PSI - Issue 38

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

28

17

The p-value of the F-statistic indicates statistical significance of the model, so it is expected to offer better predictive value than predicting the average of the output. Both parameters are statistically significant and have the same magnitude order effects on the output. The mean relative absolute error of this model is 38.2%, which is an improvement over the model based only on Rear_Axle_Inclination. A further enhancement to the model is attempted by including yet another input dimension (the longitudinal acceleration). The results are presented in Table 11:

Table 11. OLS results.with Rear_axle_inclination, Accel_Lat_Chassis and Accel_Long_Chassis Damage inputs Dep. Variable: StrainDamage R-squared 0.508 No. Observations 1220 Adj. R-squared 0.507 Df Residuals 1216 Prob (F-statistic): 9.44e-187 Df Model 3

coef

P value

[0.025 0.114 0.399 0.089 0.039

0.975] 0.137 0.467 0.161 0.135

const

0.1251

0.000 0.000 0.000 0.000

Rear_axle_inclinationDamage 0.4332

Accel_Lat_ChassisDamage Accel_Long_ChassisDamage

0.1250 0.0867

Further analysis of this model is not justified because, while the additional parameter is statistically significant, its influence on the output is far smaller than the other ones. We can deduce that this third parameter is practically superfluous for predicting relative damage. We can conclude that the relative damage associated with the attached strain gauge can be alternatively predicted by one or two sensors with an error of 38-40%. The choice of instrumentation may take into account factors such as feasibility or desired accuracy. It is interesting to note that the best results are obtained from sensor quantities, which require explicit instrumentation. We attempted to correlate damage with the steering wheel angle, which does not require additional instrumentation, since this quantity is readily available from the CAN bus, but this did not help improve the results. A linear model with only the steering angle as the independent variable suggests that this is statistically significant, but the overall model RAE is disappointingly high at 75.1%. Combining the steering angle with transversal acceleration yields in a model with both components deemed as statistically significant, but an RAE score of 49.6%, which is only 0.7% less than the transversal-acceleration-only model, making the effect of the steering angle almost negligible. In this case, the engineer – without running this data science exercise – might have considered more inputs as significant to the model. It turns out that the damage in the location where the strain gage was positioned was not sensitive to most parameters that the expert would have naturally thought of. 5. Tools used This analysis for both case studies were performed on a large number of time series files coming from a mixture of CAN data and high precision sensor data, collected using the QuantumX data acquisition system. The data was then indexed in Aqira, which gives access to nCodeDS for high speed, streamed data processing and to a Python environment. The data science aspects of this study were done in Python, with pandas and scikit-learn. The input data, analyses and results were centralized and shared between the co-authors in Aqira, enabling full collaboration.

UNRESTRICTED