← Back To Current Issues

A Transformer-Based Framework for Early Fault Detection in Industrial Sensor Networks

Lucas A. Verhoeven<sup>1,2,#</sup>, Priya N. Raghunathan<sup>1,#</sup>, Samuel K. Adeyemi<sup>1,2</sup>, Hana Lindqvist<sup>1</sup>, Marco T. Ferreira<sup>1,2,*</sup>

<sup>1</sup>Department of Intelligent Systems and Industrial Analytics, Lakemont University of Technology, Brevik, Norway; <sup>2</sup>Laboratory for Predictive Maintenance Research, Lakemont University of Technology, Brevik, Norway

Correspondence to: Dr. Marco T. Ferreira, Department of Intelligent Systems and Industrial Analytics, Lakemont University of Technology, Brevik, Norway; Email: m.ferreira@lakemont.example.edu

ABSTRACT

Early detection of equipment faults from multivariate sensor streams remains a central challenge in industrial analytics, particularly when static machine metadata must be analyzed alongside continuously recorded signals. This study proposes a transformer-based architecture, combining a sensor embedding layer, a multi-head self-attention encoder, and a metadata fusion head, for joint fault classification across heterogeneous inputs. The proposed model was evaluated against three established baselines (support vector machine, random forest, and LightGBM) on a combined metadata and sensor-stream benchmark of 62,300 records. The proposed model achieved 96.1% accuracy and an F1-score of 0.953, outperforming the strongest baseline (LightGBM) by 4.7 percentage points, at a modest increase in inference time (11.8 ms versus 7.2 ms). Ablation analysis showed that removing the transformer encoder reduced accuracy by 6.8 points and restricting the model to single-head attention reduced accuracy by 3.1 points, indicating that both sequence-level attention and multi-head feature weighting contribute meaningfully to overall performance. These results suggest that transformer-based architectures offer a practical and accurate approach for fault detection tasks involving mixed static and sequential data.

Keywords: Fault detection · Predictive maintenance · Transformer · Self-attention · Industrial sensor networks · Deep learning

Scroll to Top