Predictive Maintenance In Manufacturing: Leveraging Autoencoders And K-Means Clustering

Abstract

Predictive maintenance of manufacturing machines is essential for preventing unforeseen machine malfunctions, minimizing downtime, and improving efficiency. The traditional way of machine maintenance does not provide an efficient solution to problem detection as it involves periodic inspection or is a reactive approach. The paper provides a novel hybrid model for predictive maintenance involving autoencoders and K-Means. Autoencoders are involved in the process of extracting features and K-Means is involved in the process of anomaly detection. The sensor data on vibration, temperature, pressure, and current is pre-processed in order to solve problems related to missing values, outliers, and normalization. The high-dimensional raw sensor data is transformed into low-dimensional latent space by the use of autoencoder. Then, K-means clustering is applied to the latent features in order to cluster similar operating states and detect anomalies. The proposed hybrid approach has been compared with the other methods and shows a significant improvement with the accuracy of 95.1%, precision of 93.8%, recall of 94.5%, and F1-Score of 94.1%. The results reveal the advantages associated with using hybrid machine learning techniques within industrial settings, such as greater reliability of machines, lower maintenance costs, and increased efficiency. Further research might investigate dynamic models, multimodal data fusion, and other types of clustering methods.