Effect of Noise on Accuracy of Grain Size Evaluation by Magnetic Barkhausen Noise Analysis

Abstract

Magnetic Barkhausen noise (MBN) is a magnetic signal caused by domain wall motion and is used for non-destructive testing and evaluation of ferromagnetic materials because of its sensitivity to both mechanical and magnetic properties. Recently, machine learning models have been employed to evaluate materials based on MBN; however, the application of material evaluation to low-volume targets is challenging because of their low signal-to-noise ratio, which is due to their low volume. Therefore, understanding the influence of the signal-to-noise ratio is important, particularly for low-volume objects. However, very few reports have quantitatively assessed the influence of noise in MBN analysis. In this study, we focused on noise to improve the accuracy of MBN analysis using machine learning, investigated its impact on the prediction accuracy of machine learning models, and explored methods to mitigate its effects. A method for grain size evaluation based on MBN analysis was adopted and performed for Fe-Co alloy wires with different grain sizes. After the measurement of MBN, the relationship between the extracted features from the analysis of MBN by fast Fourier transform and grain size was learned using a gradient boosting decision tree to create a grain size evaluation model, and the coefficient of determination was used to evaluate the prediction accuracy of the grain size evaluation. The machine learning model demonstrated high prediction accuracy (R2 = 0.926) across the entire grain size range. Using the model to assess the effect of signal-to-noise ratio, experiments were also conducted using MBN time-series data with artificially applied Gaussian noise. Additionally, from the insight of the distribution of predicted grain sizes, we confirmed that a noise reduction method by averaging the MBN prediction results can improve the prediction accuracy by reducing the effect of noise as expected. This research will lead to the adoption of MBN applications, which are simple and practical methods of material evaluation, for the micro–nano discipline.