TA-FaultNet: A Temporal Attention Framework with Bidirectional
LSTM for Multi-Class Fault Detection and Health Monitoring in
Industrial Wireless Sensor Networks
Massila Kamalrudin1,* Mustafa Musa2
1 Faculty of Information and Communication Technology, Universiti Teknikal Malaysia Melaka, Malaysia
2 Center of Research and Innovation Management, Universiti Teknikal Malaysia Melaka, Malaysia
Emails: massila@utem.edu.my · mustafmusa@utem.edu.my
Received: February 03, 2026 Revised: March 12, 2026 Accepted: April 08, 2026 ⋆ Corresponding author
ABSTRACT
Industrial wireless sensor networks are central to the continuous monitoring of critical plant equipment, yet reliable
identification of multiple concurrent fault modes from heterogeneous multivariate sensor streams remains an unsolved
operational challenge. Physical failure mechanisms—pump cavitation, valve blockage, gradual sensor drift—and
wireless channel disturbances each imprint distinct but overlapping temporal signatures that render classical threshold
and rule-based detectors inadequate for automated maintenance dispatch. This paper presents TA-FaultNet, a neural
architecture designed specifically for the multi-class fault identification problem in industrial sensor deployments.
The network couples a two-stage stacked bidirectional recurrent encoder with a parallel multi-head self-attention
module and a compact temporal convolutional block, enabling simultaneous capture of long-range process dynamics
and fine-grained fault-onset localisation from raw sensor windows. TA-FaultNet is evaluated on the publicly available
Skoltech Anomaly Benchmark under five operational classes and assessed through a comprehensive battery of
experiments including baseline comparisons, systematic component ablation, cross-experiment generalisation, and
progressive noise-injection testing. The proposed architecture decisively outperforms eight competing methods
spanning classical anomaly detectors, standalone recurrent and convolutional networks, and the Transformer, while
remaining lightweight enough for edge gateway deployment. Attention weight visualisations expose fault-specific
temporal activation patterns, providing maintenance engineers with interpretable diagnostic evidence beyond bare
classification labels.
Keywords: Industrial IoT Fault detection Predictive maintenance Bidirectional LSTM Multi-head self-attention
SKAB dataset Temporal convolutional network Wireless sensor networks Time-series classification