AquaFusionNet: Edge Water Safety Framework

• AquaFusionNet is a hardware-aware neural framework that integrates high-magnification imaging and multivariate sensor data for real-time pathogen detection and water anomaly prediction.
• It uses a pruned SSD with a MobileNetV3-Small backbone and a compact 1D-CNN, fused via a gated cross-attention mechanism, optimized for edge devices like the NVIDIA Jetson Nano.
• Field deployments using the AquaMicro12K dataset demonstrate high detection accuracy, low false rates, and robust performance in resource-constrained environments.

AquaFusionNet is a hardware-aware neural framework for real-time pathogen detection and water quality anomaly prediction that integrates microscopic imaging and physicochemical sensor data within a single, edge-deployable pipeline. By fusing high-magnification organism-level imagery and compact multivariate sensor time-series through a lightweight gated cross-attention mechanism, AquaFusionNet enables concurrent microbial bounding-box detections and calibrated anomaly risk scoring on resource-constrained hardware such as the NVIDIA Jetson Nano. The framework is trained using the AquaMicro12K dataset, which contains 12,846 annotated microscopic images focused on drinking water contexts, overcoming previous limitations in joint microbial-sensor datasets. In field deployment across seven water facilities in East Java, Indonesia, AquaFusionNet processed 1.84 million frames, consistently achieving state-of-the-art detection and anomaly metrics while maintaining low power operation and robustness to field noise conditions (Kristanto et al., 7 Dec 2025).

1. Architecture and Fusion Mechanism

AquaFusionNet’s architecture consists of three principal modules designed to jointly model vision and sensor modalities. The vision branch employs a structured-pruned Single Shot Detector (SSD) with a MobileNetV3-Small backbone (width multiplier $\alpha=0.75$), resulting in a 38% reduction of channels via layer-wise sensitivity analysis. This backbone feeds a multi-scale SSD feature pyramid (levels P3–P7) for detection of submicron biological targets and larger particulate matter. The detection head applies convolutional predictors across pyramid levels and executes non-maximum suppression (NMS) during inference.

The temporal branch, AquaTemp-Net, receives multivariate sensor input $S_t \in \mathbb{R}^{6\times 60}$—pH, turbidity, TDS, temperature, DO, and ORP—sampled at 1 Hz over a 60-second window. It utilizes a compact residual 1D-CNN:

• $$x_1 = \text{Conv1D}_{(7,64)}(S_t) \rightarrow BN \rightarrow \text{ReLU}$$
• $x_2 = [\text{ResBlock}_{128}] \times 2 (x_1)$
• $x_3 = [\text{ResBlock}_{256}] \times 2 (x_2)$
• $f_t = \text{GlobalAvgPool}(x_3) \in \mathbb{R}^d$

Fusion occurs via a gated cross-attention mechanism, projecting global visual ($f_v$) and temporal ($f_t$) descriptors into a common query-key-value space:

• $Q = W_Q f_v$, $K = W_K f_t$, $V = W_V f_t$, $A = \text{softmax}(QK^\top/\sqrt{d_k}) V$ A gating vector $g = \sigma(W_g[f_v;f_t]+b_g)\in\mathbb{R}^d$ balances visual and sensor influences:
• $f_\text{fused} = g \odot A + (1-g) \odot f_v$ Final outputs are obtained via an MLP and sigmoid for anomaly score prediction in $[0, 1]$, with bounding-box detections and class scores extracted from the SSD head. Alerts are triggered when risk $>\tau_r$ ($\tau_r=0.7$) or any class score $>\tau_c$ ($\tau_c=0.6$).

2. Edge Deployment and Hardware Co-Design

AquaFusionNet is meticulously optimized for the constraints of edge devices such as the NVIDIA Jetson Nano (4 GB RAM, thermal envelope, solar-charged 12 Ah LiFePO$_4$ battery). The pruned and INT8-quantized model occupies 8.7 MB, containing 8.7 million parameters, and achieves a maximum inference throughput of 41 FPS (max-N mode) with a typical power draw of 4.8 W. Key optimizations include:

• Aggressive backbone pruning fitting within the memory footprint
• Employment of shallow 1D-CNNs for time-series sensor data, as opposed to heavier transformers
• INT8 quantization calibrated on representative samples to maintain accuracy

Continuous operation at 1–2 FPS is feasible, ensuring bandwidth for logging, wireless connectivity, and intermittent processing bursts, critical in solar-battery-powered field deployments.

