An Integration of Spectrum Analysis and Attention-based Network for Condition Monitoring of Vibration Components

High reliability of some key components in various engineering systems is crucial to ensure stable performance, safety operation and economic life-cycle management. This requires efficient and effective condition monitoring and predictive maintenance. With the availability of massive sensor data, increasing attention has been paid to data-driven methods for automatic condition recognition and fault diagnosis. This study proposes a lightweight framework combining domain knowledge and deep learning technique. A novel network incorporating 1D convolution module and multi-head self-attention mechanism is designed to receive information from frequency domain by spectrum analysis. Experiments on multiple cases indicate that the developed method is able to achieve the state-of-the-art results and has the potential generality to cover vibration scenarios for different engineering applications.