Year: 2026
Type: article
Abstract: This study presents a data‐driven deep‐learning framework for predictive maintenance in industrial equipment. The framework integrates two complementary architectures: an unsupervised autoencoder‐based anomaly‐detection model for unlabeled industrial data and a supervised autoencoder–feedforward deep neural network (AE‐FFDNN) model for fault classification in labeled datasets. The principal contribution of this work is the development of a unified dual‐mode framework that supports both unsupervised anomaly detection for unlabeled industrial data and supervised fault classification for labeled datasets within a single predictive‐maintenance architecture. In the unsupervised stage, the autoencoder learns the normal operational behavior of industrial equipment and identifies abnormal conditions based on reconstruction‐error analysis. Experimental evaluation on the refinery compressor dataset demonstrated strong anomaly‐detection capability, achieving a pseudo‐labeled anomaly‐classification accuracy of 99.32%. Since no independently verified fault labels were available, this result reflects the internal consistency of the unified dual‐mode framework rather than validation against confirmed fault events. In the supervised stage, latent features extracted by the autoencoder were used as inputs to a deep neural‐network classifier for predictive fault classification across compressors, pumps, and turbines. To improve robustness under imbalanced industrial conditions, class‐weighting and regularization techniques were employed. Experimental results demonstrated classification accuracies of 98.45% for compressors, 97.75% for pumps, and 98.58% for turbines. Furthermore, a comparative benchmark analysis using stratified 5‐fold cross‐validation was conducted against conventional machine‐learning and deep‐learning models, including random forest, XGBoost, and standard multilayer perceptron (MLP) architectures. For compressor fault classification, the proposed AE‐FFDNN framework achieved an average accuracy of 98.41%, F 1‐score of 86.25%, and AUC of 99.58%, demonstrating highly stable and competitive performance across different operating conditions. Overall, the unified dual‐mode framework supports early fault detection, enhances equipment reliability, reduces maintenance costs, and minimizes unexpected operational downtime in critical industrial processes. The proposed methodology can contribute to the development of intelligent predictive‐maintenance systems for oil, gas, and petrochemical industries.
DOI: 10.1155/jece/6014814
Institution: Islamic Azad University Kerman
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,100
OpenAlex ID: W7214486307