1,000 training images to reach production-ready inspection performance — versus 10,000 for conventional systems — is the number that makes electroluminescence inspection tractable at module-line scale. An EL camera captures the defects. A trained AI model decides what the dark spots mean. The camera is a commodity purchase. The inference is the product.
EL imaging in a PV line environment gets treated as a camera upgrade decision. The module manufacturer buys a higher-sensitivity InGaAs camera, installs it in the tabber enclosure, and expects inspection. What they have at that point is raw near-infrared emission images with dark and gray regions whose significance depends on context: the same dark area can be a dead cell, a crack shadow, potential-induced degradation onset, or a normal variation in carrier lifetime distribution. Distinguishing between these at 3,600 cells per hour is not a camera problem. It is a classification, detection, and segmentation problem running simultaneously on each frame, with inference completing before the next cell arrives.
This post covers what EL imaging captures, what the common defect signatures mean in power-loss terms, and what the inference system needs to deliver at tabbing speed.
What electroluminescence imaging captures
When a silicon PV cell is forward-biased — current injected in the direction that normally produces electricity — it emits near-infrared light through radiative recombination. The emission intensity at any point on the cell surface is proportional to the minority carrier density at that location. Areas where carriers recombine non-radiatively through defect states, grain boundaries, or broken current paths emit less light and appear darker in the EL image.
This makes EL a direct map of the cell's electrical behaviour, not just its physical surface. A camera that captures visible light shows surface contamination, scratching, and cell colour variation. An EL camera shows which parts of the cell are contributing to power output and which are not. A cell that looks clean under white light can show a complex dark pattern under EL that represents a 5–15% power loss from a crack network that formed during cell handling or laminate pressing.
The cost of detecting EL defects at the cell or string stage, before lamination, is negligible — a cracked cell is removed and replaced. The cost of the same defect in a shipped module is a warranty claim, a service call, or a field replacement at the installation site. The value-chain logic is the same as it is for any other inspection application: discover the defect at the earliest point in the value stream where it is detectable.
EL defect glossary
| EL image feature | Physical defect | Power loss mechanism | Appropriate inference task |
|---|---|---|---|
| Dark line crossing current-collecting fingers | Finger crack (type A) | Isolates PV material on the shadow side of the crack; current cannot route around a broken finger | Object detection |
| Dark line at or along the busbar | Busbar crack | Interrupts the main current collection pathway; high severity per unit length | Object detection |
| Uniformly dark cell in the string | Dead cell | No photovoltaic response; junction failure, complete string break upstream, or cell fully shunted | Classification |
| Diffuse dimming starting at cell periphery | Potential-induced degradation (PID) | Ion migration under high system voltage; sodium from glass migrates to cell surface; early-stage PID affects edges first, progresses inward | Segmentation (area progression) |
| Dark stripe along the full busbar length | Ribbon or busbar disconnect | Poor tabbing solder joint; current path interrupted at the ribbon-to-cell interface | Object detection |
| Localised bright spot on an otherwise normal cell | Shunting | Local short circuit concentrates current; excess recombination at the shunt location | Object detection |
| Even grey reduction across a cell that passes string continuity | Carrier lifetime reduction | Precursor to dead-cell state; reduced efficiency before visible junction failure | Classification |
| Irregular dark patches at cell corners | Mechanical crack from handling | Cell fracture during stringing or laminate pressing; crack orientation does not follow crystal direction | Object detection |
Finger cracks
Finger cracks form when a cell flexes during stringing, soldering, or handling. The crack runs across the current-collecting finger network, breaking electrical continuity on the crack-shadow side. In EL, the shadow side appears as a dark triangular or wedge-shaped region bounded by the crack line and the cell edge. Power loss depends on crack orientation relative to the busbar: a crack running parallel to the busbar shadows a small area; a crack running perpendicular from a busbar to the cell edge shadows the entire portion of the cell between the crack and that edge.
Finger cracks that produce less than 3% power loss per cell are typically tolerated in module specifications. Cracks above that threshold, or cracks near the busbar that risk propagation under thermal cycling, fail the module. The inference system must calculate the shadowed area from the crack geometry — a detection task returning the crack bounding box is insufficient. Accurate power-loss estimation requires segmentation to measure the shadow area in calibrated mm².
