Issue link: https://iconnect007.uberflip.com/i/1275152
80 SMT007 MAGAZINE I AUGUST 2020 tems and measurements that define the qual- ity of the products. The solution had to address use cases where workflow automation and yield loss predic- tions were due to failures caused by fixtures, test program settings, and other factors that are not due to the actual parts used in the product. These were false negatives that increased pro- duction costs. At the core of the solution, a fundamental pre- diction it makes involves almost real-time anom- aly detection for every test in production. At any one product test, there are hundreds to thou- sands of such measurements and tests. On a typical day, a few hundred products go through this mandatory process. Some of these anoma- lies will be classified as degradation patterns that relate directly to specific parts of the fixtures. Simply put, anomalies are atypical to what is normal. This presents the first challenge for manufacturing and test. Current factories may capture and store high value and critical data such as failures. It was never imperative to capture good measurements, especially down to a single test. Therefore, to predict anom- alies, the factory must be able to capture all measurements that are performed—regardless of the nature of the result associated with that measurement. Figure 4 shows a screen capture of an actual anomaly that was predicted by the system. The component is an expected 10K Ohm resistor on a complicated electronic board. The green dots are measurements within the test limits or specifications, which are the dots in white. The anomaly is the pink dot and is measured at 9.429K Ohm. In production, this will not trigger any concern as the test "passes" and there is no capability of current processes to detect such a situation. However, it is evident that the manufactur- ing process is very capable of producing such stable measurements. The mean and median are standard deviations from the test limits. In such situations, it should be alarming that such a drastic anomaly happens in production regardless of whether it is within the test limits. Measurements such as these may shift sud- denly or over time, depending on many fac- tors. The anomaly detection algorithm must have the ability to learn the "norm" as fast as possible to avoid too many false calls. Detect- ing anomalies is the first step in the process of applying ML in manufacturing test. Figure 3: Cloud-based real-time advanced analytics architecture.