13th International Conference on Fracture June 16–21, 2013, Beijing, China -5- fatigue, or corrosion. Therefore, the experiment here was mainly designed to identify this form of damage. The structural defect was simulated by a notch at different depths, which was directly machined on a milling machine, as shown in the upper left corner of Fig. 3. Based on the severity of the damage, there were four groups. The first group represents the baseline health state without any defect; the second group induces a defect whose size is 46 mm by 48 mm with a depth of 1.5 mm (50% thickness), as shown in the yellow region of Fig. 2. In the third group, the size of the defect remains the same, but the depth is increased to 2.4 mm (80% thickness). Finally, in the fourth group, the defect is completely though the plate (3 mm depth). The torque of the four bolts connecting the plate and the support was 80 Nm, which was strictly controlled by the torque wrench in each experimental group. The image of the experimental real objects can be observed in Fig. 3, and the details of the different health states simulated are summarized in Table 1. Figure 3. Picture of the experiment Table 1. Details of different health conditions State Label Description Health state State 1 Baseline state, no defects Damage state 1 State 2 Defect size: 46mm × 48mm × 1.5mm Damage state 2 State 3 Defect size: 46mm × 48mm × 2.4mm Damage state 3 State 4 Defect size: 46mm × 48mm × 3mm 3.2. Detection Results of Step 1 For each damage state, 50 sample records are collected as damage identification according to the periodicity characteristics of the response signal. In Step 1, the dynamic strain signal from FBG is first decomposed into several intrinsic mode functions (IMFs), then the AR-based model is applied on the second level IMF component to extract the damage sensitive features and Mahalanobis distance-based pattern classification method are used to detect and locate damage. Here the sample sets from State 1 (health state), State 2 (damage state 1), State 3 (damage state 2), and State 4 (damage state 3), totally 200 samples, are used for damage identification. These samples are divided into two parts. One is the first 25 samples of State 1 as the training sample set. Another is the test sample set composed of the later 25 samples of State 1 and the remaining three damage states, totally 175 samples. The identification results of the channel 19, 26, 27, and 28 are shown in Fig. 4,
RkJQdWJsaXNoZXIy MjM0NDE=