Feature Engineering Study for Signal Identification of Sealed Relay Components

Particle Impact Noise Detection is a highly reliable remainder particles detection method for components and equipment in China. In the process of remainder detection, the movable parts inside the components will also vibrate, a pulse signal similar to the remainder particles' waveform will be generated, which is called the component signal. the distinction between the component signal and the remainder signal is the premise of subsequent detection, which will also significantly affect the effect of remainder detection. Based on the theory of Feature Engineering, this paper constructs and selects the effective recognition features of component signals to improve the recognition accuracy of component signals. Firstly, based on the original recognition features, the voiceprint recognition features are selected and added, then combined the Pearson correlation coefficient method and feature importance selection method to select the top 10 features in performance. Then the ten features are constructed to obtain a large number of constructed features. In the end, these construction features are selected again to find the effective recognition feature combination of component signals. The experimental results show that the classification accuracy of the new feature combination is 96.57%. This construction and selection method can be extended to other particle signals recognition fields and effectively improve signal recognition accuracy without increasing experimental samples.