« Back To Artificial Intelligence Industry Insights
Association for Advancing Automation Logo

Member Since 1974

LEARN MORE

Content Filed Under:

Industry:
N/A

Application:
N/A

Industrial AI Applications That Can Help Your Business Today

POSTED 07/24/2024  | By: Nick Cravotta, A3 Contributing Editor

Are you a manufacturer that’s ready to begin an AI journey, but you aren’t sure how to get started?

Artificial intelligence can be successfully integrated into manufacturing with low risk — so you can evaluate how to best take advantage of the capabilities AI has to offer. This includes detecting and eliminating defects, improving safety, bringing more value to your skilled labor, and streamlining supply chain logistics.

This article explores applications that can help your business today — and get you started on deploying AI in your facilities.

Enabling People to Do More

Versace Massimilliano is co-founder and CEO of Neurala, a company that automates quality inspection through AI-based vision software. According to Massimilliano, “By endowing a machine with intelligence, you can apply human capabilities where you can’t safely or cost-effectively put a person.”

Ultimately, AI is designed to help people on the production line. Massimilliano cites the example of a machine that produces 10,000 parts a day. With human-based inspection, a person can sample just a small fraction of overall production. If this person is also managing the machine, there will be little time left for quality control.

With the assistance of AI, a person’s capacity and accuracy can be vastly improved. Instead of inspecting parts all day, the worker can manage the machines that inspect each part. This approach eliminates the repetitive, dehumanizing aspects of production and enables people to focus on work that is more engaging.

Amplifying human abilities is key to addressing the labor challenges facing manufacturing today. For example, in December 2023, 40% of available job openings for durable goods manufacturing were unfilled according to the U.S. Chamber of Commerce.

In many cases, AI is not intended to eliminate workers, but rather fill jobs people don’t want to do or tasks that are not safe or feasible for people. In addition, AI significantly boosts the output and quality of individuals. More can be done with the same number of people, enabling manufacturers to increase margins while improving quality, reducing defects, improving safety, and increasing pay for skilled labor.

AI at the Edge

One concern for many manufacturers is how AI will impact their existing infrastructure. Understandably, they are hesitant to scrap an operational production line to rebuild an AI-based system. AI, however, can be introduced to existing lines without massive redesign or investment.

Many production floors deploy older equipment that is either not connected to the network or doesn’t have integrated sensors for AI analysis. Rather than attempt an expensive and complex upgrade, these systems can be designed to take advantage of AI-based capabilities through the use of external sensors, such as:

  • Cameras (image- or infrared-based)
  • Acoustic monitors
  • Environmental sensors (temperature, pressure, humidity, chemicals, etc.)

Each sensor is an “edge device,” effectively a small computing system that can operate on its own. Using edge devices, existing production systems can take advantage of AI without expensive upgrading or retrofitting.

For example, an edge camera might have an integrated graphics processor unit (GPU) to run an AI neural network on the camera itself. This neural network would identify potential issues at a particular point on a production line and then escalate more detailed processing up to a hub. A hub might be as simple as a Raspberry PI board that monitors 10 or more sensors. If an issue is detected by the hub, a more powerful server on the factory floor could make a final decision on what action to take, including escalating the issue to a person.

Traditionally, AI has been implemented by collecting data and processing it on large servers either on site or in the cloud. However, this approach often introduces too much latency to be effective for a production line; by the time a decision has been made, the part being analyzed has already moved on.

Seth Clark is head of AI Product Innovation at Modzy. Modzy is dedicated to enabling deployment and scaling of AI at the edge. “By running AI models at the edge where they are needed,” Clark says, “we effectively bring machine learning to the data. Now it is possible to provide real-time decision-making without compromising product quality or worker safety.”

Consider an AI-based camera. AI software can be coupled to any camera to create an intelligent edge sensor that can be put anywhere on the production line. Setup and deployment of intelligent sensors often requires only minimal expertise or training. In other words, while the technology underlying the AI is complex, the implementation does not have to be.

