Interpretable CNN models for computationally efficient bearing fault diagnosis using learnable Gaussian/Sinc filters

Abstract

Real-time monitoring of bearing health is essential for maintaining rotary machinery’s operational efficiency and reliability. Nowadays, deep learning-based convolutional neural networks (CNNs) have gained popularity for diagnosing various bearing fault conditions, surpassing traditional machine learning methods that rely on time–frequency analysis, feature extraction, and supervised learning. However, the foundational architecture of CNNs—including filter size, number of filters in convolutional layers, optimizers, batch size, and more—significantly impacts model performance. Among these parameters, filter size selection is particularly crucial, as it directly affects the computational efficiency and effectiveness of the trained model. To address these challenges, this paper presents an approach integrating parameter learning for structured functions used as filter banks within CNNs architecture. In the proposed architecture, the initial layer includes parameterized learnable filters (LFs) that operate directly on raw data, producing frequency-related features through the structured design of the filter functions. Specifically, two types of parameterized LFs—Sinc and Gaussian filters—are introduced, each providing distinct bandpass filtering in the frequency domain to improve the direct classification of raw time-series data. To validate its effectiveness, we used state-of-the-art bearing fault datasets leveraging vibration signals. Experimental results demonstrate that the proposed approach successfully detects various bearing fault conditions and achieves performance similar to benchmark methods, even with limited training data. Thus, the proposed method enhances classification performance while improving the interpretability and understanding of CNNs operations, leading to more effective bearing fault diagnosis.