Potential-induced degradation
PID is notable in EL inspection because it is a process signal as much as a product defect. PID onset indicates that module operating voltage, glass sodium content, or cell surface passivation is producing accelerated degradation conditions. In EL, early PID appears as a graduated darkening from the cell edges inward — cells at the end of a string, under highest reverse voltage bias, show it first.
An AI model trained on PID-progression patterns can flag modules for accelerated ageing test before they leave the factory. It can also correlate PID incidence with glass batch or cell lot numbers — a process signal that identifies a supplier quality issue before a field warranty pattern develops.
Line-speed math: inference at tabbing speed
A standard silicon module stringer operates at approximately 3,600 cells per hour. At that rate:
- 3,600 cells per hour = 60 cells per minute = 1 cell per second
- Available EL imaging dwell time: approximately 0.7–0.8 seconds per cell, accounting for mechanical positioning tolerance
- Standard 60-cell module: 3,600 cells per hour = 60 modules per hour = 1 complete module per minute
Each cell frame captured at this speed is approximately 2 megapixels at 16-bit depth in the near-infrared — roughly 4 MB per uncompressed frame. Three inference tasks run on each frame:
- Crack detection (object detection): locate and classify crack lines by type and orientation
- Dead cell and PID classification (classification): assign severity class to whole-cell anomalies
- Crack-shadow area and PID extent segmentation (segmentation): measure shadow area in mm² for power-loss estimation
All three must complete within the 0.7-second dwell window. Cloud inference is not viable at this throughput: a 100–300 ms network round-trip consumes the majority of the available window before inference even begins. Edge processing — an inference engine co-located with the EL camera enclosure — is required. The model must also queue results asynchronously: the decision for cell N must be available before cell N+1 completes its dwell, which means the inference pipeline processes each frame independently with no waiting for a previous frame to resolve.
At 60 modules per hour, the system generates 60 inference records per module and one module-level quality record per minute. Each module-level record contains the cell map (pass/fail per cell), the defect catalog (crack type, location, shadow area, PID extent), and the estimated power-loss contribution per defect. This record travels with the module through lamination and framing as the manufacturing traceability document.
What conventional EL inspection misses
Fixed-threshold EL systems compare pixel brightness against a single threshold: pixels below the threshold fail, pixels above pass. This approach misclassifies three categories of defects:
Gradual degradation onset. A cell in early-stage PID or early carrier lifetime reduction appears as a uniform gray reduction relative to neighbouring cells — not a sharp dark region. A fixed threshold set to reject obvious dead cells passes PID onset and gradual lifetime reduction. An AI model trained on the EL appearance of cells at different PID progression stages classifies the subtle graying as degradation onset before it reaches the pass/fail threshold.
Crack type misidentification. A busbar crack and a finger crack appear as similar dark lines in EL, but their power loss implications and field failure modes are different. A threshold system flags both as "dark line — fail" without distinguishing between them. The repair action is the same (cell replacement) but the process root cause is different: busbar cracks implicate stringing pressure; finger cracks implicate handling or thermal stress. Distinguishing the crack type in the defect record feeds the process control system.
Shunt localisation. A shunt appears as a bright spot in EL — a local excess emission rather than a dark region. A system calibrated to detect darkness misses shunts entirely. Shunts that fall below visible threshold today grow under thermal cycling and UV exposure in the field. Detecting them at manufacturing requires an inference model that recognises both low-emission and high-emission anomalies.
Cross-industry architecture proof
The inference challenge in EL inspection at 3,600 cells per hour is structurally identical to the inference challenge in any high-speed, high-mix production line: multi-class detection on each frame, at sub-second latency, with consistent accuracy across the natural variation in the input images. HyperQ AI Vision, operating in a Tier-1 automotive-parts deployment running 11,520 units per day at 99% detection accuracy — handling metal and plastic components with reflective surfaces, mixed defect classes, and 8,000+ product variants — is cross-industry proof that the model architecture scales to the throughput and image variability that a PV tabbing line presents. The pixel content is different; the inference architecture requirement is the same.
