Deep learning based automatic defect identification of photovoltaic module using electroluminescence images

Defect detection and quantification in electroluminescence images of solar PV modules using U-net semantic segmentation

Efficient deep feature extraction and classification for identifying defective photovoltaic module cells in Electroluminescence images

PV module power prediction by deep learning on electroluminescence images - Assessing the physics learned by a convolutional neural network

Electroluminescence (EL) imaging is one of the most common characterisation techniques for photovoltaic cells and modules. EL images contain information on both the open-circuit voltage (Voc) and series resistance (Rs) of the device. However, separating the two effects and identifying features related to each parameter can be challenging. In this study, a novel approach for decomposing EL images into Voc and Rs maps using a convolutional neural network architecture is presented. A deep learning model was first trained on paired EL and photoluminescence images that were generated using a simulation tool. Results obtained using the validation set show that the trained model is able to accurately differentiate between features related to Voc and Rs in EL images, thus replacing the need for multiple types of measurements. The proposed method presents a unique approach to analyse EL images, unlocking new capabilities that have the potential to advance solar cell characterisation.

Accurate current–voltage (IV) characterization is essential for assessing photovoltaic (PV) module performance, yet conventional IV tracing requires physical contact and controlled conditions, limiting large-scale deployment. Electroluminescence (EL) imaging, while highly effective for detecting localized defects, remains largely qualitative and indirect in estimating actual PV module power loss. This study introduces a deep learning framework that directly predicts complete IV curves from EL images, transforming EL inspection into a quantitative, non-contact diagnostic tool. In this work, we propose a convolutional neural network (CNN) that learns the nonlinear mapping between paired EL images captured at 20% and 80% of the short-circuit current and the corresponding IV response. A total of 438 PV modules were used for model development, with performance evaluated on unseen data. The trained CNN reconstructs IV curves with high fidelity, achieving a validation accuracy of approximately 95% and low parameter deviations (<2% for key metrics such as maximum power point and fill factor). The model maintains consistent accuracy even when a single EL image is provided, supporting flexible field operation. Inference is rapid, requiring less than 0.5 s per PV module inspection, enabling real-time analysis.

This review paper presents a comprehensive analysis of electroluminescence (EL) imaging techniques for photovoltaic (PV) module diagnostics, focusing on advancements from conventional indoor imaging to outdoor and daylight EL imaging. It examines key challenges, including ambient light interference and environmental variability, and highlights innovations such as infrared-sensitive indium gallium arsenide (InGaAs) cameras, optical filtering, and periodic current modulation to enhance defect detection. The review also explores the role of artificial intelligence (AI)-driven methodologies, including deep learning and generative adversarial networks (GANs), in automating defect classification and performance assessment. Additionally, the emergence of drone-based EL imaging has facilitated large-scale PV inspections with improved efficiency. By synthesizing recent advancements, this paper underscores the critical role of EL imaging in ensuring PV module reliability, optimizing performance, and supporting the long-term sustainability of solar energy systems.

Failure signature classification in solar photovoltaic plants using RGB images and convolutional neural networks

Solar panels play a crucial role in converting solar energy into electricity, with PhotoVoltaic (PV) modules being their core components. To ensure solar panels function well, efficient and accurate defect detection of PV modules is essential. Visual-based deep learning detection methods, such as Transformer and Convolutional Neural Network (CNN) models, provide a cost-effective and adaptable solution. Although the Transformer-based detectors outperform CNN-based ones in detection tasks, their complicated model structures also lead to great computational cost and latency, limiting their application in resource-constrained environments or time-sensitive tasks. To improve the real-time performance of the DEtection TRansformer (DETR) model on these edge devices, a novel Lightweight Decoder-only DETR (LD-DETR) is proposed for defect detection of PV modules, which achieves high performance and low computational cost with an easy architecture. We remove the resource-intensive Transformer Encoder structure from the DETR baseline so that the Decoder can directly utilize the finely fused features produced by a specialized lightweight convolution module that includes an upsample operation to preprocess the features. Experiments demonstrate that the proposed LD-DETR achieves competitive performance on the PV cell dataset of ElectroLuminescence (EL) images with an accuracy of 87.4% and an inference time of 6.1ms on GPU as well as 1.1s on embedded CPU. Our source code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/SeveralYang/LD-DETR</uri>.

