Issue link: https://iconnect007.uberflip.com/i/1104607
22 PCB007 MAGAZINE I APRIL 2019 device picks up different inputs, such as irra- diance, temperature, voltage, and current. But what's really cool about it is that it can switch to- pologies by itself through the switching matrix. We've been using machine learning to help identify which faults are happening on the cur- rent-voltage (I-V) curve. With traditional clus- tering techniques, the actual data that we need is covered; the only thing that it will give us ac- curately is ground and arc faults while shading is difficult to identify. So, we need to monitor in a different way. Instead of an unsupervised technique, we need a supervised technique be- cause it's important to know the type of fault. Since we can't use the unsupervised tech- nique, instead, we can do a classification tech- nique where we have the input vector from the data that we collect from the smart monitoring device. And we create these prototypes mod- els ourselves that will tell you how similar the features are to the actual fault. For example, in arc and ground faults, these are the specific characteristics that occur rather than just look- ing at the drop in the voltage and current. The algorithm will output how similar it is to each model. Specifically, it compares the input vec- tor with each prototype, assesses output how similar it is, and then assigns a weight. This process will continue until the correct weights are found. Johnson: What are some applications that you perceive for this work? Sameeksha Katoch: Right now, what we're look- ing into is improving the power efficiency for the solar array. We have a small testbed (thanks to the NSF CPS project) with smart monitoring devices attached to each panel. Each panel can REU student poster competition by Emma Pedersen. Sameeksha Katoch