Notifying problems of a machine by using Machine Learning

In 21st century, mass production and industrial machines are required to operate for 24 hours. The swift response is essential in case of emergency or predicted failure happening in the manufacturing line. The complication of a system obscures the indication of problems. Thus, developing a specific program that can detect machine's status and notify problems is the first major step to decrease downtime. Furthermore, the highly efficient predicted notification can be archived by Machine learning. In this paper, a notification software to identify problem and monitor the system is developed based on Neural Network. The Neural Network data collection is obtained from sensors, which measured the linear acceleration and angular speed of the machine. 323 datasets, categorized into 11 states, are collected from one testing machine. Training and testing sets are randomly separated by 80/20 percent respectively. To evaluate the best prediction model, input datasets are ranked from the highest to lowest frequency and chosen in 3 situations, using all input datasets, using only the top 2 up to 10 points of the highest frequency, and using only the top 2 up to 10 points of the highest and lowest frequency. Moreover, the Neural Network setups are analyzed for each situation by the following configurations: 1 to 3 hidden layers with 100 to 500 nodes. In conclusion, the best evaluation result of 88.52 % is achieved by using the top 9 points of the highest and the lowest frequency, training with 3 hidden layers and 400 or 500 nodes of neural network.