PSI - Issue 38

Frédéric Kihm et al. / Procedia Structural Integrity 38 (2022) 12–29 Kihm, Miu, Bonato / Structural Integrity Procedia 00 (2021) 000 – 000

18

7

3.3. Data Processing

The data processing is different for each signal, it is therefore made on a per-signal basis. The acceleration signal is integrated twice with the resulting drift being removed at each step using a high-pass filter in order to obtain a reliable estimate of displacement. The temperature signal is turned into a pseudo strain by applying the material thermal expansion coefficient. A similar logic was applied to the pressure to turn it into pseudo-stress. Each signal is rainflow cycle counted and relative damage is calculated periodically on 30-second windows. This window length was carefully chosen so that it is sufficiently large to contain complete fatigue cycles, while small enough to capture short transients. The fatigue slope parameter was taken as 3.0 for this study because this value is typically used for welds in the endurance regime (that’s where the failure is expected to occur) (Bannantine (1990)) and because it results in a very steep fatigue curve, giving more relative importance to cycles with small amplitudes. The signal processing workflow for the acceleration signal is illustrated in Fig. 2.

Fig. 2. The part of the signal processing workflow dedicated to the acceleration signal

3.4. Principal Component Analysis The strain-derived damage is considered to be an output signal and the remaining acceleration, temperature and pressure induced damages are considered to be inputs to the model. Dimensionality reduction using PCA is performed on the damage figures for the three input channels. However, as can be seen in Fig. 3, all three features are needed to explain at least 95% of the variance. This means that we cannot obtain a dimensionality reduction.

Fig. 3. Level of cumulative variance explained using the various Principal Components.

UNRESTRICTED