Redefining Quality Control With AI-Powered Visual Inspection

Emerging technology, from the introduction of assembly lines to the Internet of Things, has always defined manufacturing.

With the creation of computers and early automation came traditional machine vision, in which machines analyze photos of parts and components for defects based on a set of humandefined rules. While it reduces human error, traditional machine vision lacks the capacity to solve for pain points like complex defects and changing environments.

Today, more sophisticated artificial intelligence (AI), including machine learning (ML) and deep learning (DL), allows manufacturers to use AI-powered visual inspection to enhance quality and reduce costs. But even now, only 5% of manufacturing companies have a clearly defined strategy for implementing AI.

Companies need strategies to overcome challenges in visual inspection, which still relies heavily on human inspectors or inflexible rules based machine vision. The cost of sending defective pieces to customers, both in reputation and in recalls, isn’t sustainable in a competitive global environment.

The right AI platforms offer tools that can enhance quality control and cut costs after users tackle key obstacles.

From Proof of Concept (PoC) to Production

Manufacturing companies can successfully create a proof of concept (PoC) of a visual inspection system in a few weeks or even a few days. But getting to a deployable solution ready for production and then scaling it threatens to bring manufacturers to a standstill.

Arriving at a PoC, which generally takes the form of offline tests run under highly controlled conditions, is a major milestone, but developers are still a long way from successful deployment. At this point, manufacturers only have less than 10% of the software needed for the first deployment, and the first deployment is a small fraction of the software needed to scale to multiple production lines. Teams need to carefully plan, prepare and execute each step of AI deployment.

Too often, companies fail to scale solutions beyond an initial project or two. This is particularly pervasive in manufacturing due to the complex and unique nature of each project. In conventional detection built on rigid rules, you’ll need to invest massive amounts of time and money to adapt thousands of lines of code to account for small details and variables.

Manufacturers must overcome a uniquely complex set of obstacles to deploy and scale AI visual inspection systems, causing many projects to delay or fail.

Small Data Sets

In an industry focused on preventing defects, it’s difficult to implement AI to capture actionable insights from a small dataset because defects happen a fraction of a percent of the time. Unlike consumer web companies like Google and Amazon that can apply data from billions of users to train powerful AI models, collecting massive training sets in manufacturing is often not feasible. 58% of research respondents report the most significant barrier to deployment of AI solutions was a lack of data resources.

Manufacturers may have 100, 10 or even fewer images of a particular defect they want to detect. For example, in automotive manufacturing, where Lean Six Sigma practices are nearly universal, most OEMs and tier-one suppliers strive for fewer than three to four defects per million parts. Tools designed to work for big data can’t function with this small amount of data.

When there are only a few examples of a problem, AI models are difficult to train. This prevents companies from scaling or
solving for natural variance in environments.

Ambiguous Defect Requirements

Identifying a defect can be subjective, and it’s common for two inspectors to disagree on qualifications. One inspector may consider a scratch to be problematic, while another thinks the same scratch is small enough to be ignored. When even experts are in disagreement, how can we expect AI trained by humans to perform?

There’s a high rate of error between inspectors: Many studies set error rates for manual inspections between 20% and 30%, meaning as few as 70% of defects are caught by human inspectors.

Unlike big data settings like the software industry, where users can average the responses of millions of inspectors, a manufacturing setting may take the average of a judgement call from two or three experts and it is ineffective. Often, manufacturers opt for using automation such as traditional machine vision to compensate for the shortcomings of human inspection, but this comes with its own tradeoffs. Given its low accuracy, traditional machine vision forces users to permit high overkill rates (the percent of products marked defective that are actually acceptable) to prevent escape (the number of bad parts that aren’t caught).

The rate of overkill or false positives could be as high as 40% for many manufacturers, which forces inspectors to physically inspect rejected parts on the manufacturing line.

Modernizing manufacturing means accelerating efficiency through the implementation of new tools without undermining the importance of human employees. Creating AI-powered models that accurately detect defects empowers workers by giving them more time to solve the root causes of defects. Here, workers add more value, while companies save costs and boost efficiency.

While an end-to-end AI platform that enables better quality control may sound like a dream come true, don’t implement AI for the sake of keeping up with the masses. Speeding up digital transformation without completing the necessary research and preparation often results in failed projects.

The above is an extract from a white paper published by Landing AI. Click to download the full white paper.

For more information: www.landing.ai