3. Training Paradigm and Dataset Construction

AquaFusionNet is jointly trained for the detection and anomaly scoring tasks using the AquaMicro12K dataset, which comprises 12,846 RGB micrographs (1,000× magnification) from a diverse array of raw and treated drinking-water samples. The classification schema encompasses eight categories: E. coli, total coliform, P. aeruginosa, Enterococcus, Giardia lamblia, microplastics, algae, and inorganic particles. Annotations were produced by three microbiologists (Fleiss’ $\kappa=0.89$), with consensus majority voting. Sample splits were 70% training, 15% validation, 15% test. Data preprocessing included 416×416 px resizing, geometric augmentations (flips, crops), Gaussian blur, and HSV color jitter.

The loss function was $$\mathcal{L} = \mathcal{L}_\text{det} + \lambda \mathcal{L}_\text{anom}$$ with standard SSD detection loss (smooth-$L_1$ for localization, focal loss for classification) and binary cross-entropy for anomaly labels, $\lambda=1.0$. Model optimization employed Adam (lr=$10^{-3}$, cosine decay), batch size 32, for 200 epochs in FP32 precision, alternating pruning and fine-tuning rounds, followed by post-hoc INT8 quantization.

4. Evaluation and Comparative Analysis

On the AquaMicro12K test set (INT8 models on Jetson Nano), AquaFusionNet achieved:

• $[email protected] = 94.8\%$
• Anomaly prediction accuracy = $96.3\%$ (precision $95.4\%$, recall $96.9\%$, ROC-AUC $0.982$)

Comparative results versus other lightweight detectors showed superior performance at lower or equivalent power consumption:

• YOLOv5n: $89.3\%$
• YOLOv8n: $91.7\%$
• YOLOv10n: $92.4\%$
• PP-PicoDet-L: $90.8\%$
• RT-DETR-Lite: $93.1\%$

Ablation studies revealed a consistent performance uplift for full gated fusion plus pruning and INT8 quantization (vision-only $[email protected] = 91.2\%$, temporal-only anomaly accuracy $89.7\%$, late fusion $93.1\%$/ $93.8\%$, cross-attention without gate $94.3\%$/ $95.1\%$, full gated $94.8\%$/ $96.3\%$). Stress tests under real-world fouling, sensor noise, and illumination variability demonstrated that cross-modal coupling substantially reduced error rates compared to unimodal detectors.

5. Field Deployment and Practical Outcomes

During a six-month study across seven sites (five refill depots, two river intakes) in East Java, Indonesia, AquaFusionNet processed 1.84 million frames with alert accuracy benchmarked against laboratory confirmation:

• False positive rate: $2.3\%$
• False negative rate: $1.1\%$
• Uptime: $100\%$ (including grid outages with solar-battery backup)
• Average power draw: $4.8$ W

Vision-only operation induced a false positive rate of $\sim4.5\%$, and sensor-only anomaly detection suffered a false negative rate of $\sim3.8\%$ for visually evident but chemically subtle contaminations. The joint approach enabled earlier, more discriminative alerts (e.g., distinguishing bacterial flocs from inorganic turbidity) and provided calibrated risk scores with expected calibration error (ECE=$0.024$ after temperature scaling), allowing operator control at the deployment level.

6. Impact and Open Science Contribution

AquaFusionNet advances decentralized water safety infrastructure by providing a replicable, open-source end-to-end pipeline for edge device pathogen surveillance in resource-limited contexts. Its hardware/software co-design and cross-modal learning paradigm mitigate key failure modes observed in unimodal detectors and support robust, adaptive operation under environmental stressors. The open release of models, annotated datasets, and hardware schematics facilitates downstream research and deployment for similar small-scale water systems facing rapid microbial fluctuation and incomplete traditional monitoring (Kristanto et al., 7 Dec 2025).

A plausible implication is that future water-quality systems integrating joint visual/sensor models may surpass existing methods in both detection accuracy and operational resilience, especially in the context of decentralized and low-resource applications.