부산대학교 도서관

The widespread availability of sensing technology, such as detectors, has led to the generation of enormous volumes of time-series data by equipment in manufacturing warehouses and industries. There is a ton of knowledge accessible that may be put to use in order to forecast the failure of machinery and the decline of its quality in a particular plant. The downtime of manufacturing machinery is responsible for significant monetary losses, which may be mitigated by accurately predicting when the system will cease to function by analyzing sensor data. Actual data collection from sensors is now technically viable because to the advances that went into creating the Internet of Things (IoT). Our research has shown that combination research has the potential to provide reliable estimates since it is able to accurately represent the abstract features that are essential to the development of more consistent reading. This makes hybrid modeling a useful tool. In addition, it might be challenging to establish an efficient optimization approach due to the complicated nature of the many sensor data that is seen in real-time circumstances. A technique for nonlinear time prediction for predictive maintenance (PdM) is proposed in this study. The method utilizes a combination of convolution operation, extended memory recall, and skips connections (CNN-LSTM). For the purpose of predicting when machines may break down, we try out several forecasting methods, one at a time, including CNN, LSTM, and CNNLSTM. The data were taken from the case study conducted by Microsoft and utilized in this research. The database contains information on the history of breakdowns, the service background, the error circumstances, as well as the machinery attributes and telemetry, which includes information such as the output power, stress, motion, and roster accelerometer gathered during years 2015 and 2016. The combination CNNLSTM architecture which has been suggested is a program that integrates two distinct end-to-end techniques. In this particular model, the Long Short-Term Memory (LSTM) is put to use to analyze the correlation between various time analysis different factors by making use of its superior memory, and 1-D Convolutional Neural Networks (CNNs) are inclined to take on the task of retrieving raised qualities from of the information in an effective manner. The lengthy structures of the time series are learned by our technique, which does this by isolating the short-term dependence patterns that exist between the multiple factors of the time series. According to our findings, CNN-LSTM was the method that produced the most dependable results and the best forecasting reliability.