Explainable AI & Diagnostic Modeling – Root Cause Analysis Of Failures in Machine Engineering

This case study provides an overview of how an explainable machine learning model can uncover critical unexplained reasons for trucks failures for an equipment manufacturer.

• Data Science
• Decision Trees
• SVDD
• Dashboards
• Analytics Platform
• Explainable AI (XAI)
• Root Cause Analysis

Context & Key challenges

Our client is a German-Swiss multinational equipment manufacturer that produces, among others, mining trucks.

In this context, it is very important for their customers that their trucks do not break down in the mines where they are deployed, which would cause a complete shutdown of their activity (considering that a truck can block a whole mine).

Since the mining trucks are based on a central electronic system, our client was able to collect lots of data on each failure and was looking to identify the causes of the failures in order to prevent them.

Our Approach

In order to meet our client’s expectations, we used a RaaS (Result-as-a-Service) approach: our client provided us with its data (250 sensors, 5-10 trucks over 10 years, 20-30 electronic failures examples) and we integrated it directly into our own SAS-based analytics platform.

Our approach was to identify unusual events that took place right before a truck failed. We have thus:

• Aggregated and classified historical data into comparable analysis format of good and bad behaviors
• Compared the differences between the behaviors to identify events common to breakdowns
• Built an explainable machine learning model highlighting critical reasons for failures
• Provided our client with an interactive dashboard showing the results and the model used for communicating and discussing with business teams and management
• It turned out that two-thirds of the breakdowns were simply due to drivers pushing the gas pedal with the handbrake on at the same time.

Benefits

• Centralized view of available data highlighting simple and unsuspected links between the information. In this case, the data was generated by two separate systems that were themselves managed by two different departments of the company. The conclusion we came to was therefore impossible to reach without the holistic view that we built.
• Identification and explanation of about 2/3 of all failure patterns of the past years.
• Potential savings in the range of more than 100K euros, by simple elimination of the most critical failure patterns.
• High acceptance of the machine learning model by the business thanks to the interactive and explainable dashboard we provided.

Team involved

An industry Expert, a Data Scientist and a Data Engineer collaborated during 4 months on this project.

Technologies & Partners

Share

Thanks to our
Author

Nicola Z.