The FP-SENSING-AI1 package contains example code for classifying accelerometer using a neural network.
The example code runs directly on the SensorTile hardware, so that is the easiest way to get started. You can use the Sdcard or USB to a computer to collect your training data.
For your first model, you can try to reuse the basic HAR model included in FP-AI-SENSING1. In general, a CNN or recurrent (LSTM,GRU) model is commonly used for these kinds of problems.
Also check this nice tutorial with some example data, https://developer.ibm.com/tutorials/iot-deep-learning-anomaly-detection-5/
Machine Learning Engineer & Independent Consultant