Hybrid modelling for linear actuator diagnosis in absence of faulty data records

•Hybridization is proposed to deal with lack of faulty data for diagnosis.•A physical model of an electromechanical actuator and a real actuator are built.•Differences among data sources (physical model and test rig) data are reduced with a feature selection method.•Spalling and lubrication faults can be diagnosed with 90% kappa in the actuator without prior faulty records. The advantages of condition based maintenance over alternative maintenance strategies have been widely proven. Detection, diagnosis and prognostic algorithms enable the optimization of repair schedules while avoiding breakdowns and downtimes. However, some industrial limitations complicate the development of diagnostic monitoring algorithms, particularly in scenarios with unique or non-mass-produced machines, as obtaining faulty data records is difficult. This work proposes an approach that combines the data from physical models with data-based models (or hybrid modelling) to sort out the lack of faulty data records in the condition monitoring of a linear actuator. A test rig was built and used to collect data from healthy and faulty cases (the later only used for validation purposes), in addition to a physical model that simulated nominal (healthy) and faulty conditions to generate synthetic data. Synthetic and real measured data were combined with an improved fusion by means of a feature selection method. A diagnostic model was developed and the algorithm was validated in the detection of real faulty cases. Additionally, this approach is also valuable to detect unseen operating conditions. The results obtained in this work prove the validity of hybrid models for those cases in the industry where there are physical or economical limitations to obtain data records that difficult the implementation of diagnostic algorithms.