Companies with a lot of industry-grade systems have large incitements for planning and predicting maintenance on their systems. Maintenance is an important field with established routines working in stages: Corrective, Preventive and Predictive. Predictive maintenance can be achieved by the means of forecasting future anomalous states of systems through machine learning and manual feature engineering models. Current methods commonly use manual feature engineering, which requires a lot of domain expertise and may be over-optimized to already known anomaly cases. This thesis is written in collaboration with Elvaco, a company specializing in metering and collecting data. Together, we seek to implement machine learning to further improve their predictive maintenance in one of their field, district heating. Predictive maintenance can be achieved through unsupervised anomaly detection. When the problem domain is fuzzy and all types of anomalies are not known beforehand, this makes more sense than a supervised model.Our main contribution in this thesis is learning a useful representation for anomaly detection based on multiple domains, where conventional methods and single-system autoencoders would need more data to produce good results. We propose a novel autoencoder-based model that automatically learns features useful for anomaly detection, where all healthy data is represented similarly. An anomaly detector is then fitted on these features to detect deviating data. We also make an investigation of the application of Transfer Learning within the field of district heating. Here, we train an autoencoder on the data from many substations and then try fine-tuning the trained autoencoder using similar substations to that of the target substation.

The result shows that there was no overall performance increase for any method. As both of the baselines achieved the same performance, but the learnt representation did show that it could represent healthy data similarly. A further investigation of each model showed that selection of target substation had great impact of performance. A last investigation if increasing number of substations for training increased performance. The result showed that it had no impact on performance.