MultiSolSegment: Multi-channel segmentation of overlapping features in electroluminescence images of photovoltaic cells
Authors: Ojas Sanghi, Norman Jost, Benjamin G. Pierce, Emma Cooper, Isaiah H. Deane, Jennifer L. Braid
Organizations: Sandia National Laboratories · Sandia National Laboratories, 1515 Eubank Blvd SE, Albuquerque, New Mexico, 87123, USA · Case Western Reserve University · Case Western Reserve University, 10900 Euclid Ave, Cleveland, Ohio, 44106, USA · University of Colorado, Boulder · University of Colorado, Boulder, Boulder, Colorado, 80309, USA
Electroluminescence (EL) imaging is widely used to detect defects in photovoltaic (PV) modules, and machine learning methods have been applied to enable large-scale analysis of EL images. However, existing methods cannot assign multiple labels to the same pixel, limiting their ability to capture overlapping degradation features. We present a multi-channel U-Net architecture for pixel-level multi-label segmentation of EL images. The model outputs independent probability maps for cracks, busbars, dark areas, and non-cell regions, enabling accurate co-classification of interacting features such as cracks crossing busbars. The model achieved an accuracy of 98% and has been shown to generalize to unseen datasets. This framework offers a scalable, extensible tool for automated PV module inspection, improving defect quantification and lifetime prediction in large-scale PV systems.
Figures & tables
Figure 1: High-level overview of model flow. Made with Iqbal [2018] . Yellow/orange blocks denote convolutional feature maps at each resolution, with block width indicating the number of channels and block height indicating spatial resolution. Darker orange/red inner regions highlight the activated feature representations after convolution and ReLU. Blue-tinted blocks in the decoder indicate upsampled feature maps produced by transposed convolutions. Semi-transparent blocks represent feature concatenation via skip connections between corresponding encoder and decoder stages. The final purple block denotes the output layer with per-channel sigmoid activations, producing independent probability maps for each defect class. Arrows indicate the forward flow of information, while long horizontal connections visualize encoder-decoder skip connections.
Dataset Name
# Images
# Total Images (Augmented)
# Images (Training Set)
# Images (Validation Set)
Arizona State University
92
368
1,872
468
Case Western Reserve University - DuPont
123
492
Case Western Reserve University - SunEdison
250
1000
Lawrence Berkeley National Laboratory
120
480
Total:
585
2340
Table 1: Overview of Datasets Used to Train Model
Class Name
# Images
Dark
364
Busbar
2340
Crack
1688
Non-Cell
2340
Table 2: Representation of different EL image features in our augmented dataset.
Property
Value
Cardinality (mean number of labels per pixel)
0.241
Density (mean number of labels per pixel / number of labels in the dataset)
1.971e-09
Single label pixels (percentage of pixels in the dataset with only one active label)
Table 3: Dataset characteristics and frequencies for each mask channel
Figure 2: High accuracy results from our model
Metric
Value
Accuracy (Test Set)
0.982
5-Fold CV Accuracy (mean ± SD)
0.986740 ± 0.000461
Precision
0.550
Recall
0.507
Dice
0.503
Intersection over Union (IoU)
0.384
Table 4: Performance metrics for the best model, averaged across all classes
Figure 3: Predictions of MultiSolSegment (MSS) compared to PV-Vision (PVV)
Figure 4: Metrics of MultiSolSegment compared to PV-Vision
Figure 5: Quantitative detected crack count of MultiSolSegment compared to PV-Vision
Figure 6: Example of a test case where the model performs poorly
Figure 7: Low accuracy of a transfer-learned pv-vision
Figure 8: High accuracy of a transfer-learned pv-vision model with the non-cell class
Figure 9: Lack of multi-labeling in our single-channel model. (a) Original EL image. (b) Model prediction using single-channel classification; each pixel assigned to only one of five mutually exclusive classes. (c) Crack target mask from the "double mask" training setup, where overlap pixels are labeled as both crack and busbar. (d) Busbar target mask from the "double mask" training setup. (e) Model activiation map for the crack class, showing predicted probability at each pixel. (f) Model activation map for the busbar class.
Figure 10: Activation values of pixels at y = 130 for the different classes. On the right is the raw model output for the entire image, with the dashed line indicating y = 130. On the left are the class activation values along this horizontal line.