Nowadays, condition monitoring is becoming crucial to increase productivity and reduce maintenance costs and asset downtime. However, one of the greatest limits is that to obtain a robust and reliable algorithm a large amount of data is required. Getting those data from the experiments is costly both in terms of time and money. Damaging components to measure faulty data is not always possible. In this thesis, a lumped parameter model has been used to generate different fault levels in an axial piston pump. In this way, a large number of faults and their combination has been available to train the condition monitoring model with a reduced computational effort.