Machine Fault Diagnosis based on Vibration Analysis and Convolutional Neural Network

We developed an intuitive fault diagnosis method related to human auditory characteristics by applying sound quality parameters to the vibration signal. Abnormal noise was generated from a fault in a machine, and the operator used this noise to detect the abnormal condition. Although these acoustic characteristics differ with the fault conditions, diagnosing various fault conditions is limited by the auditory characteristics of the operator. The sound quality parameters represent the human auditory characteristics as physical quantities. However, the sound signal is vulnerable to various external noises generated in the environment. The vibration signal is robust to various external noises generated in the process, and the sound signal and vibration signal have a high correlation. Therefore, the vibration signals of the normal condition and various fault conditions were measured using a laser Doppler vibrometer (LDV). The sound quality parameters were applied to the vibration signal, and the characteristics of the sound quality parameters for each machine condition were analyzed. A convolutional neural network (CNN) was used to extract important features from the sound quality parameters for pattern recognition. The classified features facilitated a clear demarcation between the conditions of the machine. The classification performance of the proposed method was verified through comparison with other classification models.