Advanced technology to predict equipment health condition

• Main Author: Yang, Chule
• Other Authors: Wang Dan Wei
• Format: Final Year Project
• Language: English
• Published: 2014
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Computer hardware, software and systems
• Online Access: http://hdl.handle.net/10356/60379
• Institution: Nanyang Technological University

Objective: Maintenance was an important issue in almost all manufacturing companies. In order to keep equipments working in good condition and reduce maintenance expense for better profit, modern factories had always been seeking a effective way to take place of the traditional maintenances. This dissertation aims at building a special approach to analyze machinery log database and help the maintenance engineers make proper and efficient decision when should they do the maintenance work. Methods: Considering the great evolutions of Artificial Intelligence and the developments of MATLAB Toolbox these years, Artificial Neural Network was chosen as my project method. I used BP Algorithm to build the net model for Parameter Based Approach and Recipe Based Approach respectively, divided the industrial data into training group and testing group. Next, input the training data and used two kind of training functions(Gradient Descent method and Levenberg-Marquardt method) to train the net and applied the best net parameters to test the testing data. Results: We could see from the output data that the prediction results are highly acceptable, the maximum training and testing accuracy of both Parameter Based Approach and Recipe Based Approach reached around 94%. Besides, Gradient Descent method generated a higher accuracy in Recipe Based Approach while Levenberg-Marquardt method generated a higher accuracy in Parameter Based Approach. Conclusion: Neural Network is an effective method to predict the occurrence of equipment failure. But it still depends on the structure of the dataset. For instance, the number of pattern, the form of output and whether the dataset is balanced. These issues will be worked out in my future work.