Machine learning is being utilized in service businesses to run standard, routine, repeatable parts of processes. During the recent OPEX Summer virtual conference, the daily sessions were filled with service companies presenting their approach to using machines to run the core business processes that are executed a dozen to a hundred times a day.
Manufacturing organizations can take a lesson from this approach. As we discussed in our earlier article, by leveraging a mixed-initiative approach and combining the best of Black Belt process expertise and machine learning systems, we can operationalize machine learning in a meaningful way and drive digital transformation into the manufacturing operation.
Tough Going with Edge Cases
Machine algorithms are good at running repeatable processes—those that do not require human judgement to accomplish. However, the experts are still required to handle the “edge” cases, those that are non-standard and require some human intelligence to interpret and resolve. Edge cases in manufacturing involve non-routine things that happen infrequently and, on the surface, do not appear to be repeatable.
Some of these are extremely rare changes such as starting new production lines, qualifying next-generation equipment, replacement of outdated machinery, catastrophic equipment failure, etc. Other edge cases arise more frequently, such as when producing new products—on restoration from failure and maintenance activities—or when new operators are onboarded. In either case, the edge cases require some human intervention to resolve, re-optimize the process and bring it back to a stable state.
Getting machine-learning-based systems to handle edge cases is complex for several reasons:
- Good data is not available about these edge cases due to their infrequent occurrence.
- The knowledge base about how to respond to an edge case typically resides in the head of the experienced expert.
- ·Mapping between actions taken and outcomes achieved is incomplete, limiting any learning-based approach.
Providing enough data to train a machine-learning-based approach requires experts to manually capture all actions used to manage the edge-case event and furthermore link these actions to the outcomes. This is problematic in manufacturing environments, where people are busy. Their value is not usually associated with data-entry tasks, but in units of output produced. Asking a person to manually input responses about an event that they have been busy recovering from is not likely to produce a quality dataset of responses.
In order to overcome these challenges, we require non-intrusive but continuous capture of actions and outcomes associated with an edge case event. There are several intelligent products out there with potential to bridge the gap. These include wearable technologies, as well as passive and intelligent interfaces. Google Glass is an example of the class of intelligent wearables that could be employed to bridge the gap. However, in this case, as opposed to providing real time assistance to the wearer to handle the edge case, we instead use the device to capture data, actions, and outcomes about edge cases. Similarly, we could also use an interactive and passive interface similar to the contact tracing approach adapted by Apple and Google. This has been used to enable a Bluetooth mesh network to trade data about Covid positive interactions without sharing privacy information, and can be repurposed for the factory floor to trace and record data tags when an “edge” case response is in process.
In addition to the non-intrusive capture of data, actions and outcomes, we also need advances in machine learning to be able to leverage this data to train models that can start to handle edge cases. An interesting area of research in machine learning is apprenticeship learning. The idea behind this is that the ML agent behaves like an apprentice—observing the actions taken by the expert, and learning to mimic them to accomplish the appropriate task. These ideas have primarily been explored in robotics, where human experts are used to teach a robot agent how to take certain physical actions.
The underlying learning algorithms use inverse reinforcement learning—where the model needs to estimate the objective an expert is trying to achieve from observing their actions, and then try to optimize it when it tries to accomplish a similar task. Recent applications of this approach have been shown to work in gaming environments (e.g. Atari game play) as well as in real-world settings such as helicopter control and animation. Adapting these approaches to the manufacturing environment would allow the ML agent to learn about actions needed to handle edge cases by observation.
Teaching the Machines
The labor pinch that is the current reality will not abate for the remainder of this decade and into the next decade. Asking workers, of whom there is an ever-dwindling pool, to take time away from recovering from an event as fast as possible to enter data is a losing proposition. As the Great Resignation continues, the pressure on manufacturers will increase, as will turnover and demands for training as people filter through organizations in search of their “ideal” work situation.
As the available workforce dwindles, the machine needs to be able to absorb more and more of the “edge” content into the machine paradigm. Through a wearable monitoring product, passive tracking and inverse reinforcement-based learning approaches, the person can “teach” the machine about edge cases, which the machine can use to expand the understanding of the elements of response to edge cases that are routine, picking out elements that are repeatable even though edge cases don't happen every day.
As we march forward into the future, there will be population shrinkage. It is already happening in many countries. The portion of that future population that is willing to work in manufacturing will be a subset of a subset of a dwindling population, yet our demand for products seems to be increasing. Technology tools need to be assembled in such a way to bridge the gap.
The current state of manufacturing has several challenges to achieve the vision of machine- directed operations, with the digital aide concept at work. The economics of making the technology leap will change as the availability of cheap labor tightens. Many organizations have struggled for years to staff their operations, causing production outages and idle time, which is costly as the investment is underutilized. Additional challenges surround the comfort level of leaders with technology, ability to understand the potential for technology to solve their particular problems and patience as the technology approaches are put together into a seamless integration.
Manual data entry is a non-starter on the journey to enhancing the machine’s ability to learn the edge cases. Active monitoring tools that provide the data without the human having to stop their work on the edge case is the solution to achieve a learning machine. The imperative for the next decade is to set up the machine to learn from humans and absorb more of the “edge cases” by revealing the underlying routines and absorbing those routines in the library of Golden Runs.
Deepak Turaga is senior vice president of data science at Oden Technologies, an industrial IoT company focused on using AI to monitor, optimize and control manufacturing processes. He has a background in academic and industry research specializing in using machine learning based tools to extract insights from streaming and real-time data. He is also an adjunct professor at Columbia University, and teaches a course on this topic every spring.
James Wells is principal consultant at Quality in Practice, a consulting and training practice specializing in continuous improvement programs, and specializes in quality fundamentals, including the application of digital solutions to common manufacturing challenges. He has led quality and continuous improvement organizations for over 20 years at various manufacturing companies. Wells is a certified master Black Belt and certified lean specialist.