As of May 2025, China’s solar photovoltaic installed capacity has surpassed 1,000 GW, marking a tremendous milestone in the global energy transition. The rapid expansion of large-scale photovoltaic plants—from distributed systems of a few tens of megawatts to utility-scale bases exceeding 3 GW—has introduced critical challenges in operation and maintenance. Traditional manual inspection of solar panels is increasingly inadequate due to the vast geographic footprint, harsh environmental conditions, and rugged terrain. Manual inspection methods suffer from low efficiency, high labor costs, and safety risks, particularly in high-altitude or mountainous regions. To address these issues, we have systematically investigated and implemented artificial intelligence (AI)-powered inspection technology for solar panels, integrating advanced sensors, unmanned aerial vehicles (UAVs), data preprocessing, and fault detection algorithms. This paper presents our research and engineering practice at a 500 MW photovoltaic plant in Qinghai Province, demonstrating the feasibility and economic benefits of intelligent inspection systems.
1. Common Faults and Detection Methods for Solar Panels in Operation
Solar panels operating in the field are continuously exposed to environmental stressors such as UV radiation, wind, hail, sandstorms, and biological contamination. They are also subject to ground subsidence, landslides, and seismic events. Consequently, various defects can emerge, including short circuits in cells or strings, string disconnection, bypass diode activation, missing or cracked glass, cracks in silicon cells, dust and bird droppings, and hot spots. Among these, hot spots are the most frequent manifestation, caused by partial shading, cell cracks, broken fingers, connection issues, or aging. When a shaded or defective cell is forced to conduct current from normal cells, it dissipates excessive power as heat, leading to localized overheating—a phenomenon known as the hot-spot effect.
We classify solar panel defects into four severity levels based on their impact on system performance and repair priority:
| Defect Level | Type | Typical Examples |
|---|---|---|
| I (Critical) | Short circuit in cells/strings | Cell short, string short |
| II (Severe) | String disconnection, bypass diode activation | Open string, diode bypass |
| III (Moderate) | Missing or broken panel | Cracked glass, shattered panel |
| IV (Minor) | Hot spots, cracks, dust, bird droppings | Local hot spot, micro-crack, shading |
Traditional detection methods include visual inspection, electroluminescence (EL) testing, and infrared thermography. Visual inspection relies on human eyes to identify cracks, discoloration, or debris. EL testing uses a portable EL camera to “see inside” the panel, revealing micro-cracks, broken fingers, and poor soldering. Infrared thermography employs an IR camera to detect temperature anomalies caused by resistive heating. Hot spots appear as bright regions in thermal images, with temperature differences of 5–10 °C compared to normal areas. Common thermal signatures include point-like hot spots (due to cell cracks or shading like bird droppings), streak-like hot spots (due to broken fingers or poor soldering), and junction box overheating (due to loose terminals or diode failure).
2. Fundamentals of Artificial Intelligence for Solar Panel Inspection
Artificial intelligence (AI) enables machines to simulate human intelligence. Key technologies relevant to solar panel inspection include data acquisition and processing, machine learning, deep learning, and computer vision. Data acquisition involves collecting images and sensor data from various sources, followed by preprocessing, cleaning, and organizing. Machine learning allows computers to improve algorithms by analyzing data. Deep learning uses multi-layer neural networks to automatically learn complex patterns from massive datasets, excelling in non-structured data such as images and speech. Convolutional neural networks (CNNs) are a core deep learning model designed for grid-like data (e.g., image pixels). CNNs efficiently extract local features through convolution operations and compress information via pooling layers, achieving high accuracy in image recognition and classification.
A typical CNN architecture consists of the following layers:
- Convolutional layer: applies a small kernel (filter) sliding over the input image to compute weighted sums, capturing basic features like edges, textures, and colors. For a 2D input image X and kernel K, the convolution operation is defined as: $$(X * K)(i,j) = \sum_{m}\sum_{n} X(i+m,j+n) \cdot K(m,n)$$
- Activation function: introduces non-linearity, commonly using ReLU: $$\text{ReLU}(x) = \max(0, x)$$
- Pooling layer: downsamples the feature map, reducing spatial dimensions while retaining important information. Max pooling selects the maximum value in each window: $$P(i,j) = \max_{p,q \in window} F(i+p,j+q)$$
- Fully connected layer: integrates all extracted features to make final classification decisions. The output is computed as: $$\mathbf{y} = \sigma(\mathbf{W} \cdot \mathbf{x} + \mathbf{b})$$ where \(\mathbf{W}\) is weight matrix, \(\mathbf{b}\) bias, and \(\sigma\) activation function.
For solar panel fault detection, CNN can automatically learn hierarchical features from IR or visible images, distinguishing normal cells from defective ones such as hot spots, cracks, or shading.
