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

493

7

based and AE-based fault detectors. The architectures of (a) the CNN-based DDF extractor, (b) DNN-based fault detector, and (c) AE-based fault detector are illustrated in Fig. 5. Note that the CNN-based DDF extractor aimed to classify unknown inputs into three classes such as normal, no-helium, and bad-oil states while the subsequent DNN based fault detector classified the extracted DDF inputs into normal and faulty states. During decision in DNN-based fault detection, unknown inputs with the posterior probability of more than the pre-defined threshold were determined to be faults. In AE-based fault detection, a fault can be detected using a reconstruction error between the input and output layers; unknown inputs with the errors more than the pre-defined threshold were determined to be faults.

SOFT-MAX FC (3)

SOFT-MAX FC (2) RELU FC (8)

SIGMOID FC (32) RELU FC (16)

RELU CONV RELU CONV RELU CONV RELU CONV RELU FC (32) RELU FC (16) RELU FC (64) FLATTEN

RELU FC (16)

RELU FC (4)

DDF (32)

RELU FC (32) DDF (32)

RELU FC (16) RELU FC (32) DDF (32)

175

(2,5,2)

(10,5,3)

(10,5,3)

(6,5,3)

IN

(a) CNN-based DDF extractor

(b) DNN-based detector

(c) AE-based detector

Fig. 5. Network architectures of (a) data-driven feature (DDF) extractor, (b) DNN-based normal-anomaly classifier, and (c) AE-based novelty detector. CONV and FC express convolution and fully-connected layer, respectively. FLATTEN layer yields a vector from multichannel CONV layer outputs. Number in parenthesis in FC layer is unit size and numbers in parenthesis in CONV layer are filter size, kernel size, and stride size, respectively.

4.2. Experimental Results

Experimental comparions were conducted using vibration data sets listed in Table 2 to demonstrate the effectiveness of the developed fault detection system. Four-fold cross validation experiments were carried out and the averaged performance over four validation sets was used for comparisons as in Section 3. Data samples in the no helium and bad-oil states were assigned to the fault class during development of the DNN-based normal-anomaly classifier. Table 4 lists performances of detecting faults for the DNN-based normal-anomaly classifier and AE-based novelty detector. This table contains the numbers of data samples in the normal, no-helium, and bad-oil states, identified into the normal and faulty state, together with precisions and a recall. The result showd that the DNN-based and AE-based systems yielded comparable performance, but the novelty detection approach performed slightly better in terms of the recall and precisions for faulty inputs. Especially during the opration with the bad-oil condition, faulty data were perfectly detected by the AE-based detector.