Real-time Predictive Maintenance in Manufacturing using Streaming Sensor Data and Machine Learning

Abstract

The rationale for the new framework of real-time predictive maintenance is that it directly addresses some of the major shortcomings of traditional maintenance methods; in particular, those that utilize reactive and scheduled routines. These normal practices often result in unplanned failure or unnecessary repairs, both of which increase operating costs and reduce the overall output of the system. With the integration of industrial-type sensors and real-time continuous monitoring, there is now the ability to collect accurate data from high-value equipment, such as motors, pumps and compressors. The data that can now be collected includes vibration, temperature, pressure, and electrical current, providing a complete picture of the condition of the equipment, and such data serves as a fundamental data source for the use of predictive analyses. The ability to utilize streaming data analytics through platforms such as Kafka Streams and Apache Spark Streaming allows for the real-time ingestion and processing of data. This means that the capability of doing window-based feature extraction and producing near-instantaneous predictive results is necessary for effective early detection or prevention of failures. Additionally, through the use of machine learning models, the reliability of the predictions of failure can be enhanced. For example, Long Short-Term Memory (LSTM) models are particularly effective for predicting Remaining Useful Life (RUL) by examining temporal relations, and utilizing Random Forest and anomaly detection methods can improve the accuracy of classifying faults. This use of predictive models will enable the ability to quickly detect short-term anomalies and foresee long-term failures. The dual implementation method - edge-based and cloud-based - brings great benefit to the structure of the framework. Edge computing works to support real-time, low-latency decision-making by processing data closer to where the data is created, whereas the cloud provides scalability, storage, and higher-level analytics. Together, these two forms form a hybrid architecture that can meet the needs of Industry 4.0 Smart Manufacturing Systems. The experimental validation of the framework against a “simulated” dataset that accurately represents a multitude of “real-world” conditions further provides credibility to this framework. The efficiency of the prototype confirms that the proposed framework will reduce both unplanned downtimes and maintenance costs while also improving the reliability and accuracy of prediction results.