For example, the introduction of AI to a production line may start with a proof-of-concept camera system that needs only a small number of images of good product to train the AI models. As defects are detected through traditional manual quality control methods, images of these defects are used to retrain the AI model (see Figure 1). When the manufacturer feels confident in the AI system, manual quality control can be stopped.

Figure 1: Proof-of-concept AI camera systems can operate in parallel with manual quality control until manufacturers feel confident the AI system is as or more effective than manual inspection.

Detecting Defects

A single person is limited in how many products they can manually check for defects. For many applications, AI makes it possible to achieve 100% quality control, not just across a small sample.

Eigen Innovations implements defect detection and prevention technology using a computer-aided design (CAD) approach. Each part is transformed into a standardized model with unwanted variations removed (i.e., position, rotation, color, lighting, etc.). For example, captured images are aligned to the CAD model to eliminate all variations in placement (see Figure 2).

Josh Pickard, senior director product and innovation at Eigen Innovations, explains how CAD-based analysis of data greatly simplifies AI-based inspection. As the AI model is not burdened with learning point of view variations such as angle and placement, AI models can be trained with a smaller data set. AI models also run faster because they are less complex, and they don’t require expensive GPUs to run on edge devices.

“The CAD model provides a ‘ground truth’ for what you desire in a part. It allows comparison of a part back to its original CAD definition,” Pickard says.

This approach has several significant benefits. Consider that parts being evaluated on two different lines will have cameras in different places, creating variations in size and lighting (just to name a few factors) that traditionally make it difficult to scale AI models from one line to others. Transforming images to a standardized CAD model effectively removes these variations. This means that different lines can improve together rather than run as independent systems that must be maintained separately.

Figure 2: Transforming images to a standardized CAD models effectively removes environmental variations such as size, lighting, and placement, making it easier to scale AI models across different production lines.

A key advantage of AI systems is that they are always learning. For example, an AI system may identify a defect where one isn’t (i.e., a “false positive”). When a false positive occurs, the AI system can mark the image and show what data in the model it used to incorrectly identify the defect. A correction can be made to the data and worked back into the AI model to eliminate such false positives in the future, further improving accuracy.

Defect Elimination

An important aspect of AI is that, not only can it detect defects, it can prevent them from occurring in the first place. Consider how Eigen Innovations worked with a tier one automotive manufacturer to improve the quality of their plastic welding. Thermal cameras capture data on the individual plastic welds. This data is consolidated with the digital part records for the welds, which contains everything known and collected during the manufacturing and inspection of this particular part.

 “We integrate AI analysis with process data as well. Process data is any data we can get out of a machine or connected device that gives insight into the manufacturing process,” Pickard says.

 What this means is that when the AI detects a defect, the defect is analyzed in the context of all the available process data to see if the cause of the defect can be identified.

With this data, the auto manufacturer discovered an important variable: heater hold time correlated to overall quality. Several factors – including whether the heater required maintenance, variations in materials, and environmental conditions – materially impacted weld quality.

With these factors, the manufacturer could eliminate many manual checks they used to perform on every part. In addition, they could also reduce destructive testing which was performed on a sample of parts to confirm quality. Put another way, the manufacturer moved from being 100% dependent upon destructive testing to a completely inline AI-based monitoring approach. Now they could predict quality without the costs associated with destructive testing.

With Eigen vision systems running on 19 welding machines across 5 facilities, the manufacturer is generating comparable data and leveraging the same or similar models on each machine. They can use this data to glean insights and drive improvements across these operations.

There are a surprising number of ways that reducing defects can increase return on investment (ROI). Many are called “intangible” because they can be difficult to measure.

Neurala’s Massimilliano offers the following example: “We worked with a shampoo manufacturer to reduce defects in labeling. Even though there was nothing wrong with the product, consumers perceive product with a misprinted label as defective. These defects became returns, which incurred multiple costs that ate into profits, including shipping the product back.”

A manufacturer who ships many defects gains a reputation for high returns, which costs the store as well. Reducing defects reduced returns, directly impacting the bottom line. Reducing returns also improved the manufacturer’s reputation with its distributors.

