Numerous studies have addressed fault diagnosis in photovoltaic systems using image processing and machine learning. Early methods relied on thermal imaging and handcrafted features such as color histograms or texture descriptors. However, these methods lacked robustness against environmental variability and required expert-designed features.

The domain of fault diagnosis in photovoltaic (PV) systems has indeed progressed significantly, moving from classical image processing and handcrafted feature extraction techniques to modern data-driven machine learning and deep learning approaches. Early strategies, as you noted, relied on thermal imaging to detect anomalies—such as hotspots resulting from cell degradation, partial shading, dirt accumulation, or electrical faults—using features crafted by experts, like color histograms, edge detectors (e.g., Sobel, Canny), and texture descriptors such as Local Binary Patterns (LBP) and Gray-Level Co-occurrence Matrices (GLCM)[1]. While these handcrafted approaches laid the groundwork for automated PV inspection, they suffered from several intrinsic limitations: primarily, their susceptibility to environmental variability (e.g., changes in solar irradiance, ambient temperature, soiling, or atmospheric conditions) and their dependence on domain expertise for feature design, both of which reduced robustness and limited generalizability in diverse real-world scenarios[1][2].Addressing Environmental Variability and Feature DesignOne of the central criticisms of early methods was their lack of resilience to the aforementioned environmental changes. Thermal and visible light image characteristics in PV modules can be highly variable, confounding the ability of hand-engineered features to consistently discern faults under different operational contexts[1]. This limitation has driven the field towards the adoption of machine learning (ML) and, more recently, deep learning (DL) methodologies, which aim to learn distinguishing representations of faults directly from data, thereby reducing the need for manual feature engineering.Machine Learning AdvancementsTraditional ML approaches have typically leveraged features extracted from infrared thermographic images or electrical measurements (such as I-V and P-V curves) and used classifiers like Decision Trees, K-Nearest Neighbors, and Support Vector Machines[3]. These have demonstrated effectiveness for initial fault classification tasks, especially under controlled conditions. For example, Badr et al. systematically compared several ML classifiers under various climatic and operational scenarios, illustrating that while tuned ML models can achieve reasonable performance, their effectiveness is still gated by the quality and relevance of the engineered input features[3].Deep Learning and End-to-End Feature LearningDeep learning, especially using Convolutional Neural Networks (CNNs), has dramatically improved the landscape of PV fault diagnosis. CNNs are capable of autonomously learning multi-level feature hierarchies from images—thermal, electroluminescence, or even dynamically generated data representations—resulting in improved robustness against environmental variability and greater accuracy in fault detection and classification[1][2][4]. For instance, state-of-the-art studies have reported CNN-based approaches attaining classification accuracies as high as 98-99.5% in multi-class PV fault scenarios, outperforming traditional ML methods by a significant margin[1][4].

Moreover, CNNs are effective for both aerial inspection (using drone-acquired visual or thermal imagery) and for signal-based representations, such as chaos dynamic error maps or processed I-V curve images[2][4][5]. These models have excelled, not only in recognizing obvious anomalies but also in capturing subtle and localized fault signatures that traditional texture or color-based features would miss.Emergence of Hybrid and Sequential Deep ArchitecturesRecent advancements involve hybrid architectures that integrate CNNs with recurrent networks (such as Bidirectional Gated Recurrent Units, Bi-GRU, or LSTM) or optimization algorithms to handle temporal dependencies and multivariate sequence data, further enhancing the capability to distinguish among a broad set of fault types and environmental scenarios[6][7]. For example, the combination of CNN and Bi-GRU leverages both spatial feature extraction and sequential (temporal) analysis in PV monitoring data, offering higher fault recognition accuracy and resilience to noise and varying operating conditions[7].

