Autoencoder-based multivariate time series anomaly detection and clustering for diagnosis of mechatronic systems

In this thesis, the ability of a specific type of Deep Artificial Neural Network (DANN) called the Autoencoder to quantify the degree of fault, failure or malfunction (anomaly) of a mechatronic system using raw sensory data is analyzed. The Autoencoder approach follows the one-class classification paradigm and is trained unsupervised by reconstructing its own input using fault-free data only. For quantification of the faults or deterioration of the system, a health index is computed from an anomaly score, which is usually based on the reconstruction error. Four different distance metrics, Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mahalanobis distance and Cosine similarity are evaluated for the error calculation. Additionally, the error calculation is performed and evaluated using the latent variables, also called representation, of the Autoencoder. Time windows with different lengths from multivariate time-series data as its input are used during training and inferencing of the Artificial Neural Network (ANN) for analyzing the fault detection performances’ time dependence with transient data. The fault detection performance is evaluated using the Area Under Curve Receiver Operating Characteristic (AUCROC) on the three publicly available datasets New Commercial Modular Aero-Propulsion System Simulation (N-CMAPSS), Tennessee Eastman Process (TEP) and Gas Sensor Array Drift (GSAD) and compared against the Principal Component Analysis (PCA) as the model of the normal system instead of the Autoencoder. Various Autoencoder regularization terms and pre-layer e.g. Recurrent Neural Network (RNN) and Convolutional Neural Network (CNN) in combination with different activation function, objective function, batch size and other hyperparameter are varied on all datasets and health index computation methods in an extensive experimental series. With the best-performing combinations of Autoencoder types and hyperparameter the fault detection performance is further optimized by applying automated clustering on the training data. The clustering is done by the self-developed Autoencoder Based Iterative Modeling and multivariate time-series Clustering Algorithm (ABIMCA) approach, which shows a strong increase in the detection quality and can be used as a fault detection approach itself.

Author dc:creator

• Köhne, Jonas

Advisor dc:contributor.advisor

• Gühmann, Clemens

Licence dc:rights.uri

• https://creativecommons.org/licenses/by/4.0/

Language dc:language.iso

en

Identifier URI

https://doi.org/10.14279/depositonce-24677

OAI identifier oai:identifier

oai:depositonce.tu-berlin.de:11303/25852