Registration: European Humanoid Robot Summit 2025
Abstract:AI + Drones + Ground Robots! Solar panel maintenance solution: CNN-LSTM fault detection accuracy of 92.3%, reinforcement learning cleaning efficiency of 91.3%, edge latency of 47.2ms, cost reduction of 34.9%, Jaipur field test recovery of 31.2% in power generation, suitable for high dust and high temperature areas.
Introduction: The “Efficiency Dilemma” of Solar Panel Maintenance and Technological Breakthroughs
The global installed capacity of solar photovoltaic (PV) has exceeded 1185GW (2022 IEA data), but performance degradation due to dust accumulation, high surface temperatures, and maintenance delays remains a pain point in the industry. According to research by the U.S. National Renewable Energy Laboratory (NREL), uncleaned solar panels in desert areas can see power generation drop by 40%; a 250MW PV power station in Rajasthan, India, suffered a 32% reduction in annual power generation due to insufficient maintenance, resulting in losses exceeding $4 million. Traditional maintenance methods have three major shortcomings: manual cleaning is time-consuming and labor-intensive (large power stations require 2-3 people to collaborate, taking 4-6 hours to disassemble a single panel), semi-automated systems lack coordination (drones can only inspect without cleaning capabilities, and ground robots rely on preset programs), and cloud processing has high latency (averaging 180ms, unable to respond to faults in real-time).
To address these pain points, a team from Manipal Academy of Higher Education in India proposed an “AI-Integrated Autonomous Robot System” in the journal Scientific Reports (DOI:10.1038/s41598-025-17313-6) — through the collaboration of drones (thermal imaging + LiDAR inspection) and ground robots (reinforcement learning cleaning), combined with CNN-LSTM fault detection and edge AI real-time decision-making, achieving a closed loop of “detection – diagnosis – cleaning”. Field test data shows that the system achieves a cleaning efficiency of 91.3% (dust density reduced from 3.9mg/m³ to 0.28mg/m³), fault detection accuracy of 92.3%, edge inference latency of 47.2ms (63% faster than cloud processing), and a maximum recovery rate of 31.2% in power generation for heavily polluted panels, while reducing energy and water consumption by 34.9%, providing a feasible solution for solar maintenance in high dust and high temperature areas.
1. Technical Architecture: A Three-Dimensional Collaborative System of Drones + Ground Robots + Edge AI
1.1 Core System Modules and Hardware Configuration
The system adopts a three-layer architecture of “Aerial Inspection – Ground Execution – Edge Decision”, focusing on precision, real-time performance, and environmental adaptability in hardware selection. Key component parameters are as follows:
| Module | Core Hardware | Specifications | Function |
|
Drone System |
FLIR Vue Pro R Thermal Camera |
640×512 pixels, ±5℃ temperature accuracy, 2FPS acquisition |
Detect thermal spots, cell degradation, and other faults |
|
Livox Mid-40 LiDAR |
±2cm accuracy, 260m detection range |
Panel structure alignment, 3D modeling |
|
|
Ublox NEO-M8N GPS+IMU |
±2.5m positioning accuracy |
Autonomous inspection path planning |
|
|
Ground Robot |
Velodyne Puck VLP-16 LiDAR |
±3cm accuracy, 100m ranging |
Navigation and obstacle avoidance |
|
Adaptive Mechanical Brush |
Dynamic pressure adjustment (0-50N) |
Clean dust, avoid scratching panels |
|
|
Optical Dust Sensor |
0-5mg/m³ detection range, ±0.1mg/m³ accuracy |
Real-time monitoring of dust concentration, optimizing cleaning intensity |
|
|
Edge AI and IoT |
NVIDIA Jetson Nano |
472 GFLOPS computing power, 128-core GPU |
CNN-LSTM inference, reinforcement learning decision-making |
|
Raspberry Pi 4B |
Quad-core Cortex-A72, 4GB RAM |
Sensor data preprocessing, MQTT communication |
|
|
Communication Protocol |
MQTT over TLS 1.3 |
latency<50ms, QoS 2 (Exactly Once) |
Secure data transmission between drone – robot – edge |
1.2 Core Algorithms: AI-Driven Fault Detection to Cleaning Optimization
1) CNN-LSTM Hybrid Fault Detection Model
For solar panel faults (thermal spots, cracks, dust accumulation), the model adopts a “CNN extracts spatial features + LSTM captures temporal trends” architecture:
-
CNN Layer: 4 convolutional layers (64→128→256→512 filters) + 2×2 max pooling, input 224×224×3 thermal/visible light images, extracting spatial features such as thermal spots and cracks on the panel surface;
-
LSTM Layer: 2 layers of bidirectional LSTM (128 hidden units each), analyzing temporal correlations of continuous frame data (e.g., temperature trends of thermal spots over time), avoiding misjudgment from single detections;
-
Training and Performance: Trained on 2000 labeled data sets (60% field data + 40% simulation data), Adam optimizer (learning rate 1e-4), binary cross-entropy loss function, achieving a final fault detection accuracy of 92.3%, F1-score of 0.915, significantly outperforming traditional CNN (84.4% accuracy) and SVM (78% accuracy).
