Big Data
Article

Go Beyond Machine Learning to Optimize Manufacturing Operations

by
Andrew Silberfarb, SRI International
March 31, 2021
Download PDF

Summary

SRI International has developed a system called Deep Adaptive Semantic Logic (DASL) that uses adaptive semantic reasoning to fill in the data gaps.

Reduce costs and downtime by predicting errors using existing data with DASL

Machine learning depends on vast amounts of data to make inferences. However, sometimes the amount of data needed by machine-learning algorithms is simply not available. SRI International has developed a system called Deep Adaptive Semantic Logic (DASL) that uses adaptive semantic reasoning to fill in the data gaps. DASL integrates bottom-up data-driven modeling with top-down theoretical reasoning in a symbiotic union of innovative machine learning and knowledge guided inference. The system brings experts and data together to make better, more informed decisions.

What’s so dazzling about DASL?

Deep Adaptive Semantic Logic (DASL) goes one step beyond machine learning to optimize manufacturing. DASL draws upon human knowledge and limited example data to detect rare events and convey what it finds in a way that is understandable to human operators and maintainers. Operators can then use these intelligent statements to make better decisions when maintaining and improving manufacturing systems and processes.

But why is DASL so unique? Unlike standard machine learning systems that solely make inferences from structured, carefully curated data, DASL is a next-generation intelligent system that learns from unstructured, real-world data. It combines this data with human knowledge to build contextual models and to reason from these models to explain decisions. Andrew Silberfarb, Senior Computer Scientist at SRI, describes DASL as being “at the crossroads between statistical learning—such as standard deep learning and machine learning—and traditional expert systems that are based on human-intelligible rules.”

How DASL works to focus on a task

DASL is trained using heuristic rules, which are then compiled into the DASL system. DASL then fills in the gaps by building and training low-level neural networks. For example, DASL can be trained to recognize the structure of an airplane. Training would typically involve DASL learning to recognize the wings, fuselage and so on, then combining these individual detections through heuristic rules. Training options are varied and can include any mixture of geometric clues, logical entailment, direct annotations, and distant supervision. DASL data needs are minimal and focused as it is able to share information learned in one context to improve overall performance by leveraging structure within the rules.

DASL is well-suited to identify problems within a manufacturing pipeline and to predict maintenance needs. The relevant and understandable insights generated by DASL can help prevent significant disruptions, such as a manufacturing process shutdown. DASL sends predicted performance and information status to expert personnel, allowing users to better allocate their time and focus on the most relevant issues to avoid disruption. DASL has been designed to create human-understandable output, such as, “A failure is predicted to occur in … system after approximately 30 minutes because:…” This output can be in the form of alerts, textual reports, or structured reports. DASL collaborates with the technicians and engineers to quickly identify problems and coordinate the most efficient response. Ultimately, DASL can significantly save maintenance time and resources, while increasing up-time.

Why use DASL and not just machine learning?

Pure machine learning (ML) models have certain limitations:

• ML requires a lot of data to train the underlying algorithm, but often critical events are rare and produce limited amounts of data.
• ML is not always suited to the nuances of business problems, being a “black box” technique that can’t be guided or understood.
• ML experts, which may not be available, are needed to update the ML model as conditions change.

DASL is explicitly designed to deal with situations where the amount of data needed for ML is simply not available. In these cases, DASL combines human knowledge and example data to make decisions and generate human-understandable reports and alerts. DASL’s lower data requirements and understandability allows a broader range of applications than other ML systems. DASL also allows manufacturers to leverage existing maintenance processes and methods that are proven to work, adding machine learning in a way that enhances, rather than replaces, current practice. DASL represents an evolutionary step from machine learning and rule based systems, merging the best of both.

Sometimes, big data is not available for every possible failure mode but manufacturing still needs intelligent interactions to optimize conditions and processes. DASL provides the ability to leverage knowledge across multiple companies, assembly lines and domains. The reduced need for large data sets means that even when data is rare, DASL shows a significant average performance boost. In the competitive world of manufacturing, DASL can lead to major cost-savings and provide better planning of resources.

Read more for a deep dive on DASL technology.

Andrew Silberfarb, SRI International
Andrew Silberfarb, SRI International

Andrew Silberfarb has a Ph.D. in Physics from the University of New Mexico, and a B.S. in Physics from CalTech. He is an expert in Machine Learning and Data Fusion, having spent 7 years working at MIT Lincoln Laboratory to support government R&D. For the past 5 years he has been a working at SRI International on developing artificial intelligence and machine learning techniques for application to challenging problems in predictive analytics, medical diagnosis, and image/video detection among others.

Related
Become a Member