We propose a hybrid detection approach for pitting detection in gears. Most research on pitting detection with machine learning is done on supervised data and often on simulated pitting. However, for pitting detection in practice an unsupervised approach is required. The main idea behind the proposed solution is the ability to leverage elements of supervised machine learning models in pitting detection, while operating on unlabeled data. The training data used for this algorithm is taken from gear boxes with actual pitting failure and without prior knowledge about pitting size during different stages of operation, providing a realistic case for an operational scenario. The approach can be seen as two parts. At first we try to detect changes in the underlying structure of the vibration data using fast fourier transform to obtain frequency spectra that are fed to a sparse autoencoder. The encoded reduced feature space is the clustered to look for separability by time. In case there are significant changes in the underlying structure, an Long Short Term Memory model is trained to see if the changes can be validated as actual pitting damage. The LSTM can then further be used to adapt fast to the pitting damage with the goal of prolonging the lifespan of the gear.