주파수-진폭 변동성을 갖는 딥러닝 기반 진동신호 생성

키워드

확장요약문

Machine learning-based equipment diagnosis requires vibration data representing diverse operating conditions. However, scarce fault data and class imbalance limit model training. Existing generative models can augment datasets but offer limited control over generated signal characteristics. This study proposes a deep learning-based model that enables users to control frequency-component amplitudes and spectral variations associated with equipment conditions. The proposed model learns vibration signal spectra obtained through the fast Fourier transform (FFT) using an information maximizing generative adversarial network (InfoGAN). A convolutional block attention module (CBAM) enhances feature extraction in the discriminator, while an auxiliary recognition network estimates the latent code’s mean and variance. A Gaussian negative log-likelihood loss enables unsupervised learning of controllable spectral features. Validation using industrial fan vibration data and the Case Western Reserve University (CWRU) bearing dataset showed that generated data reproduced key frequency characteristics and feature distributions, achieving lower FID, KLD, and JSD values than a mean– variance-based generation method. With single-condition training, varying the latent code controlled amplitudes of important frequency components. With combined normal and abnormal data, generated spectra transitioned continuously between conditions. These results demonstrate controllable vibration spectrum generation through unsupervised learning and support its potential use in data augmentation for equipment condition diagnosis in industrial applications.