Thermal and Infrared Inspection Images for Solar, Electrical and Building Diagnostics

What makes a thermal image training-grade

A thermal image is training-grade when the raw radiometric values survive export and the capture conditions are recorded. Many inspection reports embed palette-rendered JPEGs (iron, rainbow, white-hot) where the color scale was auto-ranged per frame, so the same pixel value means different temperatures in different images. A model trained on those learns palette artifacts, not heat.

Ask for the native file: radiometric JPEG (R-JPEG) with the thermal payload in the metadata, 14- or 16-bit TIFF, or the camera's proprietary sequence format, plus a tool chain that converts counts to temperature. Confirm the object parameters stored per frame: emissivity, reflected apparent temperature, object distance, relative humidity and atmospheric temperature. A changed emissivity setting, for example 0.95 versus 0.85, can shift a computed delta-T enough to move a finding across a severity threshold.

Camera model, detector resolution (commonly 640x512 or 320x256 for LWIR), lens field of view and the last factory calibration date belong in the manifest. Detectors drift, and a fleet of cameras with different calibration ages produces a dataset-level shift that looks like label noise. For a broader treatment of device variation, see camera, lens and lighting diversity in image datasets.

Solar PV thermography: conditions, faults and public baselines

Solar PV thermography is only comparable when captured under stable, sufficient irradiance, so ask every supplier for the irradiance reading and the inspection procedure behind each campaign. Many PV O&M teams follow IEC TS 62446-3, the IEC technical specification for outdoor infrared thermography of modules and plants. Ask which inspection level and procedure each campaign followed, because that determines what the labels can claim about module quality. Record whether scans were drone or handheld, the flight altitude and the viewing angle, since oblique angles raise reflected sky radiance on glass.

The fault taxonomy should separate failure mechanisms rather than just "hotspot": single-cell hotspots, cracked cells, substring patterns from an open or shorted bypass diode, potential-induced degradation, interconnect failures, junction-box heating, soiling and shading. Junction-box heat in particular is easy to confuse with a cell hotspot, so it deserves its own class. Public academic PV thermal sets exist and are useful for prototyping, but they tend to be narrow: low-resolution module crops, few sites and a fixed class list.

They do not cover tracker versus fixed-tilt geometry, bifacial modules, seasonal soiling, many climates or the long tail of string-level faults. Commercial vendors also advertise thermal plus RGB solar defect sets [1]; treat such listings as market practice and verify license terms and capture metadata before relying on them.

Electrical and building thermography: different labels, different failure modes

Electrical and building inspections need different labels than PV because the anomaly is defined relative to a reference, not to a fixed temperature. In switchgear, panelboards and motor control centers, findings are usually reported as a temperature rise over a similar component under similar load (phase-to-phase comparison) or over ambient, so the label is meaningless without the load at time of scan. Request percent load or measured current per circuit where the inspector recorded it.

Building diagnostics (missing insulation, air leakage, moisture intrusion, roof wet areas) depend on indoor-outdoor temperature difference and time since sunset or HVAC change. Without those fields, a model cannot tell a thermal bridge from a solar-loaded wall. Low-emissivity surfaces such as bare bus bars and aluminum flashing reflect surroundings, a classic source of false hotspots that labelers must flag as reflections rather than findings.

The most useful labels often already exist in the inspector's report: component, finding text, delta-T, severity class and recommended action. Linking those to frames is the same pattern described in domain captions from work records, and the reports themselves can be licensed as inspection reports.

Thermal inspection data request template

A precise request names the asset type, capture conditions, file format, label source and metadata you will reject without. The template below is a starting point for a solar O&M or electrical inspection team.

Illustrative example: invented to show structure; it does not describe an available dataset.

A matching record might look like this:

Illustrative example: invented to show structure; it does not describe an available dataset.

{"image_id": "pv_0412_ir", "format": "rjpeg", "camera": "LWIR 640x512, cal 2025-11",
 "emissivity": 0.90, "reflected_temp_c": 18.0, "distance_m": 25,
 "irradiance_w_m2": 780, "paired_rgb": "pv_0412_rgb",
 "annotations": [{"class": "bypass_diode_substring", "bbox": [212, 96, 58, 140],
   "delta_t_c": 11.4, "severity": "B", "source": "inspection_report"}]}

COCO-style JSON keeps images, categories and annotations sections that annotation tools such as CVAT import directly [4]; add the temperature fields as custom attributes rather than inventing a new format.

Label quality, class imbalance and normal-only sets

Thermal labels are noisier than they look, so budget for an audit before training. Even widely used benchmark test sets carry an estimated average label error rate of at least 3.3% [3], and thermal adds its own confusions: reflections labeled as hotspots, shading labeled as cell defects, and severity classes that differ between inspection firms. Re-derive severity from delta-T where possible so classes mean the same thing across suppliers.

Severe faults are rare, so a fleet scan may yield thousands of normal modules per bypass-diode failure. Two strategies help: pool defects across many sites (see rare defect coverage) and train on normal scans with methods such as PatchCore, which models normal patches and flags deviations [2]. The second approach needs clean normal-only image sets from the same cameras and seasons. Split by site and campaign, never by frame, or adjacent overlapping frames will leak across train and test.

How this differs from RGB inspection data

Thermal data differs from RGB inspection data because the signal is physical and condition-dependent, not just visual. Corrosion, weld and surface defect pages focus on texture and lighting; thermal adds emissivity, irradiance, load and calibration as first-class metadata. If your program spans both, see corrosion detection image datasets, drone inspection imagery for utility assets and the image data hub, and check transfer options in dataset delivery formats.

How SourceX helps source thermal inspection data

SourceX sources operational datasets from US companies on request, including documents, engineering records and new recordings of hands-on work, and manages licensing and ongoing purchases. Nothing is held in stock, and a request does not guarantee a match. You describe the data you need, such as radiometric PV scans with linked report findings, and SourceX looks for US businesses that hold it; every release is approved by the supplying company. Each dataset is rights-reviewed and delivered under a license that defines records, uses, term and delivery, with personal details such as names and account numbers removed or replaced and the method recorded. Buyers anywhere can start at the SourceX buyer page, and related categories include images and inspection photos and sensor and IoT data.

Request thermal inspection image data

Describe your asset types, file formats, capture conditions and label needs, and SourceX will look for US businesses that hold matching data and manage the license if a supplier agrees. Pricing and allowed uses are agreed per deal. Request thermal inspection image data.