Impact of the STFT Window Size on Classification of Grain-Oriented Electrical Steels from Barkhausen Noise Time–Frequency Spectrograms via Deep CNNs

The Magnetic Barkhausen Noise (MBN) is a non-destructive testing method, which, due to its high sensitivity to changes in the microstructure of the material, is increasingly being applied with success as a tool for evaluation of magnetic material state and properties. However, it is no less difficult to analyze the measurement signals and their correct interpretation due to the complex, non-deterministic and stochastic nature of the Barkhausen phenomenon. Depending on the material to be examined, a signal with different characteristics can be observed. Frequently, a signal with multi-phase Barkhausen activity characteristics is obtained, like in the case of grain-oriented electrical steels. Due to the increased computational capabilities of computers, more and more advanced signal analysis methods are being used and artificial intelligence is being involved as well. Recently, the time–frequency (TF) approach for MBN signal analysis was introduced and discussed in several papers, where short-time Fourier Transform (STFT) found frequent application with promising results. Due to the automation of the search for diagnostic patterns, the stage of selecting transformation parameters becomes extremely important in the process of preparing training data for evaluation algorithms. This paper investigates the influence of the STFT computational window size on the material state evaluation results obtained using convolutional neural network (CNN). The studies were performed for MBN signals obtained from grain-oriented electrical steel with anisotropic properties. The carried out work made it possible to draw connections on the importance of the choice of the window during the implementation of CNN network training.