2) DQN Reinforcement Learning Cleaning Strategy
The cleaning behavior of the ground robot is optimized through a Deep Q-Network (DQN) with the goal of “maximizing cleaning efficiency + minimizing resource consumption”:
-
State Space (S): Dust concentration (Dₜ), solar irradiance (Iₜ), robot battery level (Eₜ), obtained in real-time through sensors;
-
Action Space (A): Low-intensity cleaning, medium-intensity cleaning, high-intensity cleaning, move to the next panel, standby, dynamically matching dust levels;
-
Reward Function (R): Cleaning efficiency > 90% and low energy consumption earns +10 points, 70%-90% earns +5 points, <50% deducts 10 points, excessive energy consumption deducts 20 points;
-
Training Results: After 10,000 training rounds, convergence was achieved in 3,500 rounds, with cleaning efficiency stabilizing at 91.3%, reducing water and energy consumption by 34.9% compared to rule-based cleaning (fixed intensity), and reducing path redundancy by 30.7% (from 18.2m to 12.6m).
1.3 System Workflow: Closed Loop Execution from Inspection to Cleaning
1) Drone Inspection Initiation:Fly along AI-optimized paths, simultaneously collecting thermal images (thermal spot detection), LiDAR (structural analysis), and visible light images (dust assessment), with data transmitted in real-time to the edge via MQTT;
2) Edge AI Fault Diagnosis:The CNN-LSTM model analyzes the data, and if the fault probability > 85% (e.g., thermal spot temperature exceeds the environment by 5℃, dust > 3.5mg/m³), generates a fault command with GPS coordinates;
3) Ground Robot Response:Upon receiving the command, the DQN model plans the optimal cleaning path, adjusts mechanical brush pressure and movement speed, and transmits dust concentration and battery data in real-time during the cleaning process;
4) Effect Verification and Feedback:After cleaning, the drone performs a secondary inspection to verify the results, the edge updates the maintenance log, and if standards are not met, a secondary cleaning is triggered.
