HybridSolarNet: Fault Detection & Energy Conversion

• HybridSolarNet is a dual-paradigm framework that combines a lightweight deep learning model for real-time solar fault detection with a hybrid solar energy conversion system.
• It employs an EfficientNet-B0 backbone enhanced by CBAM and uses focal loss with cosine annealing to effectively address class imbalance and optimize training.
• The energy conversion module integrates photovoltaic cells, photon-enhanced thermal field emission, and a Stirling engine to achieve a total conversion efficiency of approximately 35–40%.

HybridSolarNet encompasses two distinct high-impact research paradigms: (1) a lightweight, explainable deep learning architecture for real-time solar panel fault detection, and (2) an advanced hybrid solar energy conversion system that integrates spectral splitting, photovoltaic cells, photon-enhanced thermal field emission, and a Stirling engine. Both implementations demonstrate the optimization of efficiency and applicability in their domains: edge-computable visual fault detection (Hossain et al., 6 Jan 2026) and multi-modal solar energy harvesting (Nishchenko et al., 2020).

1. Lightweight Deep Learning Model for Fault Detection

HybridSolarNet’s vision-based module targets the problem of accurate, real-time solar panel fault classification under constraints suitable for UAV or edge deployment (Hossain et al., 6 Jan 2026). The architecture fuses the EfficientNet-B0 backbone with a Convolutional Block Attention Module (CBAM) for spatial and channel attention targeting.

EfficientNet-B0 Backbone:

• Input size: $380 \times 380 \times 3$
• Network Stem: $3 \times 3$ Conv (stride 2, 32 channels) $\to$ BatchNorm (BN) $\to$ Swish
• MBConv blocks: Depthwise separable convolutions with SE, as in EfficientNet-B0, expanding channel depth (24, 40, 80, 112, 192, 320, final 1280)
• Feature extraction culminates in shape $\mathbb{R}^{1280 \times 12 \times 12}$

CBAM Integration:

• Applied post-final EfficientNet-B0 Conv
• Channel Attention: Global average and max pooling $\to$ shared MLP (reduction $r=16$) $\to$ sigmoid scaling of $F$, the feature map
• Spatial Attention: $7 \times 7$ Conv over concatenated channel-refined features (from avg/max pooling) $\to$ sigmoid spatial mask

Classifier Head:

• Global Average Pooling $\to$ 1280-dim
• Dropout $(p=0.4)$
• Fully Connected (1280, 6) $\to$ Softmax over six classes (Bird-drop, Clean, Dusty, Electrical-damage, Physical-damage, Snow-covered)

Layerwise Flow:

1. Input: $380\times380\times3$
2. Stem Conv $\to 190\times190$
3. MBConv Blocks (EfficientNet-B0 sequence) $\to 12\times12\times1280$
4. CBAM attention
5. GAP $\to$ Dropout $\to$ FC(6) $\to$ Softmax

2. Training Protocols and Loss Optimization

Loss Function:

Focal loss is employed to counter the inherent class imbalance:

$$L_{\text{focal}} = -\alpha_t (1-p_t)^{\gamma} \log(p_t)$$

with $\gamma=2.0$ and uniform class weight $\alpha_t=1.0$.

Learning Rate Scheduling:

A cosine annealing schedule controls convergence:

$$\eta_t = \eta_{\min} + \frac{1}{2}(\eta_{\max} - \eta_{\min})\big[1 + \cos(\frac{T_{\text{cur}}}{T_{\max}}\pi)\big]$$

with $\eta_{\max}=1\times10^{-4}$, $\eta_{\min}=1\times10^{-6}$, $T_{\max}=25$ epochs. Scheduler is restarted once.

3. Data Handling and Validation

Split-before-Augmentation:

• Raw dataset split by stratified sampling: train 70%, validation 15%, test 15%; no augmentation leakage into validation/test.
• Only the training set is subject to augmentation (random flips, $20^\circ$ rotations, color jitter).

5-Fold Stratified Cross-Validation:

• Each fold: 1000 images/class, preserving class balance.
• Protocol: select one test fold, one validation, three training; run, repeat for all held-out test combinations.
• Final metrics: mean $\pm$ standard deviation across folds.

