 3 and 4 must be recorded when a maintainer completes the repair or investigation.

Good morning,

I wonder how this discussion would have been different if all participants had had skills in building super tables from the many subtables at our avail from the same and different systems?

What ideas would we have discussed for recovering (teasing out) failure modes from our historical data? What ideas would we have discussed for confirming that what is recorded is an "honest" attempt? What ideas would we have discussed for converting recorded honestly attempted symptoms or modes to our ideal structure of modes? What ideas would we have discussed for timely auditing to confirm that policies for modes are being complied with and then timely confirming that any remediation has occurred?

I can only speak for my own-mind's eye. Decide for yourself after learning how to build super tables from attending the SMRP webinar titled, "Build Super Tables from Operational Data," on September 5^th. Then go forth and invent.

The agenda is as follows:

• Purpose of the session.
• Big picture.
• Extract, join and mold subtables into a super table.
  • Perspective.
  • Case 1: Build a super table of work orders, order tasks and craft hours.
  • Find and cleanse bad data.
• Build aggregation variables into a super table.
  • Perspective.
  • Case 2: Identify outlier work orders by Z-Score of craft hours grouped by cost center and work type.
  • Case 3: Extend Case 2 to include a created group-orders classified by lead craft.
• SQL perspective.
• On-line help and literature.

The webinar is free to SMRP members, otherwise $35. Non-members register at https://portal.smrp.org/eweb/DynamicPage.aspx?Site=SMRP&WebCode=EventDetail&evt_key=f4bbfffc-dab6-4ae5-b351-b4776de3c4a3