Additionally, two-stage detection and classification pipelines have been proposed, where initial anomaly detection uses simple physical heuristics or thresholding, followed by deep learning-based fault categorization. Such methods benefit from lower false alarm rates and improved diagnostic specificity[5][7].Real-World Validation and Practical ImpactEmpirical studies frequently validate these approaches, not only through simulation but also in real operational PV power plants, confirming their scalability, efficiency, and practical utility for real-time and remote PV system management[2][5][6]. The ability of deep learning models to be retrained or fine-tuned for new fault classes as more data becomes available also addresses concerns about the adaptability and long-term reliability of PV monitoring systems.Summary Table: Evolution of PV Fault Diagnosis Approaches

ConclusionIn conclusion, the evolution from classical image processing and ML-based methods—limited by feature design and environmental variability—to deep learning-based approaches has greatly enhanced the robustness, scalability, and accuracy of fault diagnosis in PV systems. Modern deep learning models, especially CNNs and their hybrids, offer automated feature extraction, resilience to environmental disturbances, and superior performance, thus setting a new standard for intelligent PV maintenance and monitoring[1][2][4][6][7].

Fusion strategies for convolutional neural networks (CNNs) play a central role in advancing defect analysis, including fault diagnosis in photovoltaic (PV) systems. These strategies determine how multiple sources of information—such as different imaging modalities, signal types, or hierarchical feature maps—are integrated within deep learning frameworks to improve defect detection robustness, accuracy, and generalization. Effective fusion can be achieved at various stages of the network and can employ different architectures and mechanisms depending on the nature of the data and task complexity.1. Early (Data-Level) FusionEarly fusion integrates multiple modalities at the input level. For example, stacking thermal infrared and RGB images together as multichannel input tensors enables the CNN to learn relationships across modalities from the first layer. In PV fault analysis, this approach is practical if different sensor data can be spatially aligned and synchronized. Nevertheless, such raw integration may not fully exploit the specificities of each modality, and performance can degrade if modalities are not normalized or are unequally informative[1][2].2. Mid-Level (Feature-Level) FusionMid-level fusion, or feature-level fusion, combines features or intermediate representations extracted from separate network branches, each dedicated to a specific data stream or modality. Within PV fault diagnosis and similar industrial defect detection, dual or multi-stream CNNs process different input modalities (e.g., IR images, electroluminescence images, I-V curves), followed by concatenation, addition, or attention-based reweighting of feature maps at an intermediate layer[3][4].

For instance, mid-level fusion allows a network to extract modality-specific features—such as spatial structures from RGB and thermal anomalies from IR—before fusing them, boosting flexibility and overall discriminative power. Applications outside PV but within defect detection have demonstrated that concatenating low-, mid-, and high-level CNN features from different streams yields more comprehensive and hierarchical feature representations, which can drive gains in accuracy, as shown in textile surface defect detection and stress detection settings[3][4][5].3. Late (Decision-Level) FusionLate fusion integrates predictions (e.g., class probabilities, logits) from separate models operating on different data streams. Methods such as simple averaging, weighted voting, or ensemble meta-classifiers are used. This strategy is beneficial when independent models—perhaps trained on RGB versus IR images in PV systems—may need to operate on partially missing data or when computational resources allow for modular, independently-operated systems[1][6]. However, this approach forfeits learning cross-modal feature interactions.4. Hybrid Fusion StrategiesHybrid fusion architectures exploit the strengths of multiple fusion levels, often combining early/mid/late fusion. For example, a system might first concatenate synchronized images (early fusion) and then merge learned features with electrical parameters like power and voltage data at the feature level. This multi-stage approach caters to complex, heterogeneous PV monitoring scenarios, capturing both early pixel-level and abstract feature-level complementarities[3][7]:

5. Attention-Based FusionAttention mechanisms have become pivotal in fusion strategies, especially for defect analysis on industrial surfaces and, by extension, PV defects. Channel and spatial attention modules learn to prioritize the most relevant feature maps or spatial regions, dynamically reweighting contributions from different modalities or feature hierarchies. Self-attention or cross-modal attention can focus the network on critical defect cues, improving both interpretability and sensitivity to local or subtle fault patterns[4][5].

