drive machine learning to train data

  • is their data going to be a good predictor to roadmap failures?

generators ended up being way too difficult

  • only runs for 30 mins with super long idle time

  • analyze idle time and run time when it’s generating on the load to see if there were any coorelations with the data to figure out if we could build a goo dmodel

  • 30 mins is not good enough

  • potential multi-step approach to look outside of just the raw data

switchgear

  • much longer run times so much more data

  • very few fail instances for switchgear assets

  • challenging to get good predictive models from their data

even 1 out of 20 failure predictions is worth it - but if it caused more false positives then it creates more work - need to strike the right balance

no point in rushing the machine learning pieces

  • getting volumes of data from goole was challenging

  • they didn’t instrument their assets

  • data quality and governance issues occurred

got the non-ML products looking at health from condition-based data: less risk

  • pattern: if you have a lot of corrective maintenance on equipment > the asset is likely unhealthy

  • samples from assets (oil contaminants in an oil transformer)

go-live date on 300 generators in October

  • once you generate a score in MAS APM, how do you close the loop on service requests and work orders

  • connect both systems to consistently

    • send data from Maximo to our APM and back

      • work order data, thresholds