Anomaly detection models using big data generated from facilities and equipment have been adopted for predictive maintenance in the manufacturing industry. When facility faults or defects occur, different patterns of abnormal data are shown owing to their component behaviors. By detecting these pattern changes, it is possible to determine whether a facility abnormality occurs. This study evaluated the anomaly detection results from a combined driving system consisting of three driving motors for about six months at a manufacturing site. The learning data with an autoencoder model for about a month at the beginning of vibration data collection and continuous monitoring of anomalies using reconstruction errors showed that a component defect occurred in one driving motor, and the reconstruction error increased progressively about three months earlier than a facility manager found the failure. In addition, the micro-electro-mechanical systems sensor showed high amplitude in the entire frequency domain when high reconstruction errors occurred. However, the integrated electronics piezoelectric sensor showed different patterns as high amplitude in a specific frequency domain. The results of this study will be helpful for detecting facility abnormalities in combined driving systems using vibration sensors.

Deep learning-based fault diagnosis systems for prognostics and health management of mechanical systems is an active research topic. Notably, the absence and class imbalance of fault data (insufficient fault data compared to normal data) have been shown to cause many challenges in developing fault diagnosis systems for the manufacturing fields. Therefore, this paper presents case studies using deep learning algorithms in the absence or class imbalance of fault data. Auto-encoder-based anomaly detection method, which can be used when fault data is absent, was applied to diagnose faults in a robotic spot welding process. The anomaly detection threshold was set based on the reconstruction error of trained normal data and the confidence level of the distribution of normal data. The anomaly detection performance of the auto-encoder was verified using non-trained normal data and three sets of fault data through the threshold. As a case study for insufficient fault data, synthetic data was generated based on cGAN and applied to diagnose fault of bearing. Using the imbalanced dataset to generate synthetic fault data and to reduce the imbalance ratio, it was confirmed that the accuracy of the synthetic data generation-based 2DCNN fault diagnosis model was improved.

Due to recent development of sensor technology and IoT, research is being actively conducted on PHM (Prognostics and Health Management), a methodology that collects equipment or system status information and determines maintenance using diagnosis and prediction techniques. Among various research studies, research on anomaly detection technology that detects abnormalities in assets through data is becoming more important due to the nature of industrial sites where it is difficult to obtain failure data. Conventional machine learning-based and statistical-based models such as PCA, KNN, MD, and iForest involve human intervention in the data preprocessing process. Thus, they are not suitable for time series data. Recently, deep learning-based anomaly detection models with better performances than conventional machine learning models are being developed. In particular, several models with improved performance by fusing time series data with LSTM, AE (Autoencoder), VAE (Variational Auto Encoder), and GAN (Generative Adversarial Network) are attracting attention as anomaly detection models for time series data. In the present study, we present a method that uses Likelihood to improve the evaluation method of existing models.