In this work, deep convolutional neural networks (CNNs) are used to speed up the calculation of spatial nonuniform device parameters, voltage and fill factor (FF) losses in crystalline Si solar cells. Luminescence images on finished solar cells, specifically as‐measured electroluminescence (EL) and photoluminescence (PL) images, are used as input features to train the CNN models and losses analysis as output is from a finite‐element modeling program called Griddler. From the analysis of a small dataset of 250 commercial grade solar cells, the CNN models are able to predict the spatially nonuniform distribution of contact resistance and defect parameters of the devices under test and exhibit good performance in the inference of voltage and FF losses. As EL and PL images are widely collected in‐line production data, in the result, the possibility of deploying 2D spatial power loss analysis is implied as a fast, in‐line, and nondestructive analytic tool for solar cells manufacturing.

An Uncertainty-Aware Deep Learning Framework for Defect Detection in Casting Products

ObjectiveBrain atrophy is an important imaging characteristic of cerebral small vascular disease (CSVD). Our study explores the linear measurement application on CT images of CSVD patients and develops a fully automatic brain atrophy classification model. The second aim was to compare it with the end-to-end Convolutional Neural Networks (CNNs) model.MethodsA total of 385 subjects such as 107 no-atrophy brain, 185 mild atrophy, and 93 severe atrophy were collected and randomly separated into training set (n = 308) and test set (n = 77). Key slices for linear measurement were manually identified and used to annotate nine linear measurements and a binary classification of cerebral sulci widening. A linear-measurement-based pipeline (2D model) was constructed for two-types (existence/non-existence brain atrophy) or three-types classification (no/mild atrophy/severe atrophy). For comparison, an end-to-end CNN model (3D-deep learning model) for brain atrophy classification was also developed. Furthermore, age and gender were integrated to the 2D and 3D models. The sensitivity, specificity, accuracy, average F1 score, receiver operating characteristics (ROC) curves for two-type classification and weighed kappa for three-type classification of the two models were compared.ResultsAutomated measurement of linear measurements and cerebral sulci widening achieved moderate to almost perfect agreement with manual annotation. In two-type atrophy classification, area under the curves (AUCs) of the 2D model and 3D model were 0.953 and 0.941 with no significant difference (p = 0.250). The Weighted kappa of the 2D model and 3D model were 0.727 and 0.607 according to standard classification they displayed, mild atrophy and severe atrophy, respectively. Applying patient age and gender information improved classification performances of both 2D and 3D models in two-type and three-type classification of brain atrophy.ConclusionWe provide a model composed of different modules that can classify CSVD-related brain atrophy on CT images automatically, using linear measurement. It has similar performance and better interpretability than the end-to-end CNNs model and may prove advantageous in the clinical setting.

Production of food crops is hampered by the proliferation of crop diseases which cause huge harvest losses. Current crop-health monitoring programs involve the deployment of scouts and experts to detect and identify crop diseases through visual observation. These monitoring schemes are expensive and too slow to offer timely remedial recommendations for preventing the spread of these crop-damaging diseases. There is thus a need for the development of cheaper and faster methods for identifying and monitoring crop diseases. Recent advances in deep learning have enabled the development of automatic and accurate image classification systems. These advances coupled with the widespread availability of multispectral aerial imagery provide a cost-effective method for developing crop-diseases classification tools. However, large datasets are required to train deep learning models, which may be costly and difficult to obtain. Fortunately, models trained on one task can be repurposed for different tasks (with limited data) using transfer learning technique. The purpose of this research was to develop and implement an end-to-end deep learning framework for early detection and continuous monitoring of crop diseases using transfer learning and high resolution, multispectral aerial imagery. In the first study, the technique was used to compare the performance of five pre-trained deep learning convolutional neural networks (VGG16, VGG19, ResNet50, Inception V3, and Xception) in classifying crop diseases for apples, grapes, and tomatoes. The results of the study show that the best performing crop-disease classification models were those trained on the VGG16 network, while those trained on the ResNet50 network had the worst performance. The other studies compared the performance of using transfer learning and different three-band color combinations to train single- and multiple-crop classification models. The results of these studies show that models combining red, near infrared, and blue bands performed better than models trained with the traditional visible spectral band combination of red, green, and blue. The worst performing models were those combining near infrared, green, and blue bands. This research recommends that further studies be undertaken to determine the best band combinations for training single- and multi-label classification models for both crops and plants and diseases that affect them.