4. Performance, Efficiency, and Comparison

Metrics on Kaggle Solar Panel Images Dataset (5-Fold Mean $\pm$ Std):

Model	Accuracy	F1-Score	FPS	Size (MB)
HybridSolarNet	92.37% ±0.41	0.9226±0.0039	54.9	16.3
EfficientNet-B0	90.84%	0.9072	57.8	15.5
VGG19	87.79%	0.8780	39.9	532.6
MobileNetV3	86.26%	0.8593	59.0	16.2
ResNet50	83.97%	0.8391	43.6	89.9
Custom CNN	78.63%	0.7853	56.5	5.0

HybridSolarNet surpasses VGG19 by 4.6% in accuracy, is over 32$\times$ smaller by storage, and achieves higher inference throughput.

Ablation:

Addition of CBAM improves accuracy by +1.53%. Focal loss enhances minority-class recognition in imbalanced scenarios.

5. Inferencing, Deployment, and Hardware Suitability

• Inference speed: 54.9 FPS (NVIDIA RTX 3060, batch size 32)
• Model size: 16.3 MB (weights only)
• Designed for real-time UAV or edge deployment: model fits within flash/storage/memory constraints typical of embedded systems (e.g., Jetson Nano), maintaining low latency and power draw appropriate for aerial inspection workloads.

6. Explainability: Visual Focus and Saliency

Grad-CAM Analyses:

• Grad-CAM applied to post-CBAM convolutional features, providing class-specific saliency maps.
• Observed effect: HybridSolarNet’s activations correspond to defect regions (e.g., cracks, snow patches, bird droppings), avoiding spurious focus on image corners or watermarks—a limitation observed in VGG19.
• Example: For “Physical-damage,” Grad-CAM highlights linear/jagged micro-crack patterns; for “Bird-drop,” locates discrete splatter regions.
• Combined CBAM and Grad-CAM evidence supports deployment trustworthiness by confirming localization on semantically relevant features.

7. HybridSolarNet for Solar Energy Conversion

In a distinct context (Nishchenko et al., 2020), HybridSolarNet refers to a spectral-hybrid solar energy conversion platform integrating photovoltaic, thermal field emission, and Stirling-cycle conversion in a single system.

System Architecture:

• Incident solar flux concentrated by a parabolic dish/Fresnel lens onto a beam-splitting (dichroic) filter;
• Visible light $\to$ PV cells; UV $\to$ photon-enhanced, nano-structured cathode for thermal field emission (TFE); IR $\to$ cavity-type Stirling engine.

Key Subsystems:

• Photovoltaic Module: Typically crystalline or thin-film Si, directly bonded or mounted, achieving 15–18% visible-band conversion efficiency.
• Photon-Enhanced Gate Electrode (TFE): Cs-filled nano-structured (e.g., MWCNT) cathode, Fowler–Nordheim tunneling, tip radii 1–10 nm, cathode $650–750^\circ$C, current densities up to 8 mA (20 mW/cm$^2$).
• Stirling Engine: Absorbs IR, Carnot efficiency up to $\eta_{\text{Carnot}} = 1-T_C/T_H$; realistic net $\eta_{\text{Stirling}}\approx18\mbox{–}25\%$.

Combined Efficiency:

$$\eta_{\text{total}}\approx \eta_{\text{PV}}(\text{visible}) + \eta_{\text{TFE}}(\text{UV}) + \eta_{\text{Stirling}}(\text{IR}) - \text{losses}$$

Typical values: $\eta_{\text{PV}}\approx15\%$, $\eta_{\text{TFE}}\approx5\mbox{–}7\%$, $\eta_{\text{Stirling}}\approx20\% \Rightarrow \eta_{\text{total}}\approx35\mbox{–}40\%$.

Scalability and Practicality:

• Retrofittable to CSP dishes (0.5–2 m) or rooftop Fresnel arrays.
• All three energy conversion modalities operate simultaneously, maximizing output per unit area and ensuring power continuity across varying solar conditions.

8. Outlook and Technical Significance

HybridSolarNet constitutes a reference architecture for both advanced computer vision in solar O&M and as a platform for multi-modal solar energy conversion. For computer vision, it sets a benchmark for explainable, efficient inference suitable for edge-AI in field environments (Hossain et al., 6 Jan 2026). In energy conversion, the system architected with high-performance spectral management and nano-structured TFE modules achieves system-level efficiencies that approach theoretical multi-junction cell limits without reliance on fragile materials stacking (Nishchenko et al., 2020). Key open directions include the large-scale integration of nano-emitter cathodes, optimization of dichroic beam-splitters, extension to perovskite–PV hybridization, and long-term stability under outdoor conditions. Both variants of HybridSolarNet reflect the growing convergence between state-of-the-art machine learning and multi-physics energy system engineering.