Automation has transformed 20th-century production lines. Now, a combination of robotics and artificial intelligence is taking automation to the next level, the smart factory. Artificial intelligence in the manufacturing market is estimated to be valued at USD 1.0 billion in 2018 and is expected to reach USD 17.2 billion by 2025, at a CAGR of 49.5% from 2018 to 2025.
Artificial Intelligence in the manufacturing sector
Image recognition helps manufacturing industries to streamline quality control processes. However, detection cameras are not enough as they spot only a few defects that are predefined by programmers. The AI-powered computer-vision based solution is certainly more advanced. According to McKinsey, 50% of companies that start using AI over the next five to seven years have the potential to double their cash flow with manufacturing leading all industries due to its heavy reliance on data. In conventional image recognition systems, each defect has to be anticipated and pre-programmed by a consultant before the system intercepts it; it is thus an expensive process.
With AI-powered systems, no anticipation or pre-programming will be required, and one algorithm will be able to do the job of many. According to Forbes, automating quality testing using artificial intelligence is increasing defect detection rates up to 90%.
Zensar’s solution for automating defect detection
Our propriety solution can automate defect detection in the manufacturing assembly line using computer vision. This solution identifies whether the product is defective or not using the image of the product surface. It’s capable of finding the location and the kind of defect (class/reason of defect).
This solution is trained to detect defects on various products by pre-fed data and real-time inputs provided by the production process. It enables the system to identify and categorize the errors successfully. It helps identify/rectify the surface grit and defect over any product in a completely automated way.
Take a look at this example in the case of steel sheets.
The business impact created by Zensar
Among many industries, visual inspection on the assembly line is a very high consequence and high priority process due to the high cost of errors, possible loss of expensive equipment, or loss of customers. Our solution can be leveraged by manufacturing industries to be used with their production data with minimum support. With an AI-enabled solution, it could be useful for production in the following ways.
1) Quality Assurance – Improvement in quality and customer experience by reducing the defected products.
2) Improving Production Efficiency –Decrease in throughput time, and production can function at an exceptional speed while lowering costs.
3) Reduction in Human Errors – Reduction in human errors with the highly accurate detection module.
4) Reduction in Manual Effort – No human intervention is required.
Our solution finds defects on steel/metallic products, VLSI wafers, or inspecting automobile parts for defects. The solution can also be used in the insurance sector for car damage assessment to find out the type of error, whether it is accidental, scratches, etc.
How is Zensar solving this problem?
Our two-fold supervised model includes a classifier for finding out whether the product is defective or not and an architecture for localizing the defects. This model is generalized and works for any manufacturing data. But initially, we used the steel sheets defect detection data so that we can design a baseline model on which we can build a generalized solution later. Here’s a diagram of the model.
The surface of the product is scanned to find out whether any defects or damage was incurred during production. Computer vision techniques are fundamental here that stimulate the process of surface defect analysis through images.
We’ve used a U-Net architecture, a semantic segmentation technique for localizing the defect and classifying the class of defects.
Semantic Segmentation: The goal of semantic image segmentation is to label each pixel of an image to a class label. Since we are predicting for every pixel in the picture, it is referred to as dense prediction. The output itself is an image in which each pixel is classified into a class. Thus, it is a pixel-level classification.
UNET Architecture: The U-Net architecture is built upon the Fully Convolutional Network, which contains two parts that makes it a u-shaped architecture. The first path is an encoder which captures the context in the image using a stack of convolutional and max-pooling layers. The second path is a decoder, which exacts location using transposed convolutions. The skip connection between the first and the second path is used for getting back the image with class labels. U-Net architecture is highly used in biomedical image segmentation.
In some cases, even after implementing the U-Net architecture, there can be cases where even non-defective image gets segmented. To remove such instances, we came up with a ResNet based transfer learning classifier to mark an object as defective or not. It is a model trained on ImageNet dataset, and we use it based on transfer learning on our dataset and learn whether the product is defective or not.
The main challenge here is two-fold – the first one is to design features that are suitable for the problem. The second one is designing classifiers that can learn the boundaries in the feature space that separate defects from non-defects. We have solved both the problems with our model.
To supplement our solution, we are working on three algorithms that provide additional features like-
- Precision-based algorithm for use cases which requires quality products.
- Generalized solution which works independently of the product. Precision-based algorithm for use cases which requires quality products.
- Generalized solution which works independently of the product.
As the companies around the world automate their assembly lines, defect detection is mostly done manually due to the numerous type of defects that are hard to detect and analyze by machines. However, with the help of artificial intelligence, defects of various kinds and intensity can be discovered by training the defect detection algorithms. Zensar’s Automatic Defect detection tool is the state-of-the-art solution to help companies minimize human errors, increase efficiency, and reduce costs by eliminating human involvement and yet identify the defects with high precision.