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

490

4

the vibration signals collected in respective conditions are 111.03 hours, 23.67 hours, and 10.18 hours for normal, no helium, and bad-oil conditions. These signals were sampled at 5000 Hz and quantized into 16-bit data.

Table 1. Operational conditions of FESD. condition gas inside FESD

oil state

normal

helium

new new

no-helium

air air

bad-oil

deteriorated

3. Data-driven operation state classification

A CNN-based classifier was desinged to identify operation states of a FESD, such as normal, no-helium, and bad oil states listed in Table 1. Since the developed system yields possibilities of FESD operating in these three states, it serves for diagnosis of failure. In this system, data-driven features were extracted from vibration signals inside the neural network without any prior knowledge of experts. This section aims to investigate the effectiveness of a data driven method on operation state identification.

3.1 Operation state classification system

The neural network with three-channel, four convolutional layers followed by five fully-connected layers was employed to identify an unknown input into three operation states (i.e., normal, no-helium, and bad-oil). The network architecture is illustrated in Fig. 3. 2500-dimensional power spectra were extracted from recorded vibration signals for every one second by fast Fourier transform (FFT) and taken as inputs to the CNN-based operation state classifier. Note that mappings of input spectra and corresponding operation states were learned in an end-to-end fashion i.e., feature extraction and classification can be performed with a single neural network. In this example, convolutional layers and fully-connected layers can be considered to be a feature extractor and a classifier, respectively. For the convolutional network, the filter size, kernel size, and stride size in four respective layers are set to {(6,5,3), (10,5,3), (10,5,3), and (2,5,2)}. The number of units in each fully-connected layer was set to {64 – 32 – 16 – 3}. At the fourth convolutional layer, three-channel outputs were flatten into a 175-dimensional vector and taken as an input to the first fully-connected layer. A ReLU activation was applied to the output of each hidden layer. The network was optimized with Adam.

`

Fig. 3. Network architecture of operation state classifier that consists of four-layered convolutional network followed by five-layered fully connected network.