3. Unmanned Aerial Vehicle (UAV) Data Acquisition System
We deploy UAVs equipped with high-resolution visible cameras and infrared thermal imagers to capture images of solar panels from an altitude of 3–5 m. The UAV system must meet the following requirements: IP54 protection rating, dual-sensor payload capability, operating temperature range from -30 °C to 50 °C, flight endurance of approximately 55 minutes, and maximum wind resistance of 12 m/s. Our selected platform is the DJI M350 RTK, which provides centimeter-level positioning accuracy when combined with a Trimble PPK module. The gimbal camera is the Zenmuse H20T, integrating a wide-angle camera, zoom camera, thermal imager, and laser rangefinder, enabling simultaneous capture of visible and thermal images.
The ground station can be either a fixed or mobile hangar. A mobile hangar (e.g., Zhixun M100) offers flexibility to move across the photovoltaic plant, significantly reducing investment costs and suiting large, dispersed, or mountainous solar panel fields. The hangar provides automatic charging, storage, release, and data transmission capabilities. The operator can define virtual station positions and initiate fully automated flight missions from a remote terminal.
4. Data Preprocessing for Solar Panel Images
Raw images acquired by the UAV require thorough preprocessing before being fed into AI models. This step is critical for improving the accuracy and efficiency of fault detection. Our preprocessing pipeline includes the following stages:
| Step | Description | Example / Formula |
|---|---|---|
| Data Cleaning | Remove low-quality, irrelevant, or duplicate images to ensure dataset validity. | Eliminate blurred or overexposed frames. |
| Denoising | Reduce random noise from sensors or lighting using filters. | Gaussian filter: \(G(x,y) = \frac{1}{2\pi\sigma^2}e^{-\frac{x^2+y^2}{2\sigma^2}}\) |
| Data Augmentation | Generate synthetic samples by applying transformations to solve class imbalance and increase dataset size. | Random rotation, flipping, scaling, color jitter: \(\mathbf{X}’ = T(\mathbf{X})\) |
| Normalization / Standardization | Scale pixel values to a standard range to accelerate convergence. | \(x’ = \frac{x – \mu}{\sigma}\) |
| Label Encoding | Convert textual labels (e.g., “hot spot”) into numerical format for model training. | One-hot encoding: [1,0,0] for normal, [0,1,0] for hot spot, [0,0,1] for crack. |
Data augmentation is especially important for solar panel inspection because defect samples (e.g., hot spots, cracks) are rare compared to normal panels. By applying random rotations (\(±30°\)), horizontal flips, and brightness adjustments, we can effectively increase the diversity of training data and improve model robustness.
5. Fault Detection Algorithms for Solar Panels
After preprocessing, we extract features from infrared and visible images that distinguish normal solar panels from defective ones. Key features include temperature anomalies, brightness anomalies, shadow patterns, edge information, morphological characteristics, and multi-scale features. These features are then used to classify defects using machine learning or deep learning models.
We primarily employ two approaches:
5.1 Support Vector Machine (SVM) with Handcrafted Features
SVM is a classical supervised learning model that finds an optimal hyperplane to separate classes. For a set of training samples \(\mathbf{x}_i\) with labels \(y_i \in \{-1, +1\}\), the decision function is: $$f(\mathbf{x}) = \text{sgn}\left(\sum_{i=1}^{n}\alpha_i y_i K(\mathbf{x}_i, \mathbf{x}) + b\right)$$ where \(\alpha_i\) are support vector coefficients, \(K\) is the kernel function (e.g., radial basis function), and \(b\) is bias. In our application, we extract statistical features from thermal images (e.g., maximum temperature, temperature variance, hot area ratio) and train an SVM to classify normal vs. hot-spot panels.
5.2 Convolutional Neural Network (CNN) for End-to-End Learning
We design a custom CNN architecture to directly learn discriminative features from raw images. The forward propagation through convolutional layers is expressed as: $$\mathbf{L}^{(k)} = \sigma\left(\mathbf{W}^{(k)} * \mathbf{L}^{(k-1)} + \mathbf{b}^{(k)}\right)$$ where \(\mathbf{L}^{(k)}\) is the output of layer \(k\), \(\mathbf{W}^{(k)}\) the weight kernel, \(\mathbf{b}^{(k)}\) the bias, and \(\sigma\) the activation function (ReLU). After several convolutional and pooling layers, we flatten the feature maps and apply fully connected layers followed by a softmax output for classification: $$P(y=c|\mathbf{x}) = \frac{e^{\mathbf{w}_c^T \mathbf{x}}}{\sum_{j=1}^{C} e^{\mathbf{w}_j^T \mathbf{x}}}$$
We train the network on a large dataset of solar panel images annotated with defect types. The loss function is categorical cross-entropy: $$L = -\frac{1}{N}\sum_{i=1}^{N}\sum_{c=1}^{C} y_{i,c} \log(\hat{y}_{i,c})$$ where \(N\) is batch size, \(C\) number of classes, \(y_{i,c}\) ground truth, and \(\hat{y}_{i,c}\) predicted probability.
Our final system integrates both SVM and CNN, with CNN achieving over 92% recognition accuracy for the four defect levels, significantly outperforming manual inspection (85%).
