What question did this study set out to answer?

To develop a robust and scalable model for automatic detection and quantification of dust on solar panels.

May 9, 2026e-Prime – Nexus of Electrical Electronic and Intelligent Engineering0 citationsOpen Access

DustMambaNet: A Hybrid InceptionV3–State-Space Network for Robust Solar Panel Dust Detection

Key Points

To develop a robust and scalable model for automatic detection and quantification of dust on solar panels.
Hybrid model combining InceptionV3 and selective Mamba blocks.
Used binary classification for dust detection and introduced a Dust Severity Index.
Trained on about 12,440 images using initial training and fine-tuning phases.
Achieved 93.67% accuracy, 0.960 precision, and 0.927 F1-score on the test set.
Outperformed seven baseline models by over 3% accuracy.
Demonstrated superior local texture and global dust pattern detection.

Abstract

• Proposes DustMambaNet, a hybrid CNN-state-space model of dust-detecting on solar panels. • Combines InceptionV3 and selective Mamba blocks in order to model long-range efficiently. • Attains an accuracy of 93.67% and high F1-score compared to seven base models. • Introduces a Dust Severity Index for quantitative and actionable maintenance decisions • Scalable and computationally inexpensive monitoring of large solar farms. To develop a robust, scalable vision-based model for automatic detection and quantification of dust accumulation on solar photovoltaic panels, overcoming limitations of existing convolutional and attention-based methods and supporting proactive maintenance. We propose DustMambaNet, a hybrid model that consists of a pretrained InceptionV3 convolutional neural network as a feature extractor and two selective state space sequence modules. The state space modules use gated depthwise convolutions to represent long-range spatial dependencies that are of linear complexity, after rearranging spatial features to form sequences. The network provides a binary classification of dust with a severity index (DSI) and a continuous one. All dusty and clean panel images were resized to a size of 299×299 pixels, augmented, and divided into training and validation subsets, comprising about 12 440 images. Training consisted of two phases: initial training of the Mamba blocks and head with the backbone frozen, followed by fine-tuning of the top convolutional layers. On the test set, DustMambaNet achieved a mean accuracy of 93.67 %, precision of 0.960, recall of 0.897, and F1-score of 0.927, outperforming seven baseline models by more than 3 % accuracy and exhibiting superior Cohen’s kappa and Matthew’s correlation coefficients. The hybrid network that has been proposed is very efficient in capturing the local texture and global dust patterns distribution, and gives very dependable dust detention and severity measurement. Its scalability and efficiency allow its use in a large-scale solar farm, and the Dust Severity Index provides actionable information on how to schedule the cleaning of panels. DustMambaNet illustrates the power of combining a convolutional and state space model in the maintenance of photovoltaic systems.

问 AI

Bookmark

View Full Paper