I am dealing with a PHM RUL problem, a time series problem of machine signal. I consider to apply Fourier transform or Wavelet Transform to my sensor feature and train the LSTM model. But I have some problems when I using these transformed sensor data. my data looks like:
sensor 1 sensor 2 sensor 3 RUL 1 1.5 3 5 2 2.5 2 4 3 3.5 1 3 4 4.5 0 2 ... ... ... ...
First 3 column is my features(x), the last column is my RUL(y). The data have n rows which are time sequential. Every x mapping to one y.
I have some ideal to employ the transformed sensor data to my training model.
- Simply treat the transformed sensor data as a new feature.
- Concat the transformed sensor data after original time series data.
- Only feed the transformed sensor data to my training model.
- Create two LSTM model. The original data and transformed sensor data feed in their own model respectively, then combine the result by something like ensemble.
I can't make sure which one is more reasonable. Actually, all these ways will face a problem when I mapping my transformed sensor data(x) to my original RUL(y). That is, the transformed sensor data often have difference length(rows) from the original data. I can not mapping x to y if their length is different.
Any advise about how to apply the transformed data to LSTM model will be appreciated !