Minimise automotive defects with big data analytics

A recent report by Stout, Risius and Ross shows a surge of electrical/electronic recalls over the last 10 years ending in 2014. The report shows that Early Warning Reports (EWRs) based on carmakers' reported field incidents have risen from 2004 to 2014. EWRs are safety incidents involving injury or death.

The alarming trend in the SRR report requires more powerful tools for defect detection in manufacturing in order to prevent field failures. More importantly, automotive suppliers must implement risk mitigating tools during the design and development process.

Advanced analytics employed to find patterns in big data have delivered value to other industries. But while manufacturing has experienced one of the fastest growth of data among various industries, much of the data has been under-utilised. There is tremendous potential in using this data to combat the rise in electronic field failures.

So what techniques are effective in reducing field failures?

Machine learning, both supervised and unsupervised, and anomaly detection are suitable for this application.

Increasing complexity in devices lead to high dimensionality, a condition where the design and manufacturing have grown to thousands or tens of thousands of variables. Machine learning models are well suited to finding the critical few variables that result to low field failure rates.

In contrast, when conventional statistical techniques are used for this purpose, the resulting models are less accurate, a condition called "curse of dimensionality".

Another problem that results from increased manufacturing and device complexity is the difficulty in formulating hypothesis for design of experiments.

The number of process corners in advanced semiconductor technologies such as the FinFETs have exploded from the traditional four process corners of yesterday to dozens.

So how does machine learning alleviate this issue? Unsupervised learning techniques that discover previously unknown relationships between variables or transforms of variables can provide the solution.

Whereas, conventional statistics requires the engineer to start with the hypothesis in a deductive approach, machine learning techniques can discover new hypotheses, an inductive approach. When the unsupervised learning algorithm is ran through the combination of process and device variables, the engineer may discover models that cannot be discovered any other way.