Wear evaluation of hard disk drive head based on a converter-like neural network

Abstract

Hard disk drives (HDDs) are indispensable industrial products in modern information society, with their heads playing a critical role in read/write performance, which is directly affected by head wear. Head wear at nanoscale challenges the applicability of traditional macroscopic physical wear models, necessitating innovative solutions for accurate wear estimation. Data-driven methodologies have emerged as an effective paradigm, yet facing limitations when addressing the multiple phases of head wear degradation. Herein, building on insights into nanoscale wear and interface dynamics, we devised a "converter" module. Simultaneously, a physical constraint is incorporated into the network to guarantee the physical compliance consistency of wear outcomes. The effectiveness of this method is validated on real worn heads. An interpretive analysis is performed, involving the visualization of GRU module's hidden units and the quantification of feature’s activity contribution. Our investigations revealed that the converter results in a higher proportion of neurons becoming actively engaged and discriminative in wear evaluation, thereby enhancing the network's sensitivity to wear-related variations. Meanwhile, the trace ratios of scatter indicate that the converter improves the discriminability of the input data, thereby bolstering the network's capacity to differentiate between wear states.