Virtual sensors are software-based models that aggregates data from several sources that can perform the functions of physical sensors, improve the performance of lower quality sensors or measure data in spots where physical sensors are impractical to install. The company Komatsu Forest is interested in calculating the angle of the head of their harvester forestry machines to increase the positioning accuracy of produced wood logs and are investigating the possibility of using a virtual sensor with their already available data. In this thesis, machine learning techniques are used to produce a virtual sensor and the effectiveness of different methods is investigated. The solutions are based on LSTM networks, taking windows of data from the harvester machines and inputting them to produce a prediction for how the rotation of the harvester head has changed between timesteps. The different methods were evaluated and the highest R-Squared score reached was 0.673, which was deemed not high enough for the stated purpose. Several improvements and future research paths are suggested as the results still indicate that a solution to the problem with a virtual sensor is possible.