A Deep Learning Algorithm for Classifying Local Material Defects in Aluminum Castings Based on Signals from a Self-Excited Acoustic System SAS

This article presents a deep learning algorithm for classifying local material inhomogeneities in aluminum alloy castings based on signals recorded using a self-excited acoustic system (SAS). The aim of this study is to evaluate the potential of using acoustic signals generated in a self-excited system to distinguish between defect-free specimens and those containing internal material defects. The tests were conducted on two specimens with the same geometry: a reference specimen without defects and a specimen with internal defects confirmed by X-ray inspection. Measurements were taken at six corresponding measurement points. A preliminary frequency analysis showed that the presence of inhomogeneities affects the system’s acoustic response; however, the nature of these changes depends on the measurement location and does not allow for the use of a single universal decision criterion based solely on the dominant frequency. For this reason, a classification procedure was developed using the time-frequency representation of the signal and the Audio Spectrogram Transformer architecture. The procedure includes SAS signal acquisition, preprocessing, signal transformation into spectrograms, feature extraction using the AST model, and classification of samples as OK or NOK. The proposed algorithm was evaluated using accuracy, precision, recall, F1-score, AUC, and a confusion matrix. The results indicate that the proposed method can support the non-destructive testing of aluminum castings and serves as a basis for the further development of acoustic signal classification algorithms for the detection of localized material defects.