Community Conundrum 25:Feature Visualization is now live! Read More

Defect Detection - Watch on Demand

Community Manager
Community Manager
6 2 774

Watch @AaronCrouch (Data Analytics Manager, Atlantic Plant Maintenance) present his Dataiku DSS project aimed at preventing human injuries!

Key takeaways:

  • Measure success to convince business stakeholders,
  • Start with a True/False scenario, e.g. whether a job will have a defect rather than how many,
  • Consult legal on possibly protected information,
  • Use Partial Dependence to drive business decisions.


Presentation abstract:

Aaron Crouch (Data Analytics Manager, Atlantic Plant Maintenance) will share learnings from developing a project in Dataiku DSS to flag jobs where a safety or quality defect is likely to occur. He will walk you through how he and his team joined together data from various sources, and built a machine learning algorithm to predict injuries before they occur.

The model has an 85% accuracy in predicting which jobs will have an incident before the job starts. Outcomes include an intervention plan for job sites deemed high risk, and the team is now working on how to measure effectiveness of this plan.

Aaron Profile.png

Aaron Crouch is the Data Analytics Manager for Atlantic Plant Maintenance (a fully owned affiliate of GE). He has been working with data for over 15 years, and with Dataiku for a year and a half.

 

 

PS: if you're interested in presenting your Dataiku project at a future event, please let us know!

2 Comments
Community Manager
Community Manager

If you are interested in this event and looking for some resources to help get you started look no further: Getting started (or advanced!) with anomaly detection

Community Manager
Community Manager

At our second online event,  @AaronCrouch  presented his Dataiku DSS project aimed at detecting defects to protect workers' safety (ICYMI, here's the recording!). 

The key challenges that emerged from the follow-up discussion revolve around going beyond the predictions to drive business decisions. Once an anomaly has been identified, how to build and optimize the intervention plan?

That’s where second-order analytics come in. In Dataiku's guidebook on predictive maintenance, a few questions are proposed to help inform decisions:

pm-questions.png

Taking example on a truck from a large fleet with a part predicted to be defectuous. Once identified, it needs to be followed by a secondary report to the maintenance team, detailing the best possible options for time and place of intervention. 

Have you encountered similar challenges? Any best practices or common pitfalls which might help fellow data practitioners?