Deep learning (DL) has been the pillar of innovations in artificial intelligence. It has been providing a powerful way to compute numerous tough tasks. Now, many companies including Facebook and Google are leading DL research to strengthen their offerings.
With the ever-growing research, DL is now venturing into the manufacturing process. This article highlights a recent study which used DL along with image processing to predict a laser’s highest energy spot in a welding machine.
Convolutional Neural Network To The Rescue
Researchers at Feng Chia University, Taiwan have published a study where a convolutional neural network DL framework improves laser positioning, which positively affects machine efficiency. The study actually uses image processing to take the laser energy into account. A particular manufacturing process called template matching is the focus here.
Template matching, simply means an image (of a part) is usually matched with a template in image processing. To perform this, many imaging systems comprising a large set of images are taken. It may seem easy to do this in various applications such as creating intricate parts or models, but it takes a toll on the cost. High costs are not generally acceptable in areas of precision manufacturing.
In the study, CNNs identify laser spots of high intensity in template images obtained through a charge-coupled device (CCD) cameras. These neural networks are trained on both complete and incomplete laser light spots. Researchers highlight how the light spots are analysed:
“There are two processing modes, and whether the found light spot is complete or not is judged according to the aspect ratio of the light spot region. If it is a complete light spot, the centroid point of the light spot will be extracted using the centroid method, the distance between the CCD image centre and the light spot centre will be calculated, and the image centre will be moved to overlap the light spot. If it is not a complete light spot, the invariant moments will be calculated for the incomplete light spot region in the image and matched with the preset light spot source image to find out the position of the partial light spot region centroid in the light spot. The distance between the CCD image centre and the light spot centre is calculated, and the image center is moved to overlap the light spot centre.”
This simply means that the centres of the images are adjusted to form a consistent light spot. Also, there is another factor called ‘invariant moments’. This helps match the image to the algorithm (CNN) through weights presented as central moments. They, in turn, consider centroids of the images. A total of seven invariant moments are presented by the researchers which possess rotation, scaling, and translation invariances. This is critical because neural networks understand image qualities based on these invariant moments.
So instead of using expensive imaging systems, a CCD camera and a CNN setup does the job here. Even though this study saw a major setback such as misrecognition due to overuse of the same neural network for both the types of light spots, the accuracy of determining laser positioning through this method has a catch. Unlike regular template matching, this is way faster. If more improvement is made on accuracy, it could altogether be a new standard method in manufacturing.
Being A New Standard
This study completely changes the way template matching is perceived. In most of the applications in machining, a standard apparatus is utilised right from the inception to termination of the process. By introducing deep learning in these manufacturing processes, complexities related either internal to the processes or are external in nature (like market fluctuations and cost) and can be reduced indirectly.
In spite of the fact that all of this being seemingly abstract, incorporating DL also reduced the need for extensive image analysis. Furthermore, image processing, which was also a part of this study, gets a boost in computing and does not require to be conducted in depth. This study is a typical example. As you can see, laser welding will be way more efficient than the traditional setup used in the industry.
Try deep learning using MATLAB