An alternative to machine learning. By Steve Roemerman
The emergence of the fourth industrial revolution concept a few years ago marked a major turning point for the manufacturing industry. The goal of every industrial revolution has been to improve productivity and tap the capabilities of new technology. However, the current path has the widest range of possibilities. Today’s focus is on integration and automation. When implemented correctly, these transformative processes can improve productivity, efficiency, and improve a company’s bottom line.
Companies have attempted to deploy AI technology, such as machine learning, to solve some age-old issues with manufacturing. These practices are based on the idea machines should both learn and act without human interference. In some cases, machine learning has allowed companies to reduce bottlenecks in production, as well as downtime, and labor costs. While promising, such successes are offset by the limitations of machine learning in a manufacturing environment.
The heart of the problem is that traditional AI requires failures in order to learn. But manufacturing companies simply can’t afford or allow failures in their production. As such, machine learning cannot take place because there is not enough data for the machines to learn. However, there is an alternative to machine learning which can solve this dilemma. This alternative is Evolved Artificial Intelligence.
Developed by Lone Star Analysis, Evolved AI (EAI) is a hybrid, knowledge-driven approach. It gets its hybrid status for its inclusion of multiple modeling techniques, statistical methods and knowledge of the domain from subject matter expertise. This combination approach allows EAI to start with the known boundaries of a problem, instead of learning everything from scratch. This is one of the differences between EAI and the machine learning method. For example, if a formula incorporates multiple variables as a given, or widely accepted knowledge, EAI teaches the model this concept from the start. Other models might rely on a training set of those same variables to reach the conclusion of the already understood knowledge. In this way, EAI is naturally smarter.
How does it work?
Using adaptive discovery, EAI finds the most ideal array of conditions contributing to a scientific equation matching real-world observations. The conditions introduced are tailored to the problem by including known interactions between variables and any existing relationships, especially those less obvious at first glance. Conditions are then either emphasized or given less value based on their joint ability to produce an expected outcome.
Machine learning models typically use deterministic approaches. EAI, however, is more resilient when presented with new data, whether it’s discrete, continuous, deterministic, stochastic or a combination. Mathematically, this enables it to solve more complex problems. Practically, it means EAI is less greedy for data and processing power. This opens a whole world of challenging problems other approaches just can’t solve.
EAI was developed to address concerns about machine learning, such as its need for and reliance on large data sets, its level of accuracy and its transparency. Combining previously experienced problems with mathematical approaches like signal processing, time series forecasting, and stochastic optimization resulted in an interpretable and explainable model.
Because of mathematical and physics-based transparency, practitioners using EAI can easily interpret results and explain how the system works.
The terms within the mathematical model are explicit. This is in direct opposition to traditional neural networks which introduce unexpected variable relationships with hidden architecture which is opaque to the user at best.
Benefitting from EAI
In addition to being more understandable, EAI’s benefits range from computing power to data set sizes. Because it can run on a laptop PC, there’s no need for exorbitant amounts of computing power from graphics processing units, or extensive cloud resources. EAI can operate anywhere: edge, cloud or offline.
EAI isn’t only dependent on historical training data. This means solutions can be generated when no historical data has been collected, data is limited, or data is no longer relevant. This removes the need for big data and training data. That means there’s also no need for cleaning, scaling or tagging. Customers can instantly gain value instead of waiting around for data to become available.
Lastly, the designer of the model has control over what the system learns. This provides a built-in clarification function. A mathematical model describes the input/output relationship. This allows models to be tuned to specific equipment characteristics for applications in the industrial internet of things (IIoT).
Application in Manufacturing
Although the possibilities with EAI are endless, its biggest benefits currently to manufacturing are condition-based maintenance and operation monitoring. In both instances, EAI learns what the manufacturing process is supposed to do, and immediately predicts critical conditions and prescribes mitigating action. For condition-based maintenance, this can focus on critical pieces of infrastructure like pumps, motors and conveyors to ensure a reduction in downtime. Meanwhile, operational monitoring solutions take a more macro approach by focusing on things like inventory levels, cycle times and resourcing to forecast disruptions in production.
Long term, EAI will change how manufacturing performance is monitored. Machine-specific data can help uniquely characterize each item. This is a more detailed approach, as opposed to using generic data and static thresholds to characterize a broad cross-section of industrial equipment. Characterization will maintain relevance in often changing environments. Subsequently, machines may be checked for deviations from their normal mode of operation. Therefore, prescriptive maintenance may take place on an individual ‘need’ basis, rather than comprehensively through a ‘calendar-based’ plan. Using this approach reduces chances of performing maintenance when it’s not needed.
Evolved AI is a smarter subsect of artificial intelligence. EAI can ultimately help companies realize their technological goals by resolving the very real limitations of machine learning. This tested and proven hybrid approach is unconstrained by mathematical limitations, more accurate and more transparent. With its ability to assess data sets of any size, while also factoring in the potential for random occurrences, EAI is primed to be essential for manufacturing companies looking to truly create smarter operations.
Steve Roemerman is the chairman and CEO of Lone Star Analysis. Lone Star Analysis delivers predictive and prescriptive analytics and evolved artificial intelligence to help customers make informed decisions in the face of uncertainty. Since 2004, companies have trusted Lone Star to deliver actionable answers to complex problems in manufacturing, aerospace, defense, energy, logistics, transportation and more.