Prior knowledge-guided improved wiener filter network for gear interpretable edge diagnosis under multi-speed conditions

Abstract Benefiting from end-to-end learning, automatic feature extraction, and high-dimensional data processing capabilities, deep learning significantly improves the accuracy and efficiency of mechanical fault diagnosis. However, its diagnostic performance under varying operating conditions of gears (e.g. across different rotational speeds) remains suboptimal due to limited robustness, while its inherent ‘black box’ nature leads to poor interpretability, making it difficult for users to trust and adopt. To address these challenges, this study proposes an interpretable prior knowledge-guided improved wiener filtering network with signal processing and deep learning collaborated, where frequency-domain prior knowledge is introduced into edge computing for gear interpretable edge diagnosis under multi-speed conditions. This interpretable network comprises a lightweight three-layer architecture: an adaptive mode learning layer (AMLLayer), a prior knowledge-guided feature pooling layer (PK-FPLayer), and a classifier. Considering the principle of gear signal modulation mechanism under speed variation, multiple interpretable improved wiener filtering kernels (IWFKs) are first designed as AMLLayer of deep learning architecture, where these speed fitted-kernels dynamically adapt their center frequencies and bandwidth penalty coefficients to enable adaptive extraction of gear fault modes across multiple rotational speeds. Subsequently, the (PK-FPLayer) calculates multiple frequency-domain feature indicators from the extracted modes to capture fault-related information. Finally, the classifier makes decisions via two fully connected layers. In contrast to conventional cross-speed gear fault diagnosis models, this study integrates signal processing principles with deep learning, leveraging a speed-adaptive strategy to mine fault-related information across rotational speeds, thereby enhancing the model’s interpretability. Experimental results demonstrate that the performance of the proposed method outperforms other state-of-the-art approaches in cross-speed recognition tasks. Visualization analyses further validate the model’s interpretability, illustrating the adaptive extraction capability of the filtering kernels under multi-speed conditions. Additionally, owing to the model’s lightweight architecture with fewer parameters and lower computational cost, we successfully implement it on edge-embedded microsystems devices, achieving superior real-time diagnostic performance in edge-based condition monitoring scenarios.