Problems in applying Kalmennet to vehicle slip angle estimation

[yufei0726]

Hello author, I'm encountering a problem when using Kalmennet for vehicle sideslip angle estimation. My approach is as follows:

1. I used data collected by the high-fidelity simulation platform Carsim, processing it into approximately 130 state sequences of 5000 time series each.

2. Similar to conventional KF, I used [beta, yawrate] as the state variables and [ay, wz] as the observations.

3. I modified the source code for GPU batch training, using a normally normalized output and ground truth labels distributed in the training set during supervision to reduce the influence of dimensions; the batch size is 64, the epoch number is 150, and each epoch iterates through 5000 time steps, with the estimation for each step being computed in parallel across the batch size.

However, the current training effect is very limited. The estimated beta shows a similar trend, but the numerical magnitude still differs. I would like to ask if you could provide some suggestions.

[yufei0726]

Additionally, I noticed you mentioned in your answers to other issues that RNNs need to be correctly initialized, and that the length of a trajectory segment cannot be too long. Could you elaborate on how to ensure correct initialization, or do I need to shorten each trajectory slice? I'm currently initializing hidden variables using the open-source code torch.randn from another paper by your team, split-kalmennet.