PREDICTION OF BEARING REMAINING USEFUL LIFE USE DEEP LEARNING METHODS

• Main Author: Dwi Juniar, Heni
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/62248
• Institution: Institut Teknologi Bandung

Bearings are an important part of industrial rotating machines and their operating status directly affects machine performance and has a significant impact on machine and operator safety. With operation at certain loads and environmental factors cause the reliability of bearings to decrease over time. It can lead to considerable economic losses. In order to ensure the reliability and safety of bearings, proper maintenance is necessary. Prognostic Health Management (PHM) is a detection, diagnostic, and prognostic system to help in determining the efficient, suitable, and appropriate maintenance strategies implementation. PHM technology provides the basis for Predictive Maintenance (PdM) on industrial components and systems. In simple terms, predictive maintenance can be interpreted as maintenance based on the condition of the machine itself. There are several methods used in the PdM strategy, including vibration analysis, acoustic emission, oil analysis, particle analysis, corrosion monitoring, thermography, and performance monitoring. From these several methods, vibration analysis is one of the methods used to monitor the mechanical condition of equipment whether it is in normal condition or experiencing a decrease in operational capability. Vibration analysis can be used to predict the remaining useful life (RUL). The remaining useful life is the length of time the machine is in operating condition before requiring repair or replacement. The remaining useful life (RUL) prediction on equipment is used to determine the reliability and help in decision making on doing equipment maintenance to avoid downtime. Several methods can be used to predict the remaining useful life, such as the physics-based method, data-driven method, and hybrid method. The use of data-driven methods, including deep learning methods, has attracted many people. In recent years, research on the use of deep learning for representational learning, time series classification and prediction has received a lot of attention. Deep learning methods can handle large amounts of data with high accuracy. Several methods have been used for the development of deep learning to predict RUL. There are three categories of models that can be used in deep learning methods: generative models, hybrid models, and discriminatory models. The Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology was used in this study for developing the model and vibration analysis. At the business understanding stage, the objective of this research was defined, to predict the remaining useful life of the bearings to find out the right maintenance schedule. At the data understanding stage, the type, structure, and quality of the data used were described. The object of training data and validation data used was vibration data from bearings with three different operational conditions. The first condition had a load of 1800 rpm and a rotational speed of 4000 N. The second condition had a load of 1650 rpm and a rotational speed of 4200 N. The third condition had a load of 1500 rpm and a rotational speed of 5000 N. Pre-processing stage consisted of the data load process, the data transformation by Fast-Fourier Transform (FFT), and feature selection. RUL prediction modelling was done by using deep learning method. The deep learning methods used were convolution neural network (CNN) and Long Short-Term Memory (LSTM). After that, analyzing the accuracy of each model using a metric model and calculating the score for each model were done in order to determine the accuracy of each model. There was two inputs to study features or patterns in each direction, vertical accelerometer and horizontal accelerometer. Based on the evaluation of each model using regression model metrics, LSTM shows smaller error value than CNN to predict the remaining useful life of the bearing. The LSTM model has MAE = 0,16 and RMSE = 0,21, meanwhile for CNN, the value of MAE = 0,31 and RMSE = 0,39. Those values were obtained by using a metric model for the model as training data and validating results data. The LSTM model shows score of 0,73 with a maximum value of 1, this indicate that LSTM model was accurately enough to prove that the model can predict RUL. Besides, the score was obtained by CNN model was 0,43 with a maximum value of 1, this score indicate the less accurate for the model in predicting RUL. In addition, the time that required for the LSTM model to conduct the training model was 30 seconds with the number of epoch used was 15, while the CNN model took 100 minutes in the model training process with the number of epoch used was 30 epochs.