2. Comparative Advantages: Four Core Breakthroughs Disrupting Traditional Solar Maintenance
2.1 Performance Comparison with Traditional Maintenance Methods
This system comprehensively surpasses manual cleaning, semi-automated systems, and cloud processing solutions in efficiency, cost, and reliability, with specific comparisons as follows:
| Comparison Dimension | Traditional Manual Cleaning | Semi-Automated System (Drone Inspection + Manual Cleaning) | Cloud Processing Driven System | This AI Integrated System | Improvement Rate |
|
Cleaning Efficiency |
68.7% (high manual error) |
75.2% (fixed cleaning intensity) |
82.5% (delays lead to missed cleaning) |
91.3% (adaptive intensity) |
22.6% improvement over manual |
|
Fault Detection Accuracy |
65% (visual inspection, subjective) |
83.2% (only image recognition, no temporal analysis) |
88.1% (cloud delays lead to misjudgment) |
92.3% (CNN-LSTM spatiotemporal fusion) |
9.1% improvement over semi-automated |
|
Time Taken for Single Panel Maintenance |
45 minutes (2-3 people collaborating) |
30 minutes (1 person operating drone + cleaning) |
25 minutes (cloud decision delay 50ms) |
12 minutes (fully automated closed loop) |
73.3% faster than manual |
|
Resource Consumption (Water + Electricity) |
100% (baseline) |
85% (high drone power consumption) |
78% (cloud transmission energy consumption) |
65.1% (DQN optimization) |
34.9% reduction compared to manual |
|
Extreme Environment Adaptability |
Poor (high temperature/dust prone to overheating) |
Average (drone wind resistance < level 5) |
Poor (network fluctuations affect decisions) |
Good (wind resistance level 8, operating -4℃ to 44℃) |
Adapted temperature range expanded by 88℃ |
2.2 Core Competitive Advantages Analysis
1) “Inspection – Cleaning” Collaboration, Breaking Technical Silos
In traditional solutions, drones can only “detect problems” and require manual intervention for the cleaning phase (delays > 24 hours); this system achieves real-time communication via MQTT, dispatching ground robots within 10 seconds after drone detection of faults, realizing “problem discovery and immediate handling”, compressing maintenance response time from days to minutes. Field tests on a 3.96kW PV array in Jaipur, India, showed that this system reduced power generation losses due to faults from 28% to < 5%.
2) Edge AI Enables Low-Latency Decision Making
Using Jetson Nano and Raspberry Pi 4B to build edge computing nodes avoids the delays of uploading data to the cloud (average cloud processing 180.3ms), with edge inference latency of only 47.2ms, meeting real-time control requirements (<50ms). Additionally, the edge preprocesses data (e.g., thermal image denoising, LiDAR point cloud filtering), reducing transmission bandwidth by 78.6% (from 1.2MB/s to 280KB/s), suitable for remote areas with weak network environments.
3) Reinforcement Learning Adapts to Dynamic Environments
Unlike rule-based cleaning (fixed pressure/path), the DQN model can adjust strategies based on real-time dust concentration and irradiance: for example, during high temperatures (44℃) at noon in Rajasthan, it automatically reduces cleaning intensity (to avoid overheating panels) while prioritizing cleaning panels with dust > 4mg/m³ (these panels can lose up to 31.2% in power generation, yielding the highest cleaning benefits).
4) Multi-Dimensional Data Fusion Enhances Reliability
By integrating thermal (thermal faults), LiDAR (structural faults), and optical (dust) data, Pearson correlation analysis (r=0.87, p<0.01) confirms the strong correlation between “dust concentration and thermal spots” — when dust > 3.5mg/m³, the average panel temperature increases by 6℃, establishing a “dust – thermal spot” linkage warning to avoid misjudgment from a single sensor (e.g., during cloudy days when visible light sensors fail, thermal imaging can still locate dust areas).
3. Existing Disadvantages: Three Real-World Challenges from Laboratory to Industrialization
3.1 High Hardware Costs, High Entry Barriers for SMEs
The total cost of core system components (FLIR thermal camera, Livox LiDAR, Jetson Nano) is approximately $18,000, which is 125% higher than traditional semi-automated systems (around $8,000). Among these, the thermal camera (35% share) and LiDAR (28% share) are the main cost items, while the average annual maintenance budget for small and medium-sized PV power stations (<10MW) is about $50,000. Deploying this system would require an investment recovery period of 3-4 years (compared to about 2 years for traditional solutions), limiting its widespread adoption.