In recent years, many researchers have focused on using convolutional neural networks to perform human activity recognition as evidenced by the emergence of a number of convolutional neural network architectures such as LeNet-5, AlexNet and VGG16 and modern architectures such as ResNet, Inception V3, Inception-ResNet, MobileNet V2, NASNet and PNASNet. The main characteristic of a convolutional neural network (CNN) is its ability to extract features automatically from input images, which facilitates the processes of activity recognition and classification. Convolutional networks indeed derive more relevant and complex features with every additional layer. In addition, CNNs have achieved perfect classification on highly similar activities that were previously extremely difficult to classify. In this paper, we evaluate modern convolutional neural networks in terms of their human activity recognition accuracy, and we compare the results with the state of the art methods. In our research, we used two public data sets, HMDB (Shooting gun, kicking, falling to the floor, punching) and the Weizman dataset (walking, running, jumping, bending, one hand waving, two-hand waving, jumping in place, jumping jack, skipping). Our experimental results indicated that the CNN with NASNet architecture achieves the best performance of the six CNN architectures on both human activity data sets (HMDB and Weizman).

Electrocardiogram (ECG) recording of electrical heart activities has a vital diagnostic role in heart diseases. We propose to tackle the problem of arrhythmia detection from ECG signals totally by a deep model that does not need any hand-designed feature or heuristic segmentation (e.g., ad-hoc R-peak detection). In this work, we first segment ECG signals by detecting R-peaks automatically via a convolutional network, including dilated convolutions and residual connections. Next, all beats are aligned around their R-peaks as the most informative section of the heartbeat in detecting arrhythmia. After that, a deep learning model, including both dilated convolution layers and a Long-Short Term Memory (LSTM) layer, is utilized to recognize arrhythmia. Indeed, the segments centered around R-peaks acquired from the previous step are fed into this network to distinguish various arrhythmias. The LSTM part of the proposed network enables modeling the relation among different heartbeats in a sequence. Experiments on the MIT-BIH databases and Creighton university ventricular tachyarrhythmia show the superiority of our proposed method on arrhythmia detection in comparison with the recent methods proposed for this problem. The performance of the proposed model on test samples is 98.93%, 99.78%, and 99.58% respectively in terms of overall accuracy, sensitivity, and specificity for tackling the problem of 4-class arrhythmia classification. Thus, it outperforms other recent methods with a large margin in terms of accuracy and specificity.

In this paper, alternative direction finding methods have been proposed using deep learning techniques. Firstly, Regeression and Classification models have created by using Convolutional Neural Networks (CNNs). In the second Convolutional Neural Networks and Recurrent Neural Networks (RNNs) have been utilized in the proposed methods. Despite having vast amount of direction finding studies, utilization of neural networks is scarce in literature and past works mostly only includes usage of CNNs. In this study, direction finding is performed by learning signals reaching multiple antenna arrays by networks. Created neural networks have been fed with different data formats and their performances against noisy and no-noise data have been shown. In addition, comparative analysis of the developed models were made in the similar Signal-to-Noise Ratio (SNR) range with the subspace based MUSIC algorithm, which is frequently used in direction finding.