Boston, MA. The intersection of machine vision, deep learning, and Industry 4.0 was a topic of interest at The Vision Show. Presenters described a machine-vision approach to Industry 4.0 as well as deep-learning techniques and software with applicability to industrial environments.
Phil Jadd of Xyntek described Industry 4.0 as collaborative data-driven process optimization, drawing on the IoT revolution to interconnect what once were standalone devices. In an industrial setting, however, implanting the necessary connectivity can be an expensive process, he said, with one possibility a “rip-and-replace” strategy of installing Industry 4.0-compatible equipment. Such an approach results not only in high equipment costs but downtime plus uncertainties regarding how well the new equipment will work in the short term.
Jadd suggested a machine-vision approach in place of electrically connected systems for extracting metrics such as OEE. AI-enabled machine vision, he said, can not only read human-machine interfaces but monitor many devices at once. To track movement through a warehouse, he said, a 360-degree camera system can view the entire warehouse at once, sending data to an OPC server to monitor uptime and downtime, calculating how many times a particular piece of equipment jams during a shift, for example. In fact, he said, the necessary cameras might already be installed in a facility.
There are challenges, however. Jadd described the gathering of data using AI as a slow process, requiring lots of processing power (GPUs, for example, to implement neural networks) that’s likely not available inline on self-contained vision systems. Security is a concern as well, with some industries more susceptible to hacking than others.
Jadd suggested an approach of layering technologies—rather than rip-and-replace—as a way forward to Industry 4.0. He recommended a closed-loop approach in which data feeds back into the manufacturing process, perhaps using a collaborative robot to turn a dial to optimize a process. “That’s what Industry 4.0 is about—extracting data and making changes to real-world processes,” he said. The ultimate goal: “Become more lean, more efficient, and make more money per unit.”
Jadd described Xyntek as an integrated-services company that has worked in pharmaceutical, life-sciences, environmental-controls, HVAC, and security industries.
Deep learning for industry
Andrew Mayer, a sales engineer at Leoni Engineering Products & Services Inc., described deep learning as applied to industrial applications. A neural network that can identify dogs and cats on the Internet may have thousands of nodes and layers, but, he said, a dozen nodes and two or three layers might suffice for industrial applications.
A challenge is training—the most computationally intensive stage of implementing a deep-learning machine-vision application. For new applications, you may not know what a defect will look like. One that you can produce in a lab might not reflect what you’ll see in production. Further, a deep-learning system may flag as a defect a cosmetic characteristic that doesn’t affect functionality. For training, he emphasized, we are not interested in good parts—the most valuable part is a bad part.
Mayer cited as an example a system examining wood grain to separate oak from maple from pine. Generally, he said, 10 to 100 samples of each class will yield acceptable results. He described deep learning as a tool like any other—good for some applications but not so good for others. Deep learning isn’t so good at telling you that your oak board is 82.1 mm wide—“It doesn’t think about the world that way,” he said. Ultimately, he said, deep learning should be implemented when it can match customer expectations with reality.
Deep-learning software library
Kim Hanjun, a vice president at Sualab, described his company’s SuaKIT deep-learning software library for machine vision. He said Sualab, a South Korean company, was founded in 2013 and has until now focused on the Asian market, with, for example, Samsung having used the company’s technology for mobile-device parts inspection; LG, for automotive parts inspection; and Hanwha, for solar inspection. Now, he said, the company is beginning to address American and European markets.
Deep learning, he said, involves teaching a machine to find features by itself. Its merit is that a machine can find features humans can’t. He traced the development of deep learning from 1943, citing as milestones advances in back propagation and the rectified linear unit (ReLU) activation function.
He commented that whereas a platform like TensorFlow may need 10,000 or 100,000 training images to learn to sort cats and dogs, industry defects—dirt and scratches, for example—are much alike, minimizing the number of images needed for training. While SuaKIT training requirements will differ, in general 100 images per defect type should suffice, with typical training times averaging 20 minutes for 1,000 1,024- by 1,024-pixel images. Inspection-line real-time processing can reach 20 such images per second. SuaKIT requires 64-bit Windows and an NVIDIA GeForce GTX 980 (GeForce GTX 1080Ti recommended) GPU.