Autoencoder-Based Unsupervised Anomaly Detection for Industrial Production Line Sensor Data

I. INTRODUCTION

Novel and compound fault variants routinely evade threshold-based production-line monitoring systems, motivating unsupervised approaches that learn a model of normal operating behaviour and flag deviations, without requiring labelled examples of every fault category in advance.

II. METHODOLOGY

A five-layer deep autoencoder was trained exclusively on benign sensor records from a multi-station assembly-line dataset, using 78 statistical sensor features per record, with the reconstruction error threshold calibrated on a validation split of normal operation to achieve a target 1 percent false-alarm rate, then evaluated on a held-out test set containing five previously unseen fault categories including bearing wear and misalignment.

III. RESULTS AND EVALUATION

The autoencoder achieved an F1-score of 0.93 across the five held-out fault categories at the calibrated threshold, compared with 0.85 for an isolation forest baseline and 0.81 for a one-class SVM baseline trained on the identical feature set, with the largest relative advantage observed for slow-onset degradation patterns that departed only subtly from normal operating statistics.

IV. CONCLUSION

Autoencoder-based reconstruction error provides a more sensitive unsupervised anomaly signal than tree-based or kernel baselines for detecting previously unseen production-line fault patterns. Future work will evaluate detection latency and threshold drift under live production conditions.

V. REFERENCES