Machine-learning fault classification

Machine-learning fault classification is a supervised method: a model is trained on labelled monitoring data and then sorts new observations into categories, such as normal operation or a particular type of fault. Common algorithms include decision trees, K nearest neighbors, logistic regression, neural networks and support vector machines. The data is divided into a training set and a separate test set, so the model's accuracy can be checked against data it has not seen before. In field use, the model learns the expected AC power or power-to-irradiance ratio from inverter and irradiance data and raises a fault when measured output stays outside that range. Reported examples detected partial shading, inverter faults, blown fuses, open string circuits and overheated inverters, and research also targets diagnosis of DC series arc faults. The same approach can classify I-V curves or EL cell images, for example sorting cells into good, cracked or corroded. Accuracy depends on how much training data there is: cracked cells were classified poorly because there were few examples and crack patterns vary a lot. False alarms still happen, so a flagged fault should be confirmed on site before any repair is made.

<!-- review notes

• sentence 8: support check: partial - The passage documents a false alarm, but it does not say a flagged fault should be confirmed on site before repair. -->