AI Improves Automation System Performance and Versatility

Machine vision has long been a powerful tool in manufacturing — from inspections and quality to control and robotic guidance applications. But the latest advances in artificial intelligence are taking the power of machine vision to new levels.

AI, including machine learning (ML) and deep learning (DL), is transforming the way inspections are conducted. As AI technology advances, so does the range of inspection applications that can be automated. It’s also revolutionizing vision guidance for robotics, allowing robotics to pick and manipulate objects with more human-like precision.

The Case for Utilizing AI Vision Systems

AI vision systems have become the best technology choice for inspecting high-frequency, random, and complex defects. Traditional machine vision technologies, which use rule-based algorithms, are inferior in such applications. The algorithms require human configuration of thousands of rules, which requires thousands of examples to validate edge cases. This is incredibly time-consuming and labor-intensive, thus resulting in deployment times that are too long to be practical.

“In these types of applications, it’s impractical to deploy conventional machine vision on new lines without spending substantial time handcrafting thousands of rules,” says Keven Wang, CEO at UnitX. “Additionally, every time the environment or incoming material changes, the required reconfiguration takes too much manual overhead. For the first time, AI is now robust enough to repeatably recognize defects of random features, such as random shapes, sizes, and locations.”

Due to the limitations of traditional rule-based machine vision and despite labor shortages many manufacturers still rely on manual visual inspection. Yet, manual inspectors take time to hire and train and are hard to retain. They’re more expensive, too, with rising wages due to labor shortages brought on by Covid and an incoming labor force disinterested in such a monotonous role.

It's also difficult for factories to retain a visual inspector long-term. When visual inspectors leave, they leave with all their experience and knowledge accumulated over the years. That’s why inspection is one of the most promising applications for AI. AI not only improves product quality and the well-being of human inspectors but also reduces scrap and increases capacity.

Opportunities for AI Implementation: Inspection and Robotic Operations

AI is utilized in two primary ways in manufacturing, according to Sina Afrooze, cofounder and CEO of Apera AI.

“Firstly, it is used in inspection processes. Secondly it is employed to automate tasks through robotic operations,” he explains. “In both applications, AI significantly enhances reliability and the ability to handle challenging scenarios, including edge and corner cases. This leads to reduced downtime in automated systems, making it one of the fastest-growing expansions in manufacturing.”

The opportunity in the manufacturing industry is driven by a rapidly growing demand for automation, which stems from various macroeconomic factors, such as labor shortages, nearshoring, onshoring, supply chain disruptions, and geopolitical events. These factors have led manufacturers to reconsider where they build their products, creating a strong demand for automation even with tasks that were previously considered difficult or uneconomical to automate.

AI serves as the crucial enabler for making these automation tasks reliable and cost-effective. Additionally, the use of AI in robotic, vision-guided manipulation of objects and inspection leads to higher-quality production systems with fewer errors (either false positives or false negatives).

“Humans can easily identify issues like broken glass, cracked stamped parts, or misshaped plastic injection-molded components, highlighting the potential for AI in enhancing these processes,” says Afrooze. “AI offers the opportunity to combine vision guidance and quality inspection into a single product, providing similar value to manual operations for manufacturers. This ultimately translates to an improved customer experience for the manufacturer.”

Chris Kennedy, director of partnerships and marketing at Prolucid Technologies, shares that the most promising AI/ML applications solve narrowly defined problems that have a definitive problem statement and an addressable data requirement. “You don’t necessarily need an enormous dataset on day one to be successful,” he says, “but the generation of proper training and validation datasets needs to be factored into the timing and budgeting process for a project that includes an artificial intelligence or machine learning component.” 

In these areas, AI helps analysts and quality assurance professionals work more efficiently and productively. Properly trained AI models can quickly become more accurate than human input, so when QA or inspection professionals use AI tools effectively, productivity increases dramatically.

AI Challenges

AI is transforming manufacturing inspection and robotic guidance applications, but a number of challenges need to be addressed before it can be widely adopted. Manufacturers who can address these challenges will be well positioned to reap the benefits of AI, like improved quality, increased productivity, and reduced costs.

Time and Resource Cost

Among the challenges, cost can be a big one. “When trying to solve problems, we generally use AI/ML as a last option for when traditional rule-based algorithms and techniques cannot solve the problem alone,” explains Kennedy. “That being said, a significant increase in availability of open-source tools and libraries is making it possible to solve an increasing number of difficult problems. Over time, this is having a cost-reducing effect while speeding up AI/ML application development because there are an increasing number of pretrained models available that make more specific problems solvable. That, combined with the cloud tools and micro-services available from providers like AWS, Azure, and GCP, makes more rapid model training and deployment possible.”

