Phoneme-inspired acoustic frame embedded lightweight transformer for rolling bearing fault diagnosis
Linhao Peng, Fang Liu, Lv Zhongliang, Yongbin Liu, Min Xia
Abstract
Acoustic-based fault diagnosis has gained increasing attention due to its flexibility and applicability under diverse operating conditions. Recent studies have demonstrated that Transformer-based architectures exhibit remarkable capability in modeling long-range temporal dependencies, so they are particularly effective for complex time series representation and analysis. However, applying Transformers to acoustic fault diagnosis faces the following critical challenges. The primary hurdle lies in discriminating short-duration nonstationary acoustic features amidst complex environmental interference, Simultaneously, how to ensure model diagnostic performance under limited computing resources is a challenge. Inspired by the phoneme processing paradigm where continuous speech is discretized into meaningful units, this paper proposes a lightweight acoustic segment-frame embedded Transformer, termed ASFE-Transformer, to address these issues effectively. First, an Analytic Amplitude-Phase Representation Module is introduced to derive physically meaningful representations, ensuring that the instantaneous evolution characteristics of multidimensional acoustic signals are faithfully retained. Second, an Acoustic Frame Embedding Module is developed to aggregate continuous analytic samples into a single token during the embedding stage. This strategy generates a more efficient token sequence and explicitly enhances the discriminability of transient fault cues while substantially reducing the computational burden. Third, the feed-forward network of the Transformer is redesigned by proposing a novel Swish-Depthwise Gated Linear Unit. By embedding a lightweight depth-wise convolution within the Swish-activated gating mechanism, this module introduces local contextual awareness into the encoder. Comprehensive experiments on two bearing acoustic datasets verify that the proposed framework delivers robust diagnostic performance with a lightweight design and highlight its strong potential for real-world industrial deployment.