Assuring Manufacturing Product Quality through AI


A key difference between a company that is growing and one that is struggling is the real and perceived level of quality of its products. Although the perception of what constitutes “quality” from a customer’s perspective is subject to individual interpretation, manufacturers can establish certain process parameters and final product tolerances. Doing so can ensure the finished product comes as close as possible to the original product design and performance specifications.

AI Can Help Ensure Product Quality

Artificial intelligence (AI) is helping manufacturers achieve better process and product quality by ensuring that equipment is operating properly and that any variabilities in material or processes aren’t introduced into the manufacturing process. Using sensors attached to production machines, signals can be gathered and monitored, tracking both machine operating data and ambient sensor data such as temperature, relative humidity, or other environmental data. Then, using machine learning (ML) algorithms, this data is compared with real-time process data. When conditions affecting quality are detected, they can be corrected prior to the item being physically manufactured.

An example of this type of AI technology comes from an automobile manufacturer that completes more than 50 million welds per day. The automaker wanted to ascertain when conditions were such that a bad weld would likely occur. It then wanted to be able to redo the weld while the body was still in the weld cell, eliminating the cost and waste of completely scrapping the part or needing to put the part back into the queue.

Fifteen signals are collected off each welding robot, including signals measuring torque, current, motor revolutions per minute, arc time, and the thickness of the wire that was being fed. Using software provided by Falkonry, the company is able to detect in real time if the quality of the weld is good or bad by comparing the signals in known good welds to the instant data signals it is getting from the robots in real time. The ML algorithm is also used to detect new anomalies, which are then fed back into the system to refine and improve the algorithm. The end result is a 98% detection rate for bad welds, resulting in a decrease of 5 times the number of welds that must be re-welded. Both the cost and time savings are significant, and this sort of deep analysis could not be conducted manually.

AI Can Help Overcome Human Limitations

AI is also being used to ensure that finished products meet or exceed all quality standards. Typically, visual inspections of the product are done at certain points during production and/or once the product has been completed. Years ago, this was handled via humans who would physically inspect the products, but this process is subject to human limitations.

Humans with significant skill levels and experience still suffer from fatigue, particularly when they are doing the same task over and over again. Humans also may have temporary or slowly degrading health conditions (such as reduced vision due to allergies, illness, or stress, or undetected conditions such as cataracts, macular degeneration, or keratoconus), which may affect the visual inspection process. That is why the use of machine vision (MV) and ML can be useful in improving the efficiency, accuracy, and repeatability of the product inspection process, enabling a two-pronged approach of using both machines and humans.

Landing.AI is a new company founded by Andrew Ng, a noted ML expert who led the development of Google Brain. It is incorporating MV, object identification, and ML to help manufacturers integrate AI into their workflows. For visual inspection, Landing.AI’s system can recognize patterns of imperfections after reviewing only five product images, compared with traditional visual inspection systems that must be trained with massive datasets of around 1 million images to ensure they recognize all potential imperfections. According to Ng, Landing.AI’s deep learning (DL) algorithm takes half a second to inspect a part and, in many cases, is more accurate than humans.

Another example of using ML to reduce the amount of data and time required to train a machine to recognize images comes from Fujitsu Laboratories. When changing a manufacturing line or part, the inspection system must also be revised to ensure the database of reference image data is current. Although the company’s existing image recognition system program’s targets and purposes were set in advance, the amount of image data that was usable prior to starting new manufacturing lines was limited.

Fujitsu realized that while DL could be used to generate an original algorithm, it would require a significant amount of training data to be effective. The company instead turned to specialized genetic programming to speed up the learning process by configuring the image recognition application with dedicated templates. For example, templates can narrow the learning and recognition process to three processes: image enhancement, threshold process, and binary image handling. The program will learn and evolve automatically by preparing training data from images of normal and defective parts to make pass/fail judgments.

When tested on a parts assembly line, the new image recognition system automatically generated code for inspection and achieved a nearly 100% recognition rate. As a result, the time to develop programs for pass/fail inspection processes decreased by 80%, and the company’s parts assembly machines can now relearn as adjustments are made and still maintain recognition rates of 97%-plus. Furthermore, positioning variations were cut in half, reducing actual work hours by 33%.

Desire for Quality Is a Key Driver of AI

While the use of AI is still in a relatively nascent stage, the quality of a product is ultimately a key purchase driver for both consumer and business-to-business customers. Thus, manufacturers have made quality monitoring and control a priority use case. Tractica projects that the use of AI for quality monitoring and control is on track to generate $418.2 million in annual revenue by 2025, with Asia Pacific ($133.8 million) and North America ($117.1 million) driving the market.

Additional details can be found in Tractica’s Artificial Intelligence for Smart Manufacturing Applications report. This report contains a review and revenue forecasts of quality monitoring, yield improvement, root cause analysis, predictive maintenance, energy management, digital twins, 3D printing arm control, and other key smart manufacturing use cases.

Comments are closed.