The 10x training data advantage — 1,000 images to reach production performance versus 10,000 for conventional systems — matters in PV inspection because rare defects like PID onset and early shunting have inherently small training datasets on any individual production line. A patented training approach that builds a production-ready model from 1,000 images makes it practical to cover the full EL defect taxonomy, including low-frequency defect classes, without a year-long data collection period.
For the nearest analogous flat-substrate inspection problem in the semiconductor context, see the semiconductor micro-crack and subsurface inspection post. Specifications for running HyperQ AI Vision on EL inspection applications are on the HyperQ AI Vision solution page.
Integrating EL inspection into the tabbing line: practical constraints
Inserting EL imaging into a tabbing line requires solving two physical problems beyond the inference challenge: triggering and enclosure.
Triggering accuracy determines image quality at line speed. The EL camera must capture the frame during the precise window when the cell is forward-biased and stationary relative to the image plane. A trigger offset of more than a few milliseconds at 3,600 cells per hour produces motion blur that obscures fine crack lines. The standard approach is a hardware trigger from the tabbing machine's PLC, tied to the cell positioning signal — software polling is not accurate enough. This requires a camera with hardware trigger input, not a USB webcam-class device with software triggering.
Enclosure is necessary because EL imaging requires near-total darkness around the imaging station. The near-infrared emission from a silicon cell under 0.1–0.2 A/cm² forward bias is weak — an illuminated factory environment floods the detector. The tabbing enclosure around the EL station must achieve sufficient light attenuation to prevent ambient IR from raising the sensor noise floor above the defect signal. Thermal emission from the tabbing machine itself can also contaminate the EL image if the camera spectral response extends into the longer near-infrared range where heat emission becomes significant. Camera spectral selection (typically filtering above 1,050 nm for crystalline silicon EL inspection) is an integration decision, not a post-installation fix.
These integration constraints are consistent across module manufacturers. They are not obstacles to implementation — they are specifications that determine what the vision integrator needs to supply alongside the AI inference system.
Malaysia and SEA solar supply chain context
Malaysia's PV module manufacturing base is concentrated in Penang and Pahang, serving export markets across Europe, the US, and Japan. Penang's manufacturing cluster includes module assembly, cell stringing, and lamination operations for multiple global module brands operating through local subsidiaries and contract manufacturers.
Module makers supplying Tier-1 export markets face increasingly specific EL inspection requirements in customer quality agreements. Customers that previously accepted a sampling-based EL protocol are adding requirements for 100% cell-level EL inspection records and defect traceability to the manufacturing lot. A system that produces per-module EL defect maps and archives them with the serial number meets this requirement. A sampling protocol with a PDF report does not.
The field warranty risk for Malaysian module manufacturers supplying rooftop solar projects in Singapore and industrial solar in Johor and Selangor is also relevant: module replacement in installed arrays is labour-intensive. A warranty claim on a 200-panel rooftop installation triggers crane hire, permitting, and logistics costs that dwarf the module's manufacturing cost. Catching the crack pattern that will propagate under thermal cycling in the tabbing line, before lamination, is the only point in the value stream where the defect is both detectable and correctable at low cost.
Honest positioning: what EL inspection does not catch
EL imaging under forward bias reveals electrically active defects and electrically inactive regions. It does not detect:
- Encapsulant delamination unless it has progressed far enough to reduce cell electrical contact
- Glass breakage that has not yet stressed the cell below it
- Junction box failures that occur after lamination
- Cosmetic surface contamination that does not affect electrical performance
For a comprehensive final inspection that covers both electrical defects (EL) and visual/cosmetic defects (surface inspection), two inspection systems run in sequence: EL in the tabbing line for electrical defect detection, and an AI vision system at final framed-module inspection for cosmetic, labeling, and dimensional conformance. EL inspection alone does not replace module-level final visual inspection.
Send a representative sample of EL images from your current tabbing line — including examples of the defect classes you are trying to capture — to Hypernology. Within two weeks we will run an inference evaluation on your image set, identify which defect classes are detectable at your line speed, and return a task-type specification and estimated detection rate. No contract until the per-defect capture rate is confirmed against your quality standard.