Solar and Battery Cell Defect Data: Sourcing Electroluminescence and Cell Inspection Images

Why energy components have their own data problem

Many defects in solar cells and battery cells are invisible in a photograph. A micro-crack in a silicon cell, a broken finger in the metallization or an inactive region does not show under room light; it shows when the cell is electrically excited and imaged. Inside a battery cell, electrode alignment and contamination are typically examined with X-ray or CT. Production lines in these sectors are often instrumented, and inline inspection can generate large image streams, which makes the archives interesting. It also makes them sensitive.

This page covers sourcing and evaluating that imagery. The general X-ray topic is in X-ray and CT inspection images.

What a public solar reference shows

Deitsch and colleagues studied defect classification of photovoltaic cells in electroluminescence (EL) images, comparing a support vector machine at 82.44% and a convolutional network at 88.42% average accuracy, trained on 1,968 cells (Deitsch et al., 2018). The associated ELPV dataset README lists 2,624 grayscale cell images of 300 by 300 pixels from 44 modules, with defect probabilities rather than binary labels.

Two points follow for a buyer. First, graded labels are a reasonable choice, and the ELPV data uses them: defect probability or severity handles the fuzzy boundary between a harmless and a harmful crack. Second, the numbers are one study's results on one dataset, so they say little about accuracy on your cells. The work covers photovoltaic cells only. This page cites no battery-cell dataset, so battery-specific statements below are general practice to confirm with each supplier.

Imaging modes and what they reveal

General orientation, not drawn from a cited source; confirm what each mode shows with the supplier.

Because appearance depends on drive conditions, two EL images of the same cell at different currents are not interchangeable. Imaging conditions metadata explains the general discipline, and the table above gives the extra fields this sector needs.

Labels and taxonomies

Solar defect labels often distinguish crack types, finger interruptions, material and process faults, and shading or inactive regions, with a severity or power-loss relevance that depends on module design. Battery-cell labels may cover electrode misalignment, foreign matter, weld defects, cell deformation, coating faults and leakage evidence. Both sets need a stated taxonomy and version; see defect taxonomy and label schema.

Ask whether labels are cell-level, module-level or region-level. A module-level pass/fail aggregated from many cell images is much weaker supervision than per-cell grades. Ask also whether labels come from human experts or from the vision system's own output; if the latter, the data teaches a model to imitate that system.

Where the data lives

Cell and module manufacturers hold inline EL, X-ray and optical inspection archives, plus end-of-line test results. Test laboratories and certification bodies hold qualification test images. Installers, operators and asset owners may hold field EL or thermal drone images of operating plants. Battery makers and pack assemblers hold X-ray and CT archives, often under strict internal control.

Field and factory images are different strata and should not be merged without labeling. Factory images carry production data (batch, line, equipment); field images carry site, age and operating history.

Illustrative record for a cell-inspection package

# Illustrative only - synthetic values; not a description of any real archive
package: pv_cell_el_v0
imaging: electroluminescence
record:
  image: cell_000218.png
  module_id: mod_0441        # pseudonymous
  cell_position: {row: 3, col: 7}
  line_id: line_B
  batch_id: batch_22
  drive: {current_note: "recorded", voltage_note: "recorded"}
  exposure_note: recorded
  label_scheme: "taxonomy v2, graded"
  label:
    defect_types: [crack, finger_interruption]
    severity_grade: 0.67       # example graded value, not a standard
    labeler_role: expert_2
  test_result: {power_loss_note: "linked if available"}
split_key: module_id then batch_id
excluded: robot actions, joint states, forces

Values are examples. If you need battery-cell X-ray records, the same structure applies, with geometry and source settings replacing drive conditions.

Quality and acceptance pointers

• Split by module (or pack), then batch. Neighboring cells on one module share lighting, optics and process history, so cell-level shuffling inflates results.
• Drive-condition consistency. Check whether EL images were taken at consistent excitation, and flag mixed sets.
• Label granularity. Confirm that graded labels have a written scale and a rule for combined defects.
• Defect rarity. Many cells are good; see rare defects and class imbalance.
• Linkage to performance. Final test results make labels more meaningful; see inspection images linked to outcomes.

The thresholds are for the buyer to set; none is a universal minimum.

What this data cannot support

Cell-inspection images show component condition; they contain no robot actions, joint states or forces. They cannot train a humanoid to handle, stack or install modules or cells, which needs robot or teleoperation data. Handling cells also involves safety issues that inspection images say nothing about. They also cannot predict lifetime or safety on their own: a cell that passes imaging may still degrade, and labels do not equal reliability data unless outcomes are joined.

Process generations change quickly, so archives age; a model trained on one cell design may not transfer to a new architecture. Commercial lines treat defect data as competitive information, so coverage you can license may be narrower than coverage that exists.

Questions to ask suppliers

1. Which imaging modes are archived, and with what excitation or source settings?
2. Are images stored at original resolution and bit depth?
3. Are labels binary, categorical or graded, and who assigned them?
4. Is the label scheme versioned, and has it changed during the archive?
5. Can images be tied to module, batch, line and equipment?
6. Are end-of-line electrical or power test results linkable?
7. Are field images mixed with factory images, and are they tagged?
8. What cell or module designs are shown, and are they commercially sensitive?
9. Who can authorize licensing, and do customer or equipment-vendor terms restrict it?

Permissions and provenance

Cell and module images can reveal metallization patterns, cell architecture, process windows and equipment settings. Customer agreements and equipment-vendor terms may restrict sharing. Ask what the supplier is allowed to license, whether production-process data embedded in file headers must be removed, and whether any program has export-control or public-funding restrictions. Where cells or modules are made for a customer, ask whether that part owner must consent before images are licensed. X-ray and CT capture also uses ionizing radiation, so any commissioned capture would need to run under the provider's own radiation-safety program; ask who performed existing captures and who would perform new ones. This is not legal advice or safety guidance. Cropping or removing serial numbers does not by itself settle a confidentiality or ownership question; see confidentiality in defect data.

Discuss your requirements

If you need solar or battery cell inspection imagery, describe the component, imaging mode, label type and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources among US businesses; a request does not guarantee a match. Continue with the inspection data hub, the humanoid training data guide, or the catalog entry for images and inspection photos.

Explore sourcing and licensing options with SourceX

Sources

• Automatic Classification of Defective Photovoltaic Module Cells in Electroluminescence Images — EL-image defect classification results and the ELPV dataset description; photovoltaic cells only.