\documentclass[11pt,a4paper]{article}

\usepackage[utf8]{inputenc}
\usepackage{geometry}
\geometry{
    a4paper,
    left=25mm,
    right=25mm,
    top=25mm,
    bottom=25mm
}
\usepackage{booktabs}
\usepackage{tabularx}
\usepackage{amsmath}
\usepackage{setspace}
\usepackage{caption}

\begin{document}

% ==========================================
% HEADER 
% ==========================================
\begin{center}
    {\Large \textbf{SUPPLEMENTARY APPENDIX}} \\[0.5cm]
    {\large \textbf{Hybrid Deep Learning Framework for Predictive Maintenance in Industrial IoT Systems: A Multi-Sensor Fusion Approach with Explainability}} \\[0.5cm]
    \textbf{Syed Eirfan Atthar} \\
    Department of Mechanical Engineering, Dibrugarh University, Assam, India \\
    Email: 20wj1a03b1@gniindia.org
\end{center}

\vspace{1cm}

\noindent In this supplementary document, extended technical specifications regarding the public benchmark datasets and the hyperparameter optimization space utilized in the primary manuscript are provided. These data matrices and configuration parameters are appended to ensure complete methodological reproducibility.

% ==========================================
% DATASET SPECS
% ==========================================
\section*{S1. Extended Dataset Specifications}

\subsection*{S1.1. NASA C-MAPSS Turbofan Engine Dataset}
The Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset was utilized for continuous Remaining Useful Life (RUL) regression. The specific operational boundaries and fault modalities for the evaluated sub-datasets are quantified in Table S1.

\begin{table}[htbp]
\centering
\caption*{Table S1: C-MAPSS Sub-Dataset Operational Specifications}
\begin{tabular}{lcccc}
\toprule
Subset & Train Trajectories & Test Trajectories & Operating Conditions & Fault Modes \\
\midrule
FD001 & 100 & 100 & 1 (Sea Level) & 1 (HPC Degradation) \\
FD004 & 249 & 248 & 6 (Varying Altitude/Mach) & 2 (HPC \& Fan Degradation) \\
\bottomrule
\end{tabular}
\end{table}

\subsection*{S1.2. IEEE PRONOSTIA Bearing Dataset}
The PRONOSTIA dataset was employed for discrete fault classification. The accelerated degradation tests were conducted on a dedicated physical testbed. Vibration signals were sampled at a high frequency of 25.6 kHz. The specific physical loading conditions utilized for the validation of the proposed architecture are detailed in Table S2.

\begin{table}[htbp]
\centering
\caption*{Table S2: PRONOSTIA Loading and Operational Conditions}
\begin{tabular}{lccc}
\toprule
Condition Set & Radial Load (kN) & Rotational Speed (RPM) & Total Bearings \\
\midrule
Condition 1 & 4.0 & 1800 & 7 \\
Condition 2 & 4.2 & 1650 & 7 \\
Condition 3 & 5.0 & 1500 & 3 \\
\bottomrule
\end{tabular}
\end{table}

% ==========================================
% HYPERPARAMETER SPACE
% ==========================================
\section*{S2. Hyperparameter Optimization Space}
To mitigate network overfitting and ensure mathematical convergence, hyperparameters were not statically assigned. Instead, a Bayesian optimization protocol was executed across the training phase. The predefined search space boundaries and the final selected parameters for the HEDLF architecture are specified in Table S3.

\begin{table}[htbp]
\centering
\caption*{Table S3: Bayesian Optimization Search Space and Final Selection}
\begin{tabular}{lllc}
\toprule
Hyperparameter & Search Space & Distribution & Final Selected Value \\
\midrule
Learning Rate & $[10^{-4}, 10^{-2}]$ & Log-Uniform & $1 \times 10^{-3}$ \\
Weight Decay & $[10^{-5}, 10^{-3}]$ & Log-Uniform & $1 \times 10^{-4}$ \\
CNN Filter Count & $\{32, 64, 128, 256\}$ & Discrete & $[64, 128]$ \\
BiLSTM Hidden Units & $\{64, 128, 256\}$ & Discrete & $[256, 256]$ \\
Attention Heads ($h$) & $\{4, 8, 16\}$ & Discrete & $8$ \\
Dropout Probability & $[0.1, 0.5]$ & Uniform & $0.3$ \\
Batch Size & $\{32, 64, 128\}$ & Discrete & $64$ \\
Sequence Window ($T$) & $[20, 50]$ & Discrete & $30$ \\
\bottomrule
\end{tabular}
\end{table}

\vspace{1cm}
\noindent \textit{Note: All network training operations were executed on a standardized hardware environment to ensure computational consistency. The raw data matrices corresponding to Tables S1 and S2 are publicly accessible via the respective NASA and IEEE academic repositories.}

\end{document}