fault detection analytics

In this post on data science for all, let us explore the interesting idea of predicting machine failure or manufacturing process failure. This concept is not unlike any of the other highly successful applications of predictive analytics. In fraud detection for example, predictive machine learning algorithms first get trained on data to understand what constitutes fraudulent behavior and are then used to predict if new data (or incidents) are fraudulent.

Predicting mechanical failure is fundamentally no different from this. In fact, one can argue that because such problems are governed by laws of physics, training algorithms to recognize imminent failure should be lot easier than training them to recognize fraudsters, because people’s behavior, unlike mechanical components, is not subjected to rigid physical laws! In other words, the patterns are more repeatable and therefore easily recognizable – especially by an algorithm.

Unfortunately, according to one recent survey, the adoption of machine learning or predictive analytics for this application is very limited. Less than 10% of the respondents indicated that they are using analytics for this very real and very expensive problem. Rapid-I has been very active in promoting this technology within the manufacturing industry in Europe, and they offer a reasonable explanation for why this might be so. Manufacturing as an industry has been around for more than 100 years and the processes and technology has evolved on human ingenuity and intuition rather than artificially augmented intelligence. In other words, a machine operator, by virtue of logging hundered of hours watching and listening to a machine, can spot a problem based on his experience. So the approach seems to be “why do we need automation to do this job”?

There are a couple of problems with this thinking. One, algorithms can detect problems much faster than human operators can. Two, human senses cannot distinguish minor subleties such as small variations in temperature or frequency. By the time our senses can actually distinguish these variations, it is usually too late. For example, the recognizable high frequency whine of a drill might occur only a few seconds before it cracks. Finally, there are only a few channels a human mind can process simultaneously. Where as an algorithm can process hundreds (if not thousands) of distinct sensor data at a time.

The good news is that, the same survey mentioned above, found that 33% respondents are planning to use predictive analytics for this very critical and hugely beneficial “data science for all” class of applications. In a future article, we will quickly highlight a case study for application. 

Originally posted on Thu, Mar 29, 2012 @ 09:05 AM

No responses yet

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

simafore.ai