AMERICAN INDUSTRIAL PARTNERS Additive Manufacturing Opportunity Identification

FORD MOTOR COMPANY

Using AI to Automate and Scale Digital Twin Technology for Advanced Manufacturing

Student Team: Michael Fanelli – EGL (BSE & MSE in Mechanical Engineering) Matt Rosales – Master of Business Administration

Project Sponsors: Mark Goderis – Technology Manager, Manufacturing Technology Development Annie Zeng – Digital Twin and AI Specialist, Manufacturing Technology Development

Faculty Advisors: Vijay Pandiarajan – Ross School of Business F. Andy Seidl – College of Engineering

For over 100 years, Ford Motor Company has developed, manufactured, and distributed vehicles worldwide. Today, Ford employs 186,000 employees globally, with revenues expected to exceed $140B in 2021. To maintain its position as a world-class manufacturer, Ford is investing in advanced manufacturing technologies to drive its “Factory of Tomorrow” transformation towards an intelligent and completely connected manufacturing system. One such disruptive technology is a Factory Digital Twin, a digital representation of plant floor assets and processes. Digital Twins allow for virtual designs and validation, increased plant floor efficiency, and integrated, real-time monitoring. With each asset on the plant floor having maintenance, quality, and live Industrial Internet of Things (IIOT) data, a Digital Twin relies on connected data. Often, these systems were developed independently of each other and operate in silos, making them difficult to integrate. Ford’s Master Asset Registry Service (MARS) decodes and connects disparate systems, but the process was highly manual, constraining the Digital Twin’s pace of implementation. The Tauber team’s goal was to automate this systems integration to drive efficiency and accelerate the implementation timeline. The team pursued a two-pronged solution: Retroactive and Proactive. The Retroactive Solution integrates existing assets in plants while the Proactive Solution designs out the problem by making new assets digital and connected from the moment they are born. To develop the Retroactive Solution, the team interviewed architects and end-users of various manufacturing systems, mapping the ontology. This was used to develop a multi-layered AI solution architecture, using an Expert System and Machine Learning, designed for Ford’s diverse manufacturing systems and plants. The team demonstrated this architecture’s viability and scalability by developing a proof-of-concept focused on connecting a machine acceptance system and a maintenance system in Ford’s Powertrain plants. The team also developed a roadmap for the Proactive Solution. This roadmap detailed strategies for creating connected data when assets are born, designing simpler strategies when possible and more nuanced ones when necessary. The team demonstrated this roadmap on custom tooling used in new plants, focusing on the Rouge Electric Vehicle Center, the new plant where the all-electric F-150 Lightning will be manufactured. The Tauber team’s AI solution decreases the labor and time to implement the Digital Twin on new manufacturing lines by 65%, allowing for an accelerated pace of implementation 2.86 times greater than before. This will lead to increased efficiency on the plant floor across Ford’s Powertrain plants, directly saving the company $3.64 million from 2022 to 2024, or $1.21 million annually. Additional savings are expected from extending the solution architecture to other manufacturing systems and plants as well as from a reduction in downtime.

GENERAL MOTORS

Deep Learning For Machine Vision

Student Team: Mabel Chan – EGL (BSE & MSE in Computer Science Engineering) Konstantinos Chiotinis – Master of Business Administration Soo Yeon (Sean) Lee – EGL (BSE & MSE in Computer Science Engineering)

Project Sponsors: Amar Amad – Engineering Manager William Keller – Sr. Manufacturing Project Engineer

Faculty Advisors: Jeff Alden – College of Engineering Sanjeev Kumar – Ross School of Business

General Motors Company, a $79B company, is one of the world’s largest automotive manufacturers, producing close to 7 million vehicles a year. The company uses more than 2,000 machine vision cameras across all facilities to inspect various aspects of production. So far, the accuracy of several vision cameras has not been satisfactory to General Motors, since inaccuracy results in unnecessary downtime. General Motors wants to decrease downtime and increase throughput. A potential solution to the problem of camera inaccuracy is the application of deep learning (a form of artificial intelligence learning based on artificial neural networks in which multiple layers of processing are used to extract progressively higher level features from data) on machine vision cameras.

