One technique you might use is something I call “classification by anomaly score”. If you have a set of sources that produce data streams, train a model for each source and get anomaly likelihoods out of them over time. Once you feel that they are fully trained and not throwing false positive anomalies, turn learning OFF on all the models.
Now you have a suite of models that can classify new data streams. Just pass the new data stream into each model simultaneously. Over time, track the model with the lowest anomaly likelihood, which is the resulting classification. You can also tell how confident the classification is based on the anomaly values across all the models.