Machine Learning Can Be A Game Changer In Solar Panel Inspection

Key takeaways

Inspection methods in industrial processes have significantly evolved over recent decades, introducing advanced techniques like machine vision.
Machine learning is being explored for automatic inspection, particularly in the solar panel manufacturing sector.
A recent study integrated extreme learning machine with machine vision to identify solder-related defects in solar panels.
Defect identification at a micro scale presents challenges that demand detailed analysis for accurate results.
The application of machine learning in solar panel inspection could enhance manufacturing quality and reduce defects.

Over the past few decades, inspection methods in industrial processes have seen a sea of change. Diverse inspection techniques such as non-destructive testing (NDT), especially machine vision, have become game-changers. With research and technology fuelling these innovations, processes have become optimised and products have become almost defect-free once released onto the market. Of late, machine learning has prospected for automatic inspection in industries. In this article, we discuss a recent study in which an extreme learning machine (ELM) is integrated with machine vision to investigate solder-related defects and panel position in solar panels.

Defect Identification And Edge Detection

Heng Liu and the team from the Anhui University of Technology, China proposed a research study where they used ELM along with moving least square regression (MSLR) method to determine defect features as well as accurate panel positions when manufacturing solar printed circuit board (PCB) panels. Generally, a thorough inspection at this micro scale is challenging and requires detailed analysis. In their study, they also tackled another production challenge, which is the automatic positioning of the solder joints in the panel. “Specifically, for solder joint defect identification, we take histogram peaks distribution (HPD) with the morphological operation to extract image features and then apply ELM to classify the input images. As for solar panel position determination, accurate edges location is prerequisite. To achieve this, we extract the initial edge points by a novel signal processing method, fractional calculus and then utilize MLSR algorithm to refine initial edges.” That’s why Liu and the researchers’ paper discusses image preprocessing and feature extraction in detail before jumping into using ELM for detecting solder joint defects. This procedure is done using two methods:

Image thresholding
Morphological operations

The former method gives image binarization through HPD while the latter is used to process the images obtained through binarization. For solder joint images, image colour spaces are transformed from RGB to YCbCr. This is where HPD is used for binarization and then processed with morphological operations. Now, to perform initial edge extraction from the HPD binary image, fractional calculus is used. Fractional calculus, a variation of calculus, is used to determine the half-derivative (also called fractional derivative) of a function. In this study, Gaussian kernels are combined with fractional derivatives to obtain edges from the panel. (A detailed note on the mathematical calculation can be found here.) In order to identify defects in the binary solder images captured earlier, the authors extracted the image features on the basis of solder area, gravity, anisotropy alpha and moment of inertia in the binary images. A total of six features — area,x_g,y_g (gravity components), alpha, l_x,l_y (moment of inertia components) are derived for defect detection.

ELM For Detecting Defects

ELMs are generally single-hidden feed-forward neural networks which are primarily used in feature learning, among other ML techniques. ELM algorithm for a given set of image samples can be approximated with almost zero errors. “From the mathematical analysis, it is easy to see that ELM can be trained efficiently, and its prediction is fast and robust. In practice, when we apply ELM for solder joint defect detection, we first extract the mentioned features from solder joint images and input them to ELM. And the output prediction value is or . Here, the output “−1” indicates that the solder joint is defective, whereas the output “1” means that it is a good one.” Once ELM related calculations are evaluated, it is ready for experimentation. Sample defect examples are given below. [caption id="attachment_27553" align="alignnone" width="473"]

Sample examples showing solder defects in blue and good solders in red (Image courtesy: Heng Liu et.al)[/caption] Back to the edge extraction, the edges are processed furthermore by MSLR algorithm. This algorithm focuses on pixelation in the initial edges thus achieving refined and accurate edges in the images.

Experimentation And Results

For the experiments, 30 PCB solder joint images are captured with each image having 22 solder joints. This means a total of 660 solder joints in context. For ELM solder defect identification, 500 solder joints are used as training samples and the remaining 160 are used as test samples. After this, the extraction process (image features and edges) are performed. ELM is now evaluated by choosing an activation function as well as setting the hidden neuron number. The images are then trained to identify whether the solder joint is defective or not. The panel position is now determined for all the images through MSLR again. With accuracies being more than 95 percent in defect detection with a superior edge detection through the MSLR algorithm, this study has stood out as a novel way to explore defects in industrial processes. Future studies can even explore deep learning with this study acting as a reference. What more, ML could be the new gold standard for manufacturing- related testing.

Machine Learning Can Be A Game Changer In Solar Panel Inspection

Key takeaways

Defect Identification And Edge Detection

ELM For Detecting Defects

Experimentation And Results

Stay ahead with AI insights.