I have worked extensively with machine learning and deep learning. Since you mentioned that you have 10000 separate traces, that means a lot of data, which is perfect for building a robust model.
I have two algorithms in mind
1. Using an entire trace as a sequence of signal strengths and predicting at which sequence the anomaly exists, using Recurrent Neural Networks (LSTM's to be exact). (Kind of a classification task)
2. Using the entire image of the trace as it is, and training a Convolutional Neural Network over it. (More like a regression task.)
I have a question. What kind of an accuracy are you expecting? In terms of Mean absolute error between predicted value and the actual value?