For example, Feature Pyramid Networks with attention blocks, as used in industrial defect detection, aggregate multi-scale features and recalibrate them via channel/spatial-wise attention, leading to enhanced performance in both localization and classification tasks[5].6. Multiscale and Multiview FusionIndustrial and remote sensing applications—including aerial PV system inspection—benefit from multiscale and multiview fusion. Architectures like Atrous Spatial Pyramid Pooling or ConvLSTM layer features from multiple spatial resolutions or temporal image frames, enabling the network to recognize defects manifesting at various scales or evolving over time[2][3][4]. Pixel-wise prediction (segmentation) tasks, such as surface defect detection, often employ pyramid feature fusion to unify features extracted at different depths/resolutions, achieving state-of-the-art accuracy[4].7. Cross-Model and Graph-Based Feature FusionOutside traditional CNN pipelines, combining deep CNN features with those from other models (e.g., Graph Attention Networks) via weighted fusion has shown promise in other high-dimensional, defect-related tasks like hyperspectral image classification. Here, feature fusion exploits complementary strengths (e.g., local receptive fields in CNNs, relational information in GNNs)[8][9]. Though less reported in current PV literature, such approaches may offer meaningful enhancements in future PV fault detection pipelines.

---

In summary:Fusion strategies for CNNs in defect analysis—including PV fault diagnosis—range from data-level to decision-level to hierarchical and attention-based fusion. Mid-level and attention-guided fusion are increasingly popular, as they allow simultaneous exploitation of local, global, and cross-modal relationships within raw and derived features. Hybrid and multiscale fusion strategies, when carefully matched with the nature and synchronization of available data, consistently drive superior performance, as demonstrated in PV and broader industrial defect detection literature[1][2][3][4][5][7]. As advanced PV inspection shifts toward multi-modal, real-time, and aerial surveillance systems, effective and adaptive fusion strategies will be decisive for scalable, accurate, and interpretable defect analysis.

The application of deep convolutional neural networks such as MobileNetV2, ResNet50, and EfficientNetB0 for defect analysis in photovoltaic (PV) panels has yielded robust, state-of-the-art results for both binary and multi-class fault detection and classification from visible, infrared, and electroluminescence imagery. Recent developments in this area are driven by the need for reliable PV operation and maintenance, as timely detection of defects—such as hotspots, microcracks, delamination, soiling, bypass diode faults, and breakages—directly improves performance, prevents energy loss, and extends system life [1][2][3][4].

Model Comparison in PV Defect Analysis

MobileNetV2's strengths lie in its lightweight architecture and efficient computation, making it ideal for edge deployment scenarios such as drone-based PV farm inspection or real-time monitoring on low-power devices. Its use of inverted residual blocks and depthwise separable convolutions reduce the parameter count without severely compromising accuracy. Recent work has demonstrated consistent classification accuracy between 90–94% for binary tasks and competitive multi-class performance, provided the defect types are visually distinct and dataset size is adequate. However, MobileNetV2 may underperform on highly complex or subtle multi-class scenarios compared to deeper alternatives due to limited representational capacity [1].

ResNet50 introduces residual connections to mitigate gradient degradation with increasing depth, enabling much deeper feature hierarchies that capture subtle spatial patterns in fault images. This translates into superior performance for both binary and multi-class PV defect classification, with accuracy often reported upwards of 97–98%—notably for complex inputs such as electroluminescence or thermographic images containing fine-grained texture cues. For example, a comparative study using thermal imagery of normal and defective panels reported deep CNNs (including ResNet architectures) outperforming classical ML by a wide margin for both detection (accuracy 98.71%) and fault isolation across multiple defect types [1][4]. Although ResNet50’s higher computational demand is better suited for GPU-enabled settings or cloud-based workflows, it remains deployable on high-end edge devices.

