Eventos Anais de eventos
COBEM 2023
27th International Congress of Mechanical Engineering
Fault detection in rotating machines using LSTM and time series analysis
Submission Author:
Daniel Awada Elarrat Canto , PR
Co-Authors:
Daniel Awada Elarrat Canto, Maurizio Radloff Barghouthi, Leonardo Kaucz, Eduardo Márcio de Oliveira Lopes, Giuliana Sardi Venter
Presenter: Daniel Awada Elarrat Canto
doi://10.26678/ABCM.COBEM2023.COB2023-1198
Abstract
The need to reduce the downtime in machines has led to increased interest in process optimization in various fields, including mechanical engineering. One area of focus is predictive machine health analysis, which utilizes machine learning to predict when a serious failure may occur in a machine and minimize downtime. While predicting machine failures is a complex task, recent advances in machine learning have made this method more accessible and attracted growing interest from industries and researchers seeking to improve maintenance practices. This study focuses on the detection of faults in rotating machines caused by unbalance using machine learning. Experimental data was collected on a rotating machine test bench DAC VAD 203 under different conditions, with rotating speeds varying from 300rpm to 3300rpm and a forced unbalance faults ranging from 6.80 g to 20.4 g that, combined, resulted in 266 different experiments. Additional 242 experiments were synthetically generated by interpolating data with the same rpm and different forced unbalance, creating a new set of data that simulates the behavior of a machine when its unbalance intensity is changing. The data was collected using a triaxial accelerometer positioned on top of a selected ball bearing of a rotating system and then processed into time series by windowing the raw sensor data into equal time intervals of 0.5 seconds and extracting the time and frequency-domain features for each time segment, allowing the visualization of fault occurrences and analysis over time. A composed model consisted of a layer of a recurrent neural network and a regressor model was chosen and trained with the dataset in order o automate the task of fault and intensity detection. The recurrent model used was the Long-Short Term Memory (LSTM), as it has a good performance to make prediction based on time series. The model was tuned using an optimizer based on the Gradient Descent algorithm in Python, which iteratively adjusts the weights and biases of the model to minimize the loss function and improve the accuracy of its predictions. An margin of error lower than 6.8 g in unbalance is expected, with a short processing time, based on preliminary results. Comparing with other strategies, it is expected that this new method can significantly outperform previous approaches, highlighting the importance of tailored optimization strategies for complex machine learning models.
Keywords
machine learning, rotating dynamics, synthetic data, Process monitoring, Recurrent Neural Networks
