PSI - Issue 17

Takanori Hasegawa et al. / Procedia Structural Integrity 17 (2019) 487–494 Takanori Hasegawa et al./ Structural Integrity Procedia 00 (2019) 000 – 000

492

6

aggregated, irrespective of recordings; bad-oil data (red symbols) have a multimodal distribution in response to recordings but are far apart from normal and no-helium data distributions; and no-helium data (blue symbols) also have a multimodal distribution depending on recordings but some components were overlapped with normal data distribution and some components are close to bad-oil data distributions. This figure indicates that difference in recordings induced the mismatch between training and testing data, especially in fault (i.e., no-helium and bad-oil) states. This analysis supports the result in Table 3 that the developed system did not yield good performance, especially in classifying the normal and no-helium data. Note that the preliminary experiment demonstrated that the developed CNN-based classifier achieved approximately perfect classification of three classes (99.97%) under a recording-closed condition where training and testing data were different but sampled from the same recording. Therefore, to reduce classification errors derived from such mismatch, training data should be increased, especially under fault conditions.

Fig. 4. Scatter diagrams of two-dimensional vibration data in normal, no-helium, and bad-oil operational conditions of FESD.

4. Data-driven anomaly detection

The no-helium and bad-oil states were regarded as faults and detected using neural networks in anomaly detection fashion. Note that anomaly detection focuses on detecting a fault not identifying a mode of a fault (e.g., no-helium and bad-oil). Therefore, developing anomaly detection systems does not necessarily use fault data. Since in general, it is not easy to collect fault data in advance, this approach is more adaptable. This section aims to investigate the effectiveness of data-driven methods on fault detection for FESD condition monitoring.

4.1. Anomaly detection system

There are two options in designing anomaly detection systems; one approach is based on normal-anomaly classification, which identifies an unknown input into a normal and a faulty operation state, and the other is based on novelty detection (or also referred to as outlier detection), which exploits only normal state data to train a model and detects an faulty state with less similarity to the trained normal state model. Training normal-anomaly classifier requires fault data and this is not general in condition monitoring as aforementioned. However, in the present study, both approaches are examined for comparisons. Five-layered fully-connected deep neural network (DNN) was employed for a normal-anomaly classifier and six-layered fully-connected autoencoder (AE) was employed for a normal state model in novelty detection. In this case, the CNN described in Section 3 was employed as a data-driven feature (DDF) extractor; hidden layer outputs of the second fully-connected layer before applying non-linear activations were taken as inputs to the DNN-