EfficientNetB0, architected through compound scaling of depth, width, and image resolution, achieves an exceptional balance of parameter efficiency and accuracy. It has been shown to match or even surpass ResNet50 in PV defect tasks while maintaining a much smaller model size and faster inference—crucial for both cloud and resource-limited edge applications. Studies have confirmed that EfficientNetB0 attains high accuracy (95–98%) across both binary and multi-class classification on thermographic and aerial PV datasets, showing strong generalization and resistance to overfitting even on moderately sized or imbalanced datasets [1]. The ability to compress EfficientNetB0 for deployment via frameworks such as TensorFlow Lite or ONNX further expands its applicability, including real-time, UAV-driven panel inspection and field diagnosis.

Data Preparation and Training Considerations

Comprehensive defect classification relies not only on model choice but also on appropriate dataset curation, preprocessing, and augmentation. Typical pipelines include acquisition of IR, RGB, or EL imagery, manual or semi-automatic annotation of defect types, and application of augmentation techniques to enhance generalization. Data normalization and careful class balancing are critical due to the often limited availability of labeled PV panel fault data [1][2][3].

Transfer learning, particularly with ImageNet-pretrained weights, is standard practice, as it enables models to converge efficiently even with moderate training sets. Fine-tuning the top layers to PV-specific characteristics and employing regularization methods (like dropout or early stopping) further minimizes overfitting risks [1][3].

Advanced Architectures and Multimodal Fusion

Although standalone CNNs such as MobileNetV2, ResNet50, and EfficientNetB0 form the backbone of most PV defect analysis, there is increasing use of fusion strategies and hybrid architectures to enhance performance. For instance, integrating hierarchical or pyramidal feature fusion, as in PGA-Net, enables multi-scale representation and improves segmentation and detection of subtle or multi-size defects on PV panels [5]. Attention mechanisms or the combination of CNNs with Graph Neural Networks, while more common in hyperspectral and industrial defect tasks, represent promising directions for improving PV module fault detection under real-world variability [6][7][8].

Practical Deployment and Trade-offs

Model selection for PV panel defect analysis often reflects the target deployment setting. MobileNetV2 best suits real-time, low-power, or in-field monitoring. ResNet50 delivers maximum discrimination in data-rich, compute-rich environments with highly variable defect types. EfficientNetB0 combines broad applicability, high accuracy, and efficient resource consumption, and is increasingly recognized as the default backbone for scalable PV fault detection workflows [1][3][9].

Summary and Outlook

Leveraging deep CNN backbones such as MobileNetV2, ResNet50, and EfficientNetB0 enables highly effective PV panel defect diagnosis for classifications ranging from simple binary healthy/faulty decisions to intricate multi-class fault isolation. EfficientNetB0, in particular, stands out by delivering near-best-in-class accuracy at a fraction of the computational burden of deeper models, without sacrificing deployment flexibility. Continued progress is promised by hybrid feature fusion, attention mechanisms, and context-aware deep architectures that further cope with small or imbalanced datasets and capture multi-level spatial features inherent in PV panel imagery [1][3][4][5][8][9].

Skin cancer detection using image fusion and feature aggregation strategies has rapidly advanced and become a cornerstone of state-of-the-art computer-aided diagnosis systems, particularly in dermoscopic image analysis. The complexity of skin lesion detection—owing to factors like intra-class variability, inter-class similarity, and the presence of image artifacts—presents significant challenges that are effectively addressed through the fusion of diverse data, features, and model outputs. Here, we critically analyze methodological trends, technical advances, and empirical findings in this domain, drawing upon recent literature and benchmarking results.

---

Image-Level FusionImage-level or data-level fusion focuses on combining raw or preprocessed images from multiple modalities or sources prior to feature extraction. The rationale is to harness complementary information that different imaging modalities (like RGB, multispectral, clinical, or synthetic GAN-generated images) offer for richer lesion representation. For example, generated images using StyleGANs have been fused with real dermoscopy data to mitigate dataset imbalance and improve the feature diversity accessible to learning-based classifiers, thus enhancing model robustness and generalizability for dermoscopy image classification [1]. Such strategies, while effective in maximizing visual cues, require careful spatial alignment and quality control to prevent noise amplification or the propagation of imaging artifacts.

