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

Abstract

Machine learning-based equipment diagnostic system is emerging as an effective means of prognostics and health management. The difficulty of collecting data in industrial sites makes difficult to apply this diagnostic approach. Especially with signal data, there is a problem that it is difficult to collect enough fault-type data. To solve this problem, deep learning-based data generation such as Generative Adversarial Networks has been proposed to create insufficient signal data. However, previous deep learning-based vibration signal generation approaches can create a specific fault among multiple fault types, but the user cannot control the signal features during the data generation process. Therefore, this study proposes a generation model that allows the user to control the features of the generated vibration signal. The proposed method leveraged the signal feature information in the data. Proposed generation model was trained based on Information Maximizing GAN with an attention module, Convolutional Block Attention Module. To extract and control the signal features of the data by unsupervised learning, a loss function utilizing the probability density function of the normal distribution was applied. The generated data of the proposed model was validated through qualitative and quantitative evaluation with real data, and the controllability of signal generation was checked. By allowing the user to control the features of the generated vibration signal, proposed study contributes to reducing the difficulty of implementing machine learning-based facility diagnosis in industrial sites by improving the utility of deep learning-based data generation for signal data.