Detecting Anomalies in Mining Engine Oil: A Convolutional Autoencoder Approach

Berrospi M.; Vargas-Machuca J. · Proceedings of the 2025 IEEE 32nd International Conference on Electronics, Electrical Engineering and Computing, INTERCON 2025 (2025) · DOI: 10.1109/intercon67304.2025.11244639

Resumen

This paper presents a novel deep learning methodology for automated anomaly detection in mining engine oil analysis using Convolutional Autoencoders (CAE). Traditional oil analysis depends heavily on expert judgment and fixed thresholds, which can cause subjective decisions and delays in failure detection. The CAE model was trained exclusively on 1,320 normal oil samples, each containing 32 physicochemical variables including wear metals, additives, contaminants, and degradation products. Data were normalized using Min-Max scaling and reshaped into 8x4x1 matrices to leverage the spatial feature extraction capabilities of convolutional layers. The unsupervised learning approach enables the model to learn compressed representations of healthy oil characteristics without requiring labeled failure data. During training, the model demonstrated stable convergence with consistent reduction in reconstruction loss, indicating successful learning of normal pattern distributions. Anomaly detection was performed by calculating mean squared error (MSE) between original and reconstructed samples, with a threshold established using the 95th percentile of training error distribution. The methodology was validated using 1,346 failure samples labeled by specialists, achieving an accuracy of 81.87%. The approach provides several key advantages: elimination of dependency on scarce failure data through unsupervised learning, multivariate analysis capturing variable correlations often missed by traditional methods, variable-level error analysis enabling identification of specific parameters contributing to anomalies, and scalability for integration into real-time monitoring systems. Results demonstrate that the CAE successfully reconstructs normal samples with low error while producing significantly higher reconstruction errors for anomalous conditions, enabling early detection of potential mechanical failures before visible symptoms manifest.

Instituciones: Universidad Nacional de Ingeniería