Image Processing and Machine Learning for Engine Fault Detection

The aim of this thesis is to design a process for the automatic detection of engine faults by analysing sound spectrograms. The designed technique mimics the manual state-of-the-art analysis performed by domain experts. Each operation of our method is performed automatically, without requiring domain experts intervention. The manual process being replicated consists of analysing the spectrogram image obtained from the sound emitted by the engine under inspection. This visualized spectrogram allows highlighting engine faults, which occur with peculiar characteristics that are well known by the experts. In particular, we focused on the detection of a whistle called constant tone, which appears as a straight noisy line in the spectrogram. The inspection typically requires a visual analysis performed manually by a domain expert. The approach proposed in this thesis addresses automatically this operation by means of image processing and machine learning techniques. The results are evaluated against the performances obtained with the manual process, which is considered as our benchmark. Our model is able to generate high quality detections, which are also interpretable, as they are provided with the spectrogram region where the problem occurs. It is hope that this work will lead a spectrogram analyst to speed up his work by automatizing steps that would be otherwise performed manually.