구조 진동 분석 및 딥러닝 기법을 이용한 공작기계 상태 모니터링

Abstract

Rotating machinery is the most widely used type of machinery. It plays an important role in various industrial sites. Early fault detection technology is highly important to prevent an entire system from being deactivated by a fault in the rotating machinery. The condition monitoring of rotating machinery has been performed using various sensors such as current, temperature, and vibration sensors. Among these, the vibration signal effectively determines the machine state based on feature extraction through spectral analysis. It is important that the extracted features have the physical characteristics of the current state. In addition, advances in deep learning technology enable vibration-based real-time condition monitoring with superior performance. This dissertation presents a new methodology for machine tool condition monitoring using structural vibration analysis and deep learning methods. Chatter is the most common fault in machine tools. Chatter occurs owing to the interaction between the cutter and workpiece. The main parts such as spindle and bearings are damaged owing to the strong cutting force by chatter. Therefore, a real-time cutting monitoring system is highly important to protect various components such as the cutter, workpiece, and machine tool. The condition of the machine tool is monitored by the vibration measurement of the spindle head. Harmonics generated by the spindle rotation dominate as peaks in the frequency domain. The short-pass lifter is applied to the cepstrum to effectively remove the harmonics and extract the structural vibration modal components of the machine. The vibration modal components include information regarding the wave propagation from the cutter impact to the spindle head. The force excitation from the interactions between the cutter and workpiece induces structural vibrations of the spindle head. The vibration magnitude for the in-phase flexural modes is smaller in the chatter state compared with that in the stable state. The converse variation is observed for the anti-phase flexural modes. The liftered spectrum is used to obtain this dependence of vibration on the cutting states. Supervised learning is effective in a situation where there is sufficient data and labels for normal and fault states. The convolutional neural network (CNN) is a type of supervised learning. It extracts the required features from the liftered spectrum for pattern recognition of the cutting state. The classified features enable demarcation between stable and chatter states. In addition, transfer learning is used to secure the robustness of fault diagnosis in a dynamic environment. A trained model for the existing cutter is used as an initial value to construct a chatter detection model for another cutter. In the field of fault diagnosis, feature selection is an important task to identify a mechanical system and to achieve a higher classification accuracy. The sub-band attention CNN (SBA-CNN) is a structure that combines the sub-band CNN (SB-CNN) and attention layer. Unlike conventional CNN that addresses all frequency components with an identical filter bank, the SB-CNN uses a different filter bank for each band. The attention layer is used to evaluate and emphasize the importance of the frequency bands using features extracted from the SB-CNN. The SBA-CNN is trained to classify cutting states using a liftered spectrum. It successfully evaluates the importance of frequency bands. Phase-based motion processing is used for measuring the vibration of an entire object using a camera. The mode shape of the object is visualized by vision-based vibration measurement. The reliability of the vision-based vibration response is compared with the acceleration vibration response. In rotating machines, the rotating components and their harmonics and mode components are measured with vision-based vibration. Normal and worn cutters are analyzed with vision-based vibration. The vibration magnitude is reduced in the in-phase flexural mode of the machine head when machining is performed with a worn cutter. A liftered spectrum is used to extract the modal response regardless of the rotation frequency. The magnitude of the response at the mode frequency is used to detect cutter wear. A fault diagnosis is performed for the normal, rotor imbalance, and shaft misalignment states of a rotating machine. The vibration response of the rotating machine is measured using camera vision. The rotational component and its harmonics are investigated for feature extraction. Variational mode decomposition (VMD) is used to extract the shaft mode of the vibration response. The vibration response to the shaft misalignment state is more pronounced than that to the rotor imbalance state at the shaft mode frequency. Vision-based vibration response and CNN are used for the fault diagnosis of rotating machinery. The rotational component, its harmonics, and the axial mode response extracted from the VMD are used as the input to the CNN. The classified features identify the stable, rotor imbalance, and shaft misalignment states. Explainable artificial intelligence (XAI) is used for analyzing the diagnosis model. It can be used to identify the important features when a model is used to classify states. A convolutional autoencoder is an unsupervised learning algorithm. It is trained using only normal data in a situation where it is difficult to acquire fault data. Fault occurrence is monitored by forming a boundary for the normal data based on the convolutional autoencoder.