3.2 Insufficient Scalability for Large-Scale Deployment
The current system performs stably in a 3.96kW small array (12 panels) but faces two major issues when scaling to MW-level power stations (thousands of panels):
-
Multi-Robot Collaboration: The existing architecture only supports 1 drone + 1 ground robot; multi-robot collaboration needs to address path conflicts (e.g., multiple robots cleaning the same area) and data synchronization (aligning timestamps of sensors from different nodes), currently lacking distributed scheduling algorithms;
-
Model Generalization: The CNN-LSTM model is trained only on mainstream PV panels from brands like Ford and Tesla, with recognition accuracy dropping to 75% for niche brands (e.g., local Indian Adani Solar), requiring an additional 5,000+ training images for adaptation, and new models need 2-3 weeks of debugging (traditional solutions require 3-6 months, although advantageous, still need optimization).
3.3 Performance Degradation in Extreme Environments
The system is stable in controlled laboratory environments (20-25℃, humidity 40%-60%), but performance declines in extreme environments common in PV power stations:
-
Low Temperatures and High Humidity: At temperatures <0℃, LiDAR laser reflectivity decreases by 15%, and positioning accuracy drops from ±2cm to ±5cm; at humidity > 85%, optical sensor lenses fog up, increasing dust detection errors by 22%;
-
Sudden Environmental Changes like Dust Storms: When PM2.5 concentration > 500μg/m³, the drone’s thermal camera penetration decreases, and thermal spot detection accuracy drops from 92.3% to 81.5%, requiring downtime until weather improves, affecting continuous maintenance.
4. Application Case: 72-Hour Field Validation in Jaipur, India
To validate the system’s practicality, the team conducted a 72-hour continuous test at the Sitapura PV laboratory in Jaipur, India (26.7987°N, 75.8765°E), deploying 12 pieces of 330W monocrystalline silicon panels (total power 3.96kW), simulating high dust and high temperature environments (day-night temperature difference of 48℃: dawn – 4℃ to noon 44℃), with key results as follows:
1) Cleaning Effect:The average dust concentration on the panels was reduced from 3.9mg/m³ to 0.28mg/m³, achieving a cleaning efficiency of 91.3%. Among them, two heavily polluted panels (initial dust 4.5-4.7mg/m³) had their power generation restored from 210-214W to 276-281W, with a recovery rate of 31.2%;
2) Fault Detection:Successfully identified thermal spots (temperature 45-65℃) on 7 panels, with no missed or misjudged detections, response time of 47.2ms, 152 times faster than manual inspections (average 2 hours to detect faults);
3) Resource Consumption:Water consumption per panel was 185ml (traditional manual cleaning 310ml, a reduction of 40.3%), and electricity consumption was 3.45Wh (traditional 5.21Wh, a reduction of 33.8%), in line with Rajasthan’s “water and energy conservation” policy;
4) Reliability:During the test, there were 2 instances of WiFi disconnection (RSSI<-80dBm), and the system automatically switched to BLE 5.2 communication, with a recovery time of 1.2±0.3s, without maintenance interruption, and MTTR (Mean Time to Recovery) of 2.3s (traditional systems 8.7s, 73.5% faster).
5. Future Outlook: Upgrading from “Single Station Maintenance” to “Smart PV Farms”
The team proposed three major optimization directions in the literature to address existing disadvantages and expand application scope:
1) Multi-Agent Collaboration (MARL):Develop distributed reinforcement learning algorithms to support collaboration of 10+ drones and ground robots, adapting to MW-level power stations, aiming to control multi-machine scheduling delays within 100ms;
2) Federated Learning Cost Reduction:Adopt an “edge training + cloud aggregation” model, allowing different PV power stations to share model parameters (without transmitting raw data), reducing adaptation time for new models from 2-3 weeks to 3 days while protecting data privacy;
3) Hardware Lightweighting:Through model quantization (TensorRT optimization) and low-power chips (e.g., NVIDIA Jetson Orin NX), reduce edge costs by 40%, aiming to achieve a total system cost < $10,000 by 2027, making it affordable for SMEs.
For more details, please refer to the original article:“AI-Integrated Autonomous Robotics for Solar Panel Cleaning and Predictive Maintenance Using Drone and Ground-Based Systems”
END
Click “Read the original text” for more