How thermal signatures reveal PV faults before electrical degradation is visible

Infrared thermography detects photovoltaic faults by capturing the spatial distribution of radiant heat across module surfaces — cells operating under degraded conditions dissipate electrical energy as heat rather than converting it to power, producing characteristic thermal signatures detectable well before electrical performance visibly degrades. A power balance equation applied to IR temperature maps, as demonstrated by the University of Naples Federico II (2021), enables cell-level estimation of electrical power generated or dissipated — a quantitative diagnostic that goes beyond simple thermal contrast thresholding and was validated against 3D numerical simulations of malfunctioning panels.

The taxonomy of thermographic fault signatures was established empirically by the University of Valladolid (2019) in a comprehensive on-site study covering 17,142 monocrystalline modules. Five distinct defect modes were classified: hotspot in a cell, bypass circuit overheating, hotspot in the junction box, hotspot at the busbar connection, and general module degradation. This signature taxonomy is foundational for any automated classification system built on top of thermal imagery.

Complementary findings from Universidad de Valladolid (2020) showed that illuminated and dark-condition thermography provide complementary diagnostic information for the same defect types — a result with direct implications for utility-scale inspection scheduling, since different lighting conditions expose different fault signatures in the same module.

The most dangerous fault category — low-resistance hotspots caused by crystal defects — is analysed in depth by Yangzhou University (2022). Crystal defects create a shunt resistance path that, under partial shading, drives severe localised temperature rises capable of accelerating ageing and igniting fires. Detection of such low-resistance faults before they cause irreversible damage requires both thermal imaging sensitivity and correlation with I-V curve analysis — illustrating why multi-spectral, multi-parameter approaches outperform single-modality inspection.

A critical insight from the University of Oslo (2020), combining aerial IR thermography with inverter production data across two utility-scale plants, is that the most critical determinant of string-level production loss is the number of module substrings containing thermal anomalies — not simply the magnitude of individual hotspot temperatures. This finding directly motivates systematic multi-module, string-resolved thermal monitoring rather than spot checks of peak-temperature modules alone.

UAV multi-modal deployment: compressing 210 days of inspection into hours

UAV-mounted thermal systems reduce plant-wide hotspot inspection at a 30 MW facility from up to 210 days — the time required using conventional ground-level methods across 126,000 modules and 60 hectares — to hours or days, while enabling real-time georeferenced anomaly mapping. This scale problem, quantified by Universitas Indonesia (2020), is the central economic driver behind aerial multi-spectral inspection at utility scale.

“A 30 MW plant with 126,000 modules and 60 hectares requires up to 210 days for complete hotspot inspection using conventional methods. UAV-mounted thermal systems reduce this to hours or days while enabling real-time georeferenced anomaly mapping.”

The integration of optical RGB and thermal infrared sensors on a single UAV platform is the dominant multi-spectral deployment strategy. Chungbuk National University (2019) demonstrated that simultaneous optical and thermal infrared imaging allows production of orthographic thermal maps with accurate spatial coordinates, enabling precise localisation of faults within large string arrays. The complementarity of the two spectral modalities — RGB for structural context and soiling detection, IR for thermal anomaly identification — was explicitly leveraged by Sogang University (2019), which demonstrated improved fault localisation accuracy relative to single-sensor approaches.

A two-stage inspection strategy that combines the speed advantages of aerial IR with the resolution advantages of ground-level imaging was formalised by the University of Cádiz (2022). Stage one uses UAV-based thermal imaging for rapid plant-wide fault detection and localisation; stage two deploys ground-level high-resolution inspection only on flagged modules. This hierarchical multi-spectral approach substantially reduces total inspection time and cost at utility scale while maintaining diagnostic resolution for critical faults.

Hyperspectral imaging extends multi-spectral capability beyond the thermal band. China Southern Power Grid Energy Efficiency and Clean Energy Co. (2023) demonstrated that constrained energy minimization (CEM) and orthogonal subspace projection (OSP) applied to hyperspectral UAV imagery can detect surface stains on PV modules by exploiting known spectral signatures of clean versus contaminated panel surfaces — a capability invisible to single-band thermal cameras. This work, reported by IEEE-affiliated researchers, represents the leading edge of multi-spectral PV inspection beyond the thermal band.

Georeferencing of large-scale thermal inspection data was addressed by Forschungszentrum Jülich (2022), which used structure-from-motion to automatically assign geocoordinates to 99.3% of 35,084 modules across four large-scale plants, extracting over 2.2 million module-level IR images. This automated spatial indexing is a prerequisite for integrating aerial thermal data with plant management systems and for tracking fault evolution over successive inspection cycles.

