Hello!

I currently working on a project fo estimate the knee angle based on IMU data (IMU data consist of 9 sensor values).

I have a large dataset (100 hours walking data from 20 persons) and manually anotated the knee angle using a tracking software.

Now I want to build a machine learning model which takes the IMU data and estimates the knee angle at each time point.

I am new into machine learning and struggling how to start.

I thought about creating an input matrix like follows: (For training of the model)

• Actual IMU sensor values + 19 vales from the past
• As an expected value the knee angle from the actual step is used

This would result in overlapping windows with time step 1, each angle estimation is based on the sensor data from 19 past time steps.

Now the model part:

1. I thought of building a basic CNN regression model which takes the input matrix (20 time steps x IMU sensors) and estimates the corresponding knee angle

2. Build an LSTM model which estimates the knee angle and feeds this output back as an input for the next time step (together with IMU data)

My thoughts:

Time step 1: Input (20 x sensors) -> LSTM -> (Knee angle in degrees)

Time step 2: Input Knee Angle from t1 + sensordata -> LSTM -> Knee angle

Time step 3: ...

With this appoach I hope that the model can follow the natural change of the knee angle whereelse the CNN just predicts any number. Cause the knee angle cannot change from 60 degrees to 0 within one time step and I know so far that an LSTM can learn long term dependencies, so it knows when knee_angle (t_n) = 60 knee_angle(t_n+1) must be in the range of 60+-10.

In the picture I showed how the input data looks like and the corresponding knee angle.

If it is simplier to build a classification model it is also good for me.

Because I am mostly interested in the knee angle reversal (indexed in red in the picture)

Do you have some input for my problem?