Predictive Maintenance Analysis for Industries

Abstract

In this paper, we are focused on deriving conclusions from sensor parameter data that would enable the detection of potential faults and the prediction of failures. We used Random Forest, Decision Tree, Naive Bayes, Logistic Regression, Support Vector Machine, and Long Short-Term Memory models to predict faults for sensor data. This analysis, which predicts the failure, has been examined through the pump sensor dataset from Kaggle. It is a binary classification problem, and it performs time series analysis using historical pump sensor data to predict future observations and classify them into a positive label (normal) or a negative label (broken). The pump system must be in perfect condition to ensure continuous power supply. A failure of one of the pumps in the system can lead to a temporary drop in power generation and even a complete outage. This may be avoided if failures are predicted in advance. Therefore, it is important to anticipate failure early to avoid large financial losses. Predictive maintenance is beneficial for industries to prevent these faults and losses. Despite expectations, the Random Forest algorithm outperforms LSTM, followed by Decision Trees. Support Vector Machine and Naive Bayes algorithms show inferior performance compared to Random Forest and LSTM.