FusionNet: intelligent sequence fusion for predictive maintenance in edge-enabled industrial IoT systems
Predictive maintenance (PdM) in edge-enabled Industrial Internet of Things (IIoT) environments requires reliable rare-fault detection, low-latency inference, robustness to sensor degradation, and explanations that can be inspected by engineers. This paper presents FusionNet, a compact three-branch sequence-fusion architecture that combines one-dimensional convolution (Conv1D), bidirectional long short-term memory (BiLSTM), and multilayer perceptron (MLP) pathways for multivariate industrial fault classification. The revised evaluation extends the original MetroPT3 compressor study to AI4I 2020 and NASA C-MAPSS FD001–FD004 using documented binary fault or failure-risk formulations. An additional AI4I-PMDI diagnostic run was audited separately and excluded from predictive claims because the source diagnostic field remained in the model inputs after being used to derive the target. To isolate the contribution of the fusion design, the experiments compare single branches, pairwise branch combinations, full FusionNet, late-fusion probability averaging, MLP-replacement variants, compact variants, and modern lightweight sequence baselines, including Time-Series Mixer (TS-Mixer), a lightweight Transformer encoder, and a state-space-model-inspired (SSM-inspired) gated convolution model. Robustness is examined under \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$+3$$\end{document} dB Gaussian noise and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$50\%$$\end{document} channel dropout, while deployment-oriented evidence is assessed using parameter count, saved model size, latency, throughput, quantisation status, and execution-environment records. The main pipeline avoids the Synthetic Minority Over-sampling Technique (SMOTE) on flattened time-series windows; class weighting and focal loss are used as primary imbalance strategies, while flattened-window SMOTE is retained only as a labelled ablation. Integrated Gradients (IG) provides time–feature attribution maps, and all decision thresholds are selected from validation data using the geometric mean (G-Mean) prior to test evaluation. The experiments are regenerated from raw datasets on a high-performance computing (HPC) environment, with system specifications, thresholds, outputs, and reproducibility manifests recorded for verification. The results position FusionNet as a compact and interpretable candidate for sequence fusion in binary PdM evaluation. The reported latency and footprint measurements support deployment-oriented comparison under the tested HPC environment but do not constitute validation on a physical edge device.