---

Feature-Level Fusion and AggregationFeature-level fusion and aggregation are arguably the most prevalent and effective strategies in current skin cancer detection research. These methods involve the extraction of features from multiple sources—whether different layers within a single CNN, multiple CNN architectures, or a combination of handcrafted and deep features—and then their aggregation into a unified representation for classification or segmentation tasks.

A notable advancement involves fusion of deep features extracted from distinct CNN backbones (e.g., ResNet, NASNet, EfficientNet), leveraging transfer learning and pre-trained models to augment limited medical datasets. Mahbod et al., for instance, aggregated multi-network features using principal component analysis and concatenation, followed by SVM classification, resulting in accuracy gains across public skin cancer datasets [2]. Hybrid fusion approaches have similarly combined handcrafted features (such as texture characteristics from LBP or Haralick patterns) with deep features, further enhancing discriminatory power in challenging diagnostic settings [2][3].

Attention-based mechanisms have raised feature fusion to a new level of adaptivity. By dynamically assigning weights to features across channels or spatial regions, attention modules such as CBAM or SENet enable networks to emphasize pathologically significant areas and suppress irrelevant background or noise [4]. Recent architectures also integrate multi-scale and bi-directional feature learning to amplify lesion-boundary cues and maintain consistency across diverse image scales, with consistent gains in both accuracy and interpretability noted for segmentation and classification [4].

---

Decision-Level (Score-Level) FusionDecision-level or late fusion strategies combine outputs from multiple classifiers, each of which may be trained on different input types, feature sets, or even distinct model architectures. Methods such as majority voting, weighted averaging, or meta-ensemble stacking have been effective in improving generalization and balancing accuracy, particularly in multi-class settings for diseases such as melanoma, basal cell carcinoma, and nevus [1][2]. For example, decision fusion architectures that combine the predictions from several CNNs have outperformed individual models in the ISIC 2019 skin cancer challenge, with accuracy improvements attributed to the reduction in model-specific errors and biases [1][2].

---

Advanced Fusion Architectures and ExplainabilityFusion strategies increasingly leverage architectural innovations from adjacent fields—such as hyperspectral imaging and industrial defect detection—to address domain-specific challenges in dermatology. The adoption of multi-stream networks, pyramid feature fusion, and global context modules (analogous to PGA-Net and feature pyramid networks in other domains) enables the simultaneous capture of fine-grained local features and high-level global context, which is critical for accurate lesion boundary delineation [4][5][6].

Interpretability also remains a focal point in recent work, with attention and feature-aggregation mechanisms employed not only for performance but also for enhancing model explainability—allowing clinicians to visualize which features or regions drive specific diagnostic outcomes [4].

---

Empirical Results and BenchmarksFusion-based skin cancer analysis methods have consistently demonstrated superior performance. State-of-the-art studies report classification and segmentation accuracies exceeding 90–94% on prominent dermoscopy datasets (ISIC, PH2, HAM10000), even with lightweight or ensemble architectures, as consolidated in the following table:

---

Key Challenges and OutlookWhile fusion methods yield clear gains, challenges persist—class imbalance, data heterogeneity, interpretability, and computational cost remain pressing issues [2][3][4]. Emerging directions include lightweight multi-stream fusion for mobile diagnosis [2], robust attention-guided fusion for explainable AI [4][6], and transformer-based approaches for unified long-range feature integration [3].

In summary, image fusion and feature aggregation—at both data and model levels—represent proven and versatile solutions for automatic skin cancer detection. These methods, particularly when coupled with attention, multi-scale, and ensemble frameworks, underpin robust, interpretable, and clinically relevant diagnostic pipelines [1][2][3][4].