Asked to analyze the performance of deep learning versus traditional methods for machine vision applications within GM, the team implemented deep learning on 13 machine vision cameras that were either performing poorly with traditional vision methods or were performing especially complex inspections. Using deep learning, the team was able to improve camera performance, achieving an accuracy of 99.7% or higher, a number set as the target by GM, an improvement of more than 15 percentage points for some cameras. With deep learning, every camera performed as well as or better than traditional vision methods, indicating an opportunity for large-scale implementation.

Three process improvements were also made, with a fourth proposed possibility: (1) a decision-making process to determine whether a camera is a suitable deep learning candidate; (2) a protocol for training deep learning models, which decreases implementation time by 80%; and (3) a method to reuse deep learning models on different cameras. Fourth, the team identified several bottlenecks, which, if addressed, can decrease deep learning implementation time by up to 50% (more than 3 hours per camera) and the needed space for storing historical machine vision data by 66%, saving GM several terabytes of storage space.

The performance improvements made by the team resulted in over 30 hours of decreased projected downtime per month. The potential economic benefits for implementing deep learning across the assembly line are estimated to be roughly $9 million for the first year and $40 million within 3 years. Also, the increase in machine vision accuracy will positively affect quality control.

HERMAN MILLER, INC.

Improving Domestic Outbound Freight Utilization

Student Team: Michael Krasnor – Dual MBA & MSE in Industrial and Operations Engineering Felice Mueller – Master of Business Administration Emmett Springer – EGL (BSE Biomedical Engineering & MSE Industrial and Operations Engineering)

Project Sponsors: Jeff Kurburski – Chief Technology Officer Richard Scott – Chief Global Manufacturing and Operations Officer

Faculty Advisors: Robert Inman – College of Engineering Jim Price – Ross School of Business

Herman Miller, Inc., a $2.5B company, is a globally recognized leader in the design and manufacture of furniture, and the parent company to many notable furniture and home goods brands. The firm operates manufacturing and distribution sites in the US, UK and China. The Midwest Distribution Center (MWDC) in Holland, MI aggregates and distributes goods from all Herman Miller Group brands to all customers across the Americas. The MWDC sought to improve the utilization of truck-driven trailers, their primary mode of domestic product shipment, to reduce shipping costs.

Herman Miller’s business model complicated this endeavor; their value proposition is founded on a high level of customer service and reduced costs derived from lean manufacturing principles. Herman Miller’s logistics coordinators operated within the Customer Care department to ensure that customer requests - rapid order modification, exact times and days of delivery, maximum trailer quotas, and physical loading requirements – were met. Just-in-time manufacturing with no safety net of inventory also drastically constrained shipping timelines. These factors hampered logistics coordinators’ ability to efficiently coordinate trailer loads, directly decreasing profitability, as the firm offered free domestic shipping. Coordinators relied on tribal knowledge and timeconsuming manual processes to meet customer demands.

The Tauber team analyzed shipping patterns and identified frequently visited warehouses as prime targets for improving shipping efficiency. These customers which used these warehouses had relatively smooth demand profiles, predictable delivery preferences and an ability to unload densely packed trailers. Further analysis determined that the presence of such “high-traffic warehouses” could enable efficiency improvements for deliveries to nearby customers as well. The Tauber team programmed an automated tool to codify and improve upon coordinators’ best practices and increase shipment efficiency to these clusters. The team also implemented several communication flow improvements, data quality control procedures, and statistical risk quantification and management processes to support these operational improvements.

The team successfully began implementing these tools and operational improvements. Herman Miller’s expected annual savings are $1.6M (4% of total shipping costs) with a 90% reduction in service level related risk and a 484 US tons (3.4%) reduction in CO2 emissions.