Issue link: https://iconnect007.uberflip.com/i/1275152
AUGUST 2020 I SMT007 MAGAZINE 81 In an over-simplified ML model development process, datasets called training data are fed into the model to help the machine "learn" the patterns and trends. Then, the model is tested with real-life data and measured by a variety of methods, such as: • F1 score [3] • Mathews correlation coefficient [4] • Weighted average precision [5] • Weighted average recall [6] One important factor that may not be obvi- ous was that the predictive confidence of the ML algorithms is highly dependent on: • The amount of training data available (the bigger, the better) • How focused the data is relevant to the ML output expectations • The fusion of different ML and statistical techniques based on application expertise and experience During the algorithm testing process, we had to make significant changes to the fundamen- tal anomaly detection algorithm to achieve the scores we needed to give confidence to our cus- tomers that the anomalies were relevant and important. This customized algorithm is now a trade secret and cannot be shared here. However, you can use popular open-source algorithms, such as local outlier factor (LOF) [7] or autoen- coder neural network (ANN) [8] , and start there. Now that we established a fundamental core prediction, which is the anomaly detection, we then had to correlate and classify these anom- alies with potential causes and manufacturing process issues. This is the only way for oper- ators and technicians to take the appropriate action to avoid costly downtime or bad prod- uct escapees. Figure 5 shows how we had grouped the dif- ferent types of anomalies by the same patterns and characteristics. Each of these patterns indicates a different potential cause and effect from the manufacturing process. For example, when a degradation anomaly (Figure 6) is pre- dicted, the customer may investigate the par- ticular test probe associated with that anom- aly. Through our comprehensive testing, we have discovered that a degradation pattern is highly correlated with wear and tear of test probes in a manufacturing test process. Other anomalies that are not correlated to the wear and tear of test probes do not usually degrade over time, but rather occur suddenly in steps or shifts (Figure 7). Figure 4: Example of a measurement anomaly.