Deep Convolutional Autoencoder Architecture for Predictive Maintenance Applications

Maintenance of the machinery is a crucial task in industrial production sectors working with machinery. The most important aspect of maintenance is timing. Executing maintenances more frequently or sparsely than the necessary amount causes separate problems resulting with unnecessary expenses or halts in the production. To prevent these problems, a smart system to decide the timing of the maintenance must be established. In this study, we develop an auto-encoder extension of previously proposed deep convolutional network that is trained successfully on the modelling of electroencephalogram (EEG) signals with high performance. The auto-encoder extracts features from the vibration signals collected from the machinery. This method allows us to synthesize multi-channel vibration data which we use to classify the type of the failure that the machinery bearing is going to face, without expert field knowledge and with a high accuracy. The performance of the proposed network is tested on the publicly available Case Western Reserve University (CWRU) bearing dataset with the classification accuracy. Proposed network showed a better classification performance, allowed smaller bottleneck feature sizes and faster training times compared to the Normalized Sparse Auto-Encoder - Locally Connected Network (NSAE-LCN), which is one of the best performing networks on the same dataset.