This paper examines the effectiveness of neural network algorithms for hydraulic system fault detection and a novel neural network architecture is suggested. The proposed gated convolutional autoencoder was trained on a simulated training set augmented with just 0.2% data from the real test bench, dramatically reducing the time needed to spend with the actual hardware to build a high-quality fault detection model. Our fault detection model was validated on a test bench and showed accuracy of more than 99% of correctly recognized hydraulic system states with a 10-s sampling window. This model can be also leveraged to examine the decision boundaries of the classifier in the two-dimensional embedding space.

The problem of automatic reliability monitoring and reliability-centered maintenance is increasingly important today. In this paper, we compare the accuracy of four machine learning approaches for fault detection in a hydraulic system. The first three approaches are based on SVM classifiers with linear, polynomial and RBF kernels and the last one is a gradient boosting on oblivious decision trees. We evaluate algorithms on the synthetic dataset generated by our simulation model of the helicopter hydraulic system and show that high accuracy fault detection can be achieved.