Traditionally, predictive analytics has involved utilizing data mining approaches, statistical methods, and machine learning (ML) techniques to create predictive models of future events or outcomes. At the core of these algorithms was the utilization of data to understand the behavior of a given system. In many instances, to create accurate predictive models, large quantities of data are needed and therefore, predictive analytics is also known as “Big Data Analytics” or just “Big Data.”
THE ERA OF SMALL DATA
Physics Informed Predictive Analytics (PIPA) and its Physics Informed Machine Learning (PIML) methods integrate physics-based approaches with these data-driven predictive analytics techniques. The result is that significantly less data is required to create accurate predictive models. Because these physics-informed techniques require significantly less data, than their “Big Data” cousins, therefore these physics-informed techniques are collectively called “Small Data.”
BENEFITS OF PHYSICS INFORMED MACHINE LEARNING
By reducing the quantity of data required for prognostics and predictive maintenance and improving its predictive accuracy, Physics Informed Machine Learning:
- Reduces telemetry bandwidth & network connectivity requirements
- Diminishes network security vulnerability
- Decreases the quantity of hardware instrumentation required
- Reduces the computational overhead required
- Enables On Device/On Edge IoT, thereby eliminating data latency and enabling controls integration for autonomous operations
- Enables more accurate Systems of Systems (SoS) analysis, thereby allowing the Army to better manage its fleet & materiel availability, optimize depot level overhauls & field level repairs, and increase operational readiness
It requires less data than traditional machine learning techniques to be trained and used as a predictive tool.
The models have been constructed from theoretical knowledge; blind test validation shows higher accuracy to test data, field data, etc.
Once the PIML model is trained and validated, its response to new data is instantaneous.
It is computationally inexpensive and therefore it can be run on the edge device.
It provides insight & diagnostics for a given scenario. It gives feedback on How it happened and Why it happened.
The deployment is very flexible therefore the models can be easily integrated into existing IoT platforms.
WHAT IS POSSIBLE WITH PHYSICS INFORMED MACHINE LEARNING?
Small Data enables:
- System Level Performance
- Prognostics and diagnostics
Physics Informed Predictive Analytics leads to more informed decisions leading to millions of dollars in saving and optimized operations
HOW IS PHYSICS INFORMED MACHINE LEARNING DEPLOYED?
- Integration into an existing IoT product environment
- Flexible deployment on multiple platforms
- Requires minimum computing power