I have images that represent a fixed-length time-window of different serials. Serials have time-series of different size, so e.g. serial1 has length 30, serial2 length 110 and so on. I have multiple images for every serial. Images are obtained with Grahamian Angular Field. My problem is that i have an unbalanced dataset: 1 serial of disk failed every 100 of disk. In Google Colaboratory I don't have so much memory to train the network. I have a total 40.000 images (15 channels) for 800 serials, so 8 failed serials e 792 healthy ones. If I train the CNN with 50% healthy and 50% failed serials and, on validation set and test set I have 1% my performance are really bad. What should I do? Please, don't ask me to post code. It's too long. I need a theoretical answer.