Convolutional autoencoder based condition monitoring system for unique complex technical systems

This paper presents the development of a condition monitoring system for a unique, complex technical system represented by a drop tower. The self-developed drop tower, the Einstein-Elevator, is designed as a research platform for highly reproducible zero gravity conditions. As the Einstein-Elevator is a new and unique system with a low number of data samples, detecting wear and other faulty behavior is a challenging task. However, this is important to foresee and counteract cost and time intensive shutdowns. For this purpose, a condition monitoring system based on a neural network approach using acceleration data has been developed together with a framework. According to the state of the art, first the Einstein-Elevator is presented, followed by the developed six-stage framework for model generation. The framework includes both the pre-processing of sensor data as well as the creation and optimization of a data-driven model for diagnosing faulty operating. Additionally, the results showcase the performance of the convolutional autoencoder, which was trained to accurately reconstruct spectrograms of normal flight samples without anomalies. Subsequently, the model was evaluated based on the size of the reconstruction error that occurred with the implementation of anomalous samples. In order to reduce under- or overfitting and improve the model, data augmentation via cutout-methods has been used and validated. This approach resulted in an improved anomaly detection capability as evidenced by several metrics such as the accuracy (97.22%) or the precision (93.88%). Based on the gained research results, the framework has been implemented for the use at the Einstein-Elevator. The work concludes with an outlook for further model optimizations. The developed condition monitoring system approach for a high dynamic system with very precise quality requirements/specifications can be transferred to similarly complex technical systems of various applications, which will be part of future work.