This thesis examines the possibilities of a more efficient condition-based maintenance in the railway sector by using trend analysis with temperature data from hot box detectors. The study shows that the use of trend analysis directly on hot box measurements could induce misleading predictions due to certain detector characteristics. Because of this, multiple linear regression models were used to estimate the temperature of bearings based on several known parameters for each detector. A second type of linear model was then used to predict future values of the difference between the estimated temperature and the measured temperature. Results from this study show a large variation in the predicted values. The variations of the predictions depend on the amount of measurements included in the model, the time difference between the measurements, the uncertainty of each measurement, the mean value of the measurements included, and the weight function of the model. Large prediction variations resulted in predicted values where it was difficult to separate scenarios that could be considered "normal" from scenarios where a hot box warning would be desirable. For the prediction model to be able to reduce the risk for false alarms without increasing the risk of overlooking actual faults the model needs to be improved or combined with other models.