Machine Learning of Laser Ultrasonic Data to Predict Material Properties

Abstract: The hardness of steel is an important quality parameter for several industrial applications. Conventional mechanical testing is used in quality testing for material hardness and the method is time-consuming, can cause material mix-ups, and results in material waste. To address this issue, a possible on-line method for non-destructive testing (NDT) techniques such as laser ultrasonic (LUS) measurements has been explored to replace mechanical testing. In this thesis, machine learning models are trained to predict steel hardness using LUS measurements and data from the production process. LUS data is collected from steel samples with a measured hardness using the Brinell protocol. Measured hardness values between 250 and 700 Brinell are used as the target values for the models. The production process data includes the chemical composition and tempering temperature. The models used in this thesis are Extreme Gradient Boosting (XGBoost), Multilayered Perceptron (MLP), and Convolutional Neural Network (CNN). The first two mentioned models use feature-engineered data from LUS measurements. These features include the time-of-flight for ultrasonic waves. CNN uses the raw LUS data as a univariate time series as input. Each of the models is trained solely on data from LUS measurements and both LUS and production process data to determine the effect of adding production process data. The models are optimized and tuned based on their loss on a validation set. The models are evaluated against each other based on their root mean squared error (RMSE) on a test set to determine the best performing model. The best performing model is an XGBoost model using LUS and production process data. The results indicate that models using solely LUS data can not replace or partially replace mechanical testing. The best performing model using only LUS data has a RMSE of 69.9 Brinell, which is above the required performance of a RMSE below 50 Brinell. The results also indicate a large boost in performance if including data from the production process. However, implementing this solution in the industry without losing accuracy in measurements is a hard task. While the models are not ready for direct implementation in industry, the results demonstrate potential for further research in this area.