A Ph.D. in industrial engineering sits down for lunch with a colleague who is a Six Sigma black belt. The Ph.D. talks about her research in physics-based modeling of metal fatigue and her dozens of publications. The Six Sigmas Black Belt is a reliability engineer. He is responsible for asset uptime in his company. He talks about gear teeth breaking off and ball bearings that wear out and lubricant failure that all create headaches for his preventative maintenance program, which currently seems like so much guesswork.
The Ph.D. says all of those processes can be modeled experimentally. For instance, Paris’ Law describes the rate of crack growth in gears. She says they could use high-performance computers and finite element analysis to create experimental models describing the stress and fatigue of specific gears under various conditions. The more they talk, the more excited they get. Could they build a physics-based, predictive maintenance strategy, using her scientific knowledge and his asset subject matter expertise?
Is Physics-Based Modeling Deployable?
The imaginary conversation above between the Ph.D. and the Six Sigma Black Belt illustrates the attractiveness of this approach to developing a predictive maintenance strategy. But is physics-based modeling a strategy that can be deployed and maintained at scale. In their paper “A review of physics-based modeling in prognostics: Application to gears and bearings of rotating machinery”, Cubillo, et al. summarize the state of physics-based modeling in asset health management. They cover a wide variety of traditional and machine learning approaches to physics-based modeling. Here are some of their conclusions.
Why Physics-Based Models are Attractive
An asset in normal operation will experience a variety of types of increasing mechanical wear including abrasion, cracking, pitting, and spalling of key components such as gear teeth and drive shafts. This is called asset degradation and is the primary focus of physics-based predictive models, which attempt to use mathematical models based on the underlying physical principles that produce each type of mechanical wear. Physics-based models hold the promise of telling you how much longer you can run an asset before it fails.
Physics-Based Models Provide Limited Insight
A physics-based model will generally only describe one degradation mechanism (e.g. creep or fatigue or wear) as the mathematical principles around each mechanism are unique. As most complex assets experience many different types of degradation, a physics-based model that is useful for predicting remaining useful asset life will be very complex indeed.
Reality “Complicates” Physical-Based Modeling
Physics-based models still need significant sensor telemetry. At the outset they seem promising as they seem to need just enough telemetry to solve the systems of equations that will produce each mathematical model. However, in practice they need ongoing asset measurements to compute the error between the predicted state of the asset and its actual measured state. This error drives a new dynamic of trying to determine how the ongoing error between the model and measurements is affecting the predictive power of the model. One reason for the error is that the physics of many failures are not well understood and the measurements needed to model those failures (for instance, mechanical wear dynamics buried deep within the asset) are not accessible.
Physics-Based Models are not Scalable
Physical-based models, while very intriguing for academic research, are impractical for large-scale, commercial deployments. Creating even a single physics-based model will require hundreds of hours of PhD-level expertise, and it should not be surprising that physics-based models are not “universal” models, that is, they do not translate readily from one asset to another similar asset that is experiencing different production modes and environmental factors. (See our blog on universal models.)
Amber is the only predictive maintenance technology that captures the complex multi-variable relationships of high-value assets and turns those into actionable numerical scores.
- Amber is self-tuning: Amber starts with a blank slate and uses either live asset data or historical data to automatically find the clusters that describe compliant asset operation for that asset’s usage pattern and current maintenance history. This also means that Amber is scalable. Connecting it to 100 assets so it can train 100 individualized models is no more difficult than connecting it to one.
- Hyper-Longevity: Amber will tell you if an asset is healthy and performing compliantly beyond its expected useful life. This saves the time and money expended with needless scheduled maintenance cycles on a healthy asset.
- Premature Aging: Amber tells you if an asset is unhealthy and showing issues prior to the end of its expected useful lifespan. This avoids the cost of unplanned downtime when an asset fails much earlier than expected.
- Root Cause Analysis: An Amber model trained on a subsystem does more than just tell that the subsystem is changing. It will localize that change to the few sensors on the subsystem implicated in the change. When an asset maintenance team knows which sensors are implicated, they will often know what type of failure is imminent.
- Dynamic Trending: Amber measures not only where the change is happening, but it also quantifies that change as it evolves over time. Amber provides an overall trajectory of asset health, so it provides the earliest possible detection of changes in asset health and can show both asset health “improvement” scenarios and the increasing risk as asset health trends downward.