Key Lessons When Implementing Industrial AI in Post-COVID-19 World

COVID-19 has caused a massive interruption to factory productions worldwide. The requirement of safe distancing will definitely lead to future reorganization of work cell and process flow in the industrial and manufacturing setting. Manufacturers unable to do safe distancing due to space restrictions will need to resort to more automation. While a lot of the existing automation in factories, such as product inspection, barcode reading, and item tracking, has been well-supported by conventional Artificial Intelligence (AI), Machine Learning (ML)-based AI is expected to bring greater levels of automation, allowing manufacturers to reclaim, or even surpass, their pre-COVID-19 efficiency levels.

This explains why the announcements relating to ML-based industrial AI have never ceased, despite all the negativity surrounding COVID-19. Leading edge AI chipset manufacturers Intel and NVIDIA have unveiled key partnerships that deepen their presence in the industrial and manufacturing space. In April 2020, Intel joined hands with ADLINK and Arrow Electronics to Launch Vizi-AI Development Starter Kit. The development kit features the Movidius Myriad Vision Processing Unit (VPU), Intel’s Application-Specific Integrated Circuit (ASIC), and the OpenVINO toolkit for machine vision applications at the edge, enabling deep learning-based AI automation in manufacturing processes.

Not limited to manufacturing lines, ML-based industrial AI is also deployed in factory logistics. In May, NVIDIA announced a new addition to its EGX lineup targeting edge AI applications and unveiled a partnership with BMW in which BMW developed five logistics and warehousing robots using NVIDIA’s Graphic Processing Units (GPUs) and Isaac simulation technology.

At the same time, manufacturers can also look beyond new equipment and robots. ABI Research forecasts that the total installed base of AI-enabled devices in industrial manufacturing will reach 15.4 million in 2024, with a Compound Annual Growth Rate (CAGR) of 64.8% from 2019 to 2024. As discussed in the ABI Insight Hardware Agnostic AI Deployment to Drive the Next Wave of AI Adoption (IN-5731), the market opportunities for ML-based industrial AI go beyond new GPU- or ASIC-equipped devices. AI models deployed through runtime applications are enabling legacy industrial equipment to run predictive maintenance and ambient monitoring using their microcontrollers. All these are discussed in detail in ABI Research’s Smart Manufacturing Platforms Competitive Assessment (CA-1254) and Industrial AI Platform and Service Provider Competitive Assessment (CA-1269). Since then, the space has seen new entrants, including Neurala, One Tech, Entefy, FAIM, and Laneyes, bringing more options in terms of implementation and subscription models.

Beyond typical ML applications, U.S.-based startup Landing AI is expanding the capabilities of deep learning in industrial and manufacturing by introducing transfer learning and self-supervised learning based on synthetic data generation. The company has found strong demands in leak defect and micro particle detection, surface defect inspection, and geometric measurement.

All these announcements highlight three important lessons for manufacturers in the implementation of ML-based industrial AI:

• Leverage Existing Toolkit: While large manufacturers often have the internal capabilities and resources to attempt to create their own AI projects, it is important to leverage existing chipsets and toolkits to facilitate their Research and Development (R&D) processes. Developed on the NVIDIA Isaac Software Development Kit (SDK), BMW’s robots utilize a number of powerful deep neural networks, addressing perception, segmentation, pose estimation, and human pose estimation challenges on factory floors. This knowledge can be applied to BMW’s other factories and plants using NVIDIA Transfer Learning Toolkit.
• Recognize the Importance of Flexible Architecture: Intel‘s partnership with ADLINK addresses another set of challenges in manufacturing. For applications that are resource-constrained in terms of battery and heat dissipation, Intel’s Movidius Myriad VPU brings dedicated hardware accelerator for deep neural network and computer vision. Developers are able to use OpenVINO to develop their machine vision applications and apply them across all of Intel’s products, including Central Processing Units (CPU) and Field-Programmable Gate Array (FPGA). The ability to leverage CPU, FPGA, and ASIC for the same task, albeit with different levels of processing and energy efficiency, means factory operators can save their development effort and extend their asset lifetime for a little longer by making them a little smarter.
• Embrace External Expertise: Often, AI startups lament having to compete with the agenda of these manufacturers’ internal Information Technology (IT) teams. Ultimately, the end goal of both parties is the same—namely, to increase productivity and efficiency—but each has its own means. In certain low-demand cases, getting off-the-shelf AI solutions from machine vision hardware solutions like Cognex and Basler or cloud AI vendors such as Alibaba, Baidu, and Amazon Web Services (AWS) works perfectly, but on other occasions, it is important to rely on specialized AI solution providers that have deep knowledge in ML, such as Landing AI, Smartia, and Instrumental. A healthy and successful partnership generally involves external experts with cross-vertical knowledge and an internal engineering team with deep domain expertise.