Missclassfication of Hand Generated Signals

I have two types of time series accelerometer data from two kinds of machines - one is very fast (Type A) and another one is relatively slow in terms of number of peaks/movement (Type B). I have tried various models on the RMS values of the x y z values - like time series KNN, RISE, and even a decision tree based on the features extracted from these two types of signals. These models give fairly good results too.

Now the issue is when some random hand-generated signals get tested by these models they get classified as Type B - which I think is due to similar feature values. But it is wrong.

Hand generated signals have no pattern or consistency that I can generate and put them in a separate class and create a multiclass model. I did try this but the accuracy of my model dropped drastically. I need a way to reject such random signals from getting classified.

If someone can suggest some approach to maybe filter out these signals with or without using Machine Learning, it would be really helpful.

For context I will share images of data -

Type A (fast signals)

Type B (comparatively slow signals)

Random Signal