PMA Anomaly Detection Model: More Than Just off-the-Shelf ML Algorithms

Application Notes

Problem Statement

In a typical implementation of a big data advanced analytics platform in a smart factory, massive amounts of test and measurement data generated by manufacturing test systems on the floor are acquired, transformed and analyzed ideally in real-time to produce actionable insights. One of these insights would be test and measurement anomalies that may be a prediction of deteriorating product quality that may be due to parts, process or people. While the electronics manufacturing industry has just begun the digital transformation journey, the industry may be oblivious to the serious challenges from ‘alert fatigue’. However, the alert fatigue issue is common in other industries such as cyber-security, consumer and fintech applications that uses anomaly detection as a way to predict and produce insights. The number of alerts being sent every day can and will be overwhelming. In a case study Keysight performed using large sets of test and measurement data from a manufacturing line over an 18-day period, running through an array of different anomaly detection algorithms, these algorithms typically generated more than 1400 alerts. Such many alerts have given rise to fatigue on the user side - it is difficult for users to go through almost a hundred alerts per day. In addition, a large portion of the alerts are not severe or critical enough to require any action and would be considered false positives.