AI-driven fault classification: from million-image datasets to 90%+ accuracy

Deep learning classifiers trained on million-image infrared datasets now achieve over 90% test accuracy across ten fault categories in utility-scale PV inspection — a capability that transforms aerial thermal data from raw imagery into actionable string-level diagnostics. The volume of imagery generated by utility-scale aerial IR inspections demands automated fault classification; manual review of millions of module images per inspection cycle is operationally infeasible.

Forschungszentrum Jülich (2021) curated a dataset of 4.3 million IR images from 107,842 PV modules across seven plants and trained a ResNet-50 classifier to identify ten common module anomalies with over 90% test accuracy. According to Nature Energy research on solar asset management, the scale of this dataset — unprecedented at time of publication — underscores the data infrastructure requirements of robust AI-driven thermal fault classification at utility scale.

Università Politecnica delle Marche (2020) proposed a mask R-CNN architecture for anomaly cell detection in UAV-acquired thermal images, enabling pixel-level segmentation of fault regions within modules without manual intervention. Ghent University / imec (2022) contributed a region-based CNN for PV anomaly detection in aerial imagery, addressing the foundational step of automated panel segmentation in complex backgrounds. These architectures collectively enable the pipeline from raw UAV thermal video to string-level fault maps without human review of individual frames.

Lightweight inference models for real-time deployment in resource-constrained UAV systems were addressed by Xinjiang University (2022), which replaced YOLOv5’s feature extraction backbone with the ShuffleNetv2-based Focus structure to reduce parameter count while maintaining detection accuracy on IR image datasets. This reflects a broader trend toward edge-deployable AI for onboard UAV fault detection, reducing dependency on ground-station processing latency.

For string-level electrical anomaly detection without imaging, the University of Malaya (2019) demonstrated a wireless sensor network (WSN) that monitors real-time module operating voltage across each string, detecting faulty modules exhibiting anomalously low voltage — providing a complementary electrical diagnostic channel to thermal imaging. UAE University (2020) fused PV panel junction temperatures with electrical parameters from inverters to achieve fault nature and location identification in a dual-layer diagnostic, addressing limitations of purely thermal or purely electrical approaches at utility scale. Standards for such sensor network deployments are increasingly informed by guidelines from IEC technical committees on photovoltaic systems.

“Multi-layered fusion of thermal signatures with real-time electrical parameters from inverters enables fault localization that neither method achieves independently.” — UAE University, 2020

Key innovators and the patent landscape shaping PV thermal monitoring

Huawei Technologies is the most active patent filer in this domain, with at least three active patents (EP 2021, EP 2023, US 2024) focused on image quality optimisation and fault type identification for PV modules using light-emitting state imaging and signal-to-noise ratio (SNR) control. Huawei’s SNR optimisation method — which maximises image definition prior to fault classification — is a critical preprocessing step that reduces false positives and processing workload in automated PV fault detection systems, as protected in the US 2024 patent.

Sungrow Power Supply Co. holds active EP patents (2016, 2018) protecting current characteristic curve-based string fault identification algorithms that automate severity classification without human intervention — a key commercial capability for inverter-integrated monitoring at utility scale. These algorithms enable automated string-state determination through tangent slope differencing of current characteristic curves, without requiring thermal imaging when electrical signatures are sufficient.

Academic institutions have driven the evidence base for the field. Forschungszentrum Jülich / HI ERN contributed both the large-scale IR video dataset with ResNet-50 classification (2021) and the structure-from-motion georeferencing method (2022), establishing the infrastructure for scalable automated aerial IR inspection. The University of Cádiz formalised the two-stage aerial/ground thermographic inspection strategy (2022). The University of Oslo provided the first rigorous quantitative link between thermal anomaly distribution and string-level production loss at utility scale (2020). Universidad de Valladolid produced the largest published on-site thermographic defect classification study — 17,142 modules — in 2019.

China Southern Power Grid Energy Efficiency and Clean Energy Co. is the sole assignee in this dataset deploying hyperspectral UAV imaging (2023) for stain detection using CEM and OSP spectral unmixing — representing the leading edge of multi-spectral (beyond thermal) PV inspection. Innovation trends across the dataset show a clear trajectory from single-sensor manual inspection toward multi-modal (RGB + IR + hyperspectral), AI-automated, UAV-deployed, string-resolved monitoring systems, with patent activity shifting from algorithmic fault classification toward hardware-software integration and SNR-optimised image acquisition. This trajectory aligns with broader renewable energy digitisation frameworks tracked by organisations such as IRENA in their annual innovation reports.

Honeywell International holds an EP patent (2022) covering solar panel inspection by unmanned aerial vehicle, indicating that large industrial conglomerates are entering a space previously dominated by specialist solar and academic players. The convergence of UAV hardware, multi-spectral sensors, and AI inference at the edge — as described in patent filings reviewed through EPO databases — points toward fully autonomous, continuous PV string health monitoring as the near-term commercial endpoint for this technology cluster.