Another example is apetito, a company that produces 1M+ complete meals per week. Weight-based inspection could only flag an incomplete tray without understanding what was missing. With Neurala’s help, apetito implemented vision AI to detect cases of the five most reported missing components. By inspecting multiple regions of interest, the system can discover which components are missing and identify trends/causes to reduce them in the future.

AI model updates can be completed in 10 to 20 minutes. Now the system can detect these missing components 100% of the time. According to this Neurala and apetito project, in terms of labor savings, one “brain” alone saved £15,000 (approx. $19,357.72) in labor each year (the final system employs 30 brains).

Worker Safety

An important use case for AI is to improve overall working conditions, safety, and environmental quality of the workplace in a practical manner. AI makes it possible to scale health and safety improvements across the entire factory floor. Rather than stationing a person with a clipboard at every critical location, AI enables sensors to collect data and evaluate safety conditions in a much more cost-effective manner without compromising quality or reliability.

Modzy, for example, uses a combination of cameras and environmental sensors to identify injuries and accidents as they occur. Then, by analyzing the data collected, the system can begin to recognize what patterns or conditions are at higher risk for resulting in an injury. “These models can even implement predictive capabilities, such as identifying a trend that tends to result in an unsafe event within an hour,” Clark says. For example, Figure 3 shows how AI can confirm workers are properly protected.

Figure 3: AI can improve safety in many ways, including confirming workers are properly protected.

Measuring the benefit of improving safety, whether through AI-based systems or not, can be difficult. The people working on production lines tend to be skilled labor, which can make it challenging to replace them when an injury occurs. Thus, the avoidance of downtime due to lack of personnel should be considered in any ROI calculations.

An important factor to consider when deploying cameras is how workers will perceive their use. When productivity is the primary focus, workers often worry that cameras will be used to watch and evaluate them. However, when worker health and safety is in the forefront, it is much easier to work with organized labor to establish that cameras are a real benefit.

Streamlining Supply Chain Logistics

AI is making significant improvements in supply chain logistics. Complex systems require management of many parts and subsystems. Using indoor GPS technology to track both parts and equipment, for example, can help ensure resources are where they need to be to not create bottlenecks in the production process.

Consider the use of limited resources such as equipment racks, forklifts, and other tools used to move parts around. Through the use of what Modzy’s Clark calls “situational intelligence”, AI can track inventory, predict resource location and availability, and take action before an issue arises.

For example, equipment racks are used to hold parts and materials so they can be moved throughout the factory. However, if the night crew has used all the racks to hold raw steel, this may lead to a shortage of various parts in the assembly line. With direction from an AI, the night crew could more evenly distribute the racks so as not to create an unintended bottleneck.

AI is also changing how we interact with enterprise resource planning (ERP) systems. Through AI, these systems are becoming capable of understanding conversational language, known as large language models (LLM) or natural language processing (NLP). First, this allows users to interact with complex management systems in a more natural, conversational manner. Rather than having to navigate complex command menus to get something done, users are more and more being able to simply ask for what they want. In addition, AI makes it possible to synchronize ERPs across different locations. This enables OEMs to more easily scale improvements made at one location to other production lines.

Deploying AI

Modzy’s Clark suggests being realistic about how quickly you’ll be able to see results from introducing AI to a line.

“First, integrating and then verifying integration of edge sensors with legacy equipment and software can take a long time. Second, IT and OT tend to have a low-risk mindset and it can take time to gain their buy in if they perceive AI could have a negative outcome. Third, AI can be complex and it takes talent to implement. The tools used to implement AI are different than other types of management tools. There are only a limited number of AI experts, so it may take time to identify and employ them.”

Artificial intelligence brings many benefits to manufacturing by enabling skilled labor to bring more value, detecting and eliminating defects, improving safety, and streamlining logistics. The question isn’t whether AI will change manufacturing but rather how fast. And those manufacturers who don’t begin to embrace AI will find over time that they have more defects and lower margins than their competitors.