Development of machine learning based anomaly detection methods in predictive maintenance applications

Predictive maintenance is an approach to detect and prevent unexpected equipment failures by monitoring machine conditions. Equipment used in industrial systems should be regularly maintained. Otherwise, a sudden breakdown can lead to financial losses. Advanced sensor data acquisition technologies allow many types of sensor data to be generated and stored digitally. These data can be utilized with artificial intelligence methods to obtain information about the equipment's condition. One of the challenges in fault detection is the potential inadequacy of a single sensor's performance. Multi-sensor fusion aims to create a more robust structure by effectively combining data from multiple sensors. Furthermore, different modes of data obtained from a source may provide different information. In this thesis, a deep learning architecture including both multi-modal learning and multi-sensor fusion was developed to detect equipment faults. In order to detect the fault condition, the representations of the data obtained from the vibration and current sensors in both time space and time-frequency space are used together. The raw data from vibration and current sensors were converted into spectrogram images using the Short Time Fourier Transform (STFT). Subsequently, time-frequency images and raw time-series data were provided in parallel to the proposed method for fault detection. The effectiveness of the proposed method was investigated by comparing it with different techniques on two different datasets. The results indicate that the proposed method is successful in detecting faults.