Journal of information and
communication convergence engineering
(J. Inf. Commun. Converg. Eng.)

Framework of this study. Training and test spectra go through the same preprocessing steps (green) except for the augmentation (red) which is performed only on the training data before they are fed to the deep learning models (blue).|@|~(^,^)~|@|Schematic diagram indicating the process of Raman spectral acquisition from a small contamination particle in HDD. Raman spectra can be used to characterize and identify the particle types. Due to the relatively smallsized particles compared to the laser beam, the acquired Raman spectrum is noisy and difficult to identify.|@|~(^,^)~|@|Spectra of 10 substances after baseline correction. The horizontal and vertical axes represent wavenumber in cm−1 and signal intensity, respectively.|@|~(^,^)~|@|The structure of the CNN model consists of two parts: feature extraction and classification.|@|~(^,^)~|@|Confusion matrix according to LeNet5 trained by the augmented dataset created from the background-noise method.|@|~(^,^)~|@|Class activation maps of LeNet-5 computed by HiResCAM [26] on 10 substances. The horizontal and vertical axes represent wavenumber in cm−1 and signal intensity, respectively.

© J. Inf. Commun. Converg. Eng.