Issue link: https://iconnect007.uberflip.com/i/1509873
OCTOBER 2023 I PCB007 MAGAZINE 73 Machine Learning for Troubleshooting of Equipment and Processes Machine learning can be a valuable tool for troubleshooting and data-driven prob- lem identification in PCB manufacturing. For example, ML algorithms, such as clustering or statistical models, are used to identify anoma- lies or outliers in the data. ese anomalies can signal potential issues. ML techniques can also be applied to deter- mine the root causes of anomalies or prob- lems. is may involve correlation analysis, causal inference, or rule-based systems. Such systems would identify which factors or vari- ables are most strongly associated with the detected anomalies, and direct maintenance staff to the right equipment and even suggest a solution. In order to enable such tracing capabilities, each p a n e l m u s t c a r r y a unique identif ier, t yp - ically a 2D code (data matrix code, or DMC, is the most common type). is code will then be scanned at every produc- tion stage enabling the system to trace it back to every process equipment it went through dur- ing production process. Machine learning and AI tech- nology is also being gradually adopted within AOI and AFI/AVI processes. e ultimate goal is to reduce human operators' involve- ment in defects verification to the minimum, if not eliminate it altogether, thereby streamlin- ing the AOI process by making it faster and less prone to human errors. In this application, the AI image recogni- tion model is trained on an extensive set of real defects aiming at teaching it how to tell them apart from false calls such as non-critical con- tamination or acceptable process variations. In general, deep learning neural network AI is the most suitable technology for image recog- nition in general and optical inspection in par- ticular. Despite the promising potential, imple- menting ML for the AOI verification cycle is not without significant technological and conceptual challenges. e seemingly simple question, "What is a defect?", may not have the same answers between market segments, among different PCB manufacturers, or even individual jobs within the same company. Furthermore, defect specifications fre- quently change and evolve over time as a result of new processes, changing end user require- ments, and PCB makers' other various consid- erations. is means that one AI model may not be sufficient to cover all scenarios, and such models must be retrained on a regu- lar basis with new data that better reflects the current situation. For example, an AI system would be unable to accu- rately identify defects that it had never been trained on or effectively filter out a previously unknown type of false calls. Another challenge is determining the best tradeoff between two competing met- rics: the filter rate (the number of true false calls removed by AI) and escapes, also known as underkill or false negatives (real defects that were incorrectly classified and fil- tered out as false calls by the AI model). e higher the filter rate, the more defects will escape, and vice versa. e correct approach to resolving this problem is to take a purely eco- nomic approach, weighing the benefit of hav- ing less verification equipment and reduced manpower against the cost of lower yield due to higher escape rate. Finally, there is a type of agency dilemma where there is a conflict of interest between the PCB manufacturer's motivation to cut costs and the end user's interest for high quality. It should come as no surprise that the end user ML techniques can also be applied to determine the root causes of anomalies or problems.