Machine learning models have achieved impressive results for the problem of vehicle state estimation, however the performance of these models can not be guaranteed in regions where the algorithm isn’t trained for. Considering the problem of vehicle state estimation, we aim to investigate techniques that can ensure model performance of machine learning methods, especially a side-slip angle estimator. We explore the use of Autoencoders for novelty detection. This thesis utilizes various Autoencoder based models for novelty detection. These models can detect abnormal events where the side-slip estimation model would yield poor quality predictions.The best performing variant uses gated recurrent units to learn the normal behavior in training data and predict future steps. Our experiments with standard test cases show that autnoencoders can be used to detect abnormal events like vehicle spins and sensor failures. In the second part we apply a technique called dropout as Bayesian approximation, to extract uncertainty estimates from deep neural networks. This technique has provided promising results as it indicates low confidence in situations where the estimation of the side-slip angle is not robust. Moreover, this technique was able to detect vehicle spins and banking. Since both approaches were not able to distinguish between different surfaces, we introduce a new idea for estimating the surface medium with accuracy up to 94% on test data.