6. Engineering Practice: 500 MW Photovoltaic Plant in Qinghai
6.1 Project Overview
The demonstration site is a 500 MW photovoltaic plant located in Qinghai Province at an altitude of approximately 3,000 m. The site covers a rectangular area of 4,200 m by 4,000 m, occupying about 800 hectares. The plant comprises 150 sub-arrays, totaling 910,000 bifacial monocrystalline solar panels, 2,100 combiner boxes, 130 string inverters, 10 step-up transformers, and 140 combined inverter/transformer units. The extreme temperature range is -34 °C to 40 °C, with maximum wind speeds of 30 m/s. Due to the harsh environment and large scale, manual inspection is nearly impossible. Therefore, we decided to implement an intelligent UAV-based inspection system during the construction phase.
6.2 System Architecture
The intelligent inspection system consists of a mobile UAV hangar (Zhixun M100), a DJI M350 RTK drone with Zenmuse H20T dual-spectrum payload, and on-board edge computing servers. The software platform provides web and mobile interfaces for multi-station management, device management, route planning, remote control, defect diagnosis, statistics, and report generation.
6.3 Workflow
The inspection workflow follows these steps:
- Site modeling: A pre-flight UAV maps the entire site to create a high-resolution orthophoto and 2D model, precisely locating every solar panel.
- Route planning: The system automatically calculates optimal flight paths with 80% forward overlap and 70% side overlap to ensure full coverage.
- Pre-flight preparation: Choose a clear, low-wind day (<4 Beaufort). Calibrate the thermal camera with a blackbody and warm up for 15 minutes.
- Automated flight: The operator dispatches the mission. The UAV takes off from the mobile hangar, follows the planned route, and captures both visible and thermal images of all solar panels.
- Data transmission: After landing, the UAV automatically transfers collected images to the edge server inside the hangar via wired or wireless connection.
- Data preprocessing and analysis: The server applies cleaning, denoising, normalization, and runs the trained CNN and SVM models to detect and classify defects. The system generates a defect report that includes panel IDs, GPS coordinates, defect types, severity levels, and suggested maintenance actions.
- Review and rectification: Maintenance personnel access the report via the mobile terminal, locate the defective solar panel using coordinates and visual markers, and perform on-site confirmation and repair.
- Re-inspection: After repairs, the system can schedule a follow-up flight to verify defect resolution.
6.4 Results and Performance
During the first year of operation, the system completed over 200 autonomous inspection missions, covering 100% of the 910,000 solar panels. A total of 3,842 defects were identified, including:
| Defect Type | Count | Percentage |
|---|---|---|
| Hot spots | 2,105 | 54.8% |
| Micro-cracks | 724 | 18.9% |
| String disconnection | 312 | 8.1% |
| Bypass diode activation | 256 | 6.7% |
| Dust / bird droppings | 445 | 11.6% |
The average inspection time for the entire plant was 6 hours (including flight and data processing), compared to an estimated 30 man-days for manual inspection. The defect detection accuracy reached 94%, and the false positive rate was below 3%. The system enabled prompt maintenance, preventing an estimated 2% annual energy yield loss due to undetected faults.
7. Technical and Economic Advantages
Based on our engineering practice, we summarize the key advantages of AI-driven UAV inspection for solar panels:
| Aspect | Traditional Manual Inspection | AI-Powered UAV Inspection |
|---|---|---|
| Inspection efficiency | 0.5–1 hectare per person-day | 50–100 hectares per UAV per day |
| Defect detection accuracy | 85% (subjective, fatigue) | >92% (consistent, AI-assisted) |
| Minimum detectable flaw size | ~1 mm (visual) | 0.2 mm (high-res camera) |
| Temperature sensitivity | N/A (visual only) | 0.1 °C (thermal IR) |
| Safety risks | Working at height, electric shock, extreme weather | Zero personnel exposure |
| Data traceability | Paper records, prone to loss | Digital logs with GPS coordinates, fully auditable |
| Cost per inspection (500 MW plant) | >$15,000 (labor, equipment, travel) | $2,000–$3,000 (UAV amortized, minimal labor) |
Beyond cost savings, the system provides a foundation for predictive maintenance. By analyzing historical defect patterns, we can forecast failure trends and optimize replacement schedules, further reducing long-term operational expenses.
8. Conclusion
We have systematically developed and deployed an AI-based inspection technology for solar panels, integrating UAV platforms, dual-spectral sensors, advanced data preprocessing, and deep learning fault detection algorithms. Our work addresses the critical operational challenges of large-scale photovoltaic plants, achieving substantial improvements in efficiency, accuracy, and safety. The successful implementation at a 500 MW plant in Qinghai demonstrates that AI-powered UAV inspection can reduce inspection time by 90%, increase defect detection accuracy to over 92%, and eliminate personnel safety risks. The technology also provides valuable data for predictive maintenance and lifecycle management of solar panels. As UAV and AI technologies continue to evolve—with longer endurance, real-time onboard processing, and multi-modal sensor fusion—we envision fully autonomous, all-weather, and whole-field inspection becoming the standard for future solar panel operations. Our research offers a practical reference for similar large-scale renewable energy projects aiming to harness the power of artificial intelligence for reliable and cost-effective asset management.