Deep Learning-based Vibration Signal Generation with Frequency-amplitude Variability

Machine learning-based equipment diagnostics are increasingly important for prognostics and health management, but limited industrial fault data—especially vibration signal data—hinders practical implementation. Collecting sufficient fault-type data in real industrial settings is difficult, reducing the effectiveness of diagnostic models. Although deep learning-based data generation methods such as Generative Adversarial Networks (GANs) have been introduced to address data scarcity, existing approaches typically generate only specific fault types without allowing users to control signal characteristics. To overcome this limitation, this study proposes a controllable vibration signal generation model that enables users to manipulate generated signal features. The proposed approach incorporates signal feature information into the generation process using an Information Maximizing GAN combined with a Convolutional Block Attention Module (CBAM). To achieve unsupervised extraction and control of signal features, a loss function based on the probability density function of the normal distribution was introduced. The generated vibration data were validated through qualitative and quantitative comparisons with real data, and the controllability of generated features was confirmed. By enabling user-controlled feature generation, the proposed model enhances the practical utility of deep learning-based signal data augmentation and helps reduce barriers to implementing machine learning-based equipment diagnostics in industrial environments.