Competing Quality and Production Goals

Another challenge is aligning the right stakeholders to make the investment in AI inspection. Visual inspection, explains Wang, can be a contentious topic inside a factory because it is caught between two departments: production and quality. Production cares about meeting volume quota on time. Quality cares about zero escapes. The two departments are designed to have a separation of power and maintain a healthy tension between quantity vs quality.

AI Model Training and Process Drift

AI inferences are only as good as their models. Manufacturers using AI inspection products need to ensure that AI is trained correctly. Some AI inspection products need to train on hundreds or thousands of images to build and develop models. This can be untenable for new lines or lines where the products change frequently.

Manufacturing processes are subject to drift over time and introduce new variances in features not seen before. It can be challenging for AI solutions to recognize the drift and rapidly adapt. AI models are only as good as the data fed into them, and AI inspection solutions need high-quality images for models to perform as accurately as possible.

The quest for optimizing models will become even more critical, particularly for larger models that communicate with smaller models in creation of a blended inferencing as well as the regeneration and retraining of smaller models to be deployed at the edge. The AI models that are deployed on the edge devices should be optimized for the constraints and requirements of the edge environment as well as to minimize power and memory.     

Personnel Availability

The perception that AI experts are needed — all the time — to deploy AI is one of the biggest challenges sabotaging large-scale AI deployment. Yes, the learning curve to master AI is relatively steep. But as AI has become more understandable and user-friendly over the past few years, much progress has been made. However, AI is still the province of experts, with many nuances involved in turning a fortunate experiment/proof of concept into a real-world deployment.

“There are simply not enough AI PhDs in the world to serve all areas of the economy that need them,” says Max Versace, CEO and co-founder at Neurala. “Currently, U.S. universities graduate around 3,000 PhDs in AI-related fields per year, with a median of 5.8 years to complete a PhD. And PhDs can’t be accelerated. Moreover, a ‘fresh’ AI PhD still needs years of work ‘in the trenches’ to gain the domain-specific knowledge to be effective in an applied, complex context such as industrial manufacturing.”

Striving for Transparency

While AI is a broad spectrum and efforts have been made to make the decision-making process more explainable, it can be perceived as complex and opaque, leading to unease among those who view it as a “black box.” For this reason, Apera has consciously incorporated elements into its AI to enhance explainability.

“We strive to move toward a more transparent approach, enabling users to understand why the AI makes certain decisions in tasks such as inspection and robot guidance,” says Afrooze. “While progress has been made, making AI more explainable remains a challenge and a shared goal in the industry to increase manufacturers’ comfort with using AI.”

There is a trend toward adopting standardized acceptance tests based on actual performance, regardless of the complexity or explainability of the underlying algorithm. As long as the system reliably operates in a factory setting and manages controlled failures, manufacturers are less concerned about the specific decision-making process.

Apera provides safety checks that are less intricate than the core system, allowing for the identification of potential rare catastrophic failures. Afrooze notes, “Although these occurrences are infrequent (for example, 1 in 10,000 or 1 in 100,000), it remains important to protect the environment, robot, and tools as well as account for the human intervention that is needed to resolve such situations.”

AI Implementation Tips

Adopting AI for automated inspection is a journey, and manufacturers should take one step at a time. Manufacturers need to test the technology, build the business case, and lean on experts. Manufacturers also need to start their AI journeys early because adopting AI takes time. It takes education across the entire workforce to familiarize stakeholders with AI. This includes operators, quality, production operations, and manufacturing engineering.

Start with a specific problem that needs solving and evaluate all options, including traditional non-AI solutions. There is a wide range of technologies and solution providers, all approaching problems from different angles. If AI is indeed the correct answer, be prepared to put in a significant amount of resources early to end up with a system that is robust and improves over time. Artificial intelligence and machine learning aren’t magic, but they can be powerful tools if deployed in appropriate solution areas.

Finally, seek a partner, not just a vendor, when considering AI in manufacturing. To effectively implement AI, it’s essential to understand the risks, build a fail-safe solution, and educate the staff, especially automation engineers. A partner who is genuinely invested in the manufacturer’s success, rather than in just making a sale, can make a significant difference in the overall experience during the adoption of AI in manufacturing.