The best manufacturers in the world are using machine learning to automate, improve, and evolve their factory lines. Machine learning can reduce emissions, help monitor equipment and flag anomalies, and automate manual work: all without the need for teams of hundreds. This post will explain how to get there.
Use Cases for Machine Learning in Manufacturing
Data science plays an important role in every step of the supply chain, especially manufacturing. Factories and manufacturing are holistic operations: they’re a business, not just part of a business. That’s why machine learning has some of the most varied and effective use cases here of any vertical or industry.
It’s becoming increasingly important to monitor, decrease, and ultimately eliminate the negative externalities that your manufacturing imposes on the environment. Siemens was able to use deep learning to reduce the emissions of their gas turbines by an additional 10% on top of what human experts came up with.
Preventing Equipment Failure
Downtime is a routine and expected part of a manufacturing operation, but cutting it by a meaningful percentage can have significant impacts on output. Machine learning for manufacturing includes predictive maintenance to determine the condition of equipment so that maintenance can be performed proactively, saving time and money wasted with downtime. GE’s Predix platform is the market leader in helping companies monitor and prevent failure, and we’ll cover it more below.
Industrial robots are generally considered low hanging fruit in factory automation because of their practical bottom line impact. Fanuc, one of the Japanese leaders in industrial automation, is teaching assembly line robots to learn themselves with deep reinforcement learning.
Improving Yields and Quality
In certain complicated manufacturing processes like that of semiconductors, initial applications of Machine Learning have been able increase yields by up to 30% and reduce scrap rates significantly. ML models can carry out root cause analysis as well as reduce testing costs (one of the most frustrating parts of semiconductor fabrication, with entire companies dedicated to improving it).
In the earlier stages of company formation, it’s tremendously difficult to keep supply at pace with demand. Honeywell has been integrating machine learning into their factories to forecast product demand and adjust inventory and manufacturing accordingly.
Leadership: What is Being Tested with Machine Learning in Manufacturing?
Both for their own manufacturing and as a software producer, GE has been a leader in the factory automation and machine learning space. Their core product is called Predix, and it’s a platform for creating machine learning based analytics tailored to factory assets and based on the digital twin model.
The digital twin is a simple but powerful concept: manufacturers can create digital representations of their physical factory equipment that store all of the relevant data. Think of this as a sort of database schema: if you’ve got hundreds of thousands of sensors monitoring data in your factory, these assets give you a way of tagging, nesting, and organizing them. Digital twin schemas can get much more complex than this, but the main idea holds.
Predix helps companies build systems across multiple use cases like process automation and failure prediction. According to their website, the platform supports 300+ models and 150+ asset models for digital twins. While Predix and GE Digital may not have been blockbuster successes financially, they clearly show the promise and potential that machine learning can bring to this critical function.
Alongside the established companies in the space, startups have been tackling these problems and getting some impressive traction. Alluvium helps factory owners make sense of their data streams with machine learning, and has raised money from some of the best investors on the East Coast. Landing.AI, a higher profile company, was founded by AI pioneer Andrew Ng and is approaching factory automation as its first vertical.
Domain Problems: Challenges with Machine Learning in Manufacturing
With such a high number of relevant devices and the increasing complexity of getting them to all work together, it can be seriously challenging to successfully implement machine learning in manufacturing.
Sensors and Network
Companies with significant manufacturing operations can have hundreds of thousands of sensors in their factories, if not more. This is more than just big data we’re dealing with: it’s layers of complexity around networking, organization, and latency. Algorithms will need to deal with nuances of variety, volume, and velocity.
Impact at Scale
Like we noted about customer experience or patient outcomes, certain domains magnify the impact of your predictions because of the end product or experience. If you’re manufacturing at scale, messing with your process can be destructive if not done properly. Practitioners and modelers need to make sure that machine learning integrates smoothly, and that the impact you allow predictions to have builds up slowly over time.
For products that have a lot of intermediate steps in their manufacturing process, there are a higher number of failure points, and changes propagate in increasingly unpredictable ways. When creating a machine learning system where predictions change the way you operate, these changes can have obfuscated downstream effects on the products you’re manufacturing. Extensive testing and domain understanding are prerequisites to building these models.
Model Tradeoffs in the Manufacturing Vertical
Unlike some other verticals in machine learning, manufacturing has a wide variation of data inputs and thus a varied spectrum of models that can be applied. All different types of ML models—supervised learning, unsupervised learning, and reinforcement learning—have been applied with varying degrees of success.
General vs. Specific
For practitioners in the space, an important tradeoff is model applicability and specificity. Models that rely on specific types of data will rule out parts of the manufacturing process (e.g. a model that requires certain data points from a type of machine won’t work on other types of machines), but can be more accurate in the long run. More basic approaches like anomaly detection are more general, but may sacrifice the nuances that more specific models will pick up on.
One of the unique parts of industrial processes is that data and devices are very much on the edge and not in the cloud. A typical workflow for training might involve pulling data from devices and then training in the cloud, followed by uploading your models to edge devices for inference. But running models on smaller or less powerful devices is challenging, and opens up a whole new set of constraints for data scientists.
Manufacturing as a vertical has its own set of processes and people: it’s not uncommon for technicians to be working in factory ops for 20-30 years. In industries with rich traditions like this (see: healthcare), model explainability is key. In the words of a Forbes article on the topic, “Technicians who have been in the field for 45 years will not trust machines that cannot explain their predictions.” The ubiquitous tradeoff between simple explainable models and complex opaque models (deep learning mostly) is crucial to navigate and gain initial traction.
Machine Learning in Manufacturing – Present and Future Use-Cases (techemergence) – “Major companies including GE, Siemens, Intel, Funac, Kuka, Bosch, NVIDIA and Microsoft are all making significant investments in machine learning-powered approaches to improve all aspects of manufacturing. The technology is being used to bring down labor costs, reduce product defects, shorten unplanned downtimes, improve transition times, and increase production speed.”
10 Ways Machine Learning Is Revolutionizing Manufacturing In 2018 (Forbes) – “Machine learning algorithms, applications, and platforms are helping manufacturers find new business models, fine-tune product quality, and optimize manufacturing operations to the shop floor level.”
Machine learning in manufacturing: advantages, challenges, and applications (Paper) – “The nature of manufacturing systems faces ever more complex, dynamic and at times even chaotic behaviors. In order to being able to satisfy the demand for high-quality products in an efficient manner, it is essential to utilize all means available. One area, which saw fast pace developments in terms of not only promising results but also usability, is machine learning.”
4 Unique Challenges Of Industrial Artificial Intelligence (Forbes) – “As sexy and shiny as robots are, the bulk of the value of AI in industrials lies in transforming data from sensors and routine hardware into intelligent predictions for better and faster decision-making. 15 billion machines are currently connected to the Internet. By 2020, Cisco predicts the number will surpass 50 billion. Connecting machines together into intelligent automated systems in the cloud is the next major step in the evolution of manufacturing and industry.”