Anyone who has owned a vehicle recognizes the following simple truth: every part of any vehicle deteriorates with use. That includes the engine. Most of the vehicle owning population knows this second thing: it's cheaper to maintain an engine than it is to fix a busted engine. That's why we do regular maintenance and change our oil and
Imagine for a moment that you have a colossally expensive engine. Further, lets presume that your life, your livelihood, or both were wholly dependent on the continued functioning of that engine. You'd certainly be motivated to perform routine maintenance on that engine. You would likely go further, and find ways to monitor the behavior of your engine so that you could know when you need to wrench on it to keep it in proper working order for as long as possible.
What you'd have built is called an engine health monitoring system. They are, in practice, a little more than they claim to be. Certainly engine health monitoring systems (EHMS hereafter) monitor the health of engines, but they often do so in service of current or
impending faults and suggesting maintenance actions. The adage "an ounce of prevention, a pound of cure" applies here; the thought is by noticing and repairing faults before they become critical, the total cost of maintenance of the engine can be substantially reduced.
While you may not own an incredibly expensive engine that drives your existence, plenty of companies do. These can range from small fleets of very intricate and expensive engines of the sort we see in power generation to large fleets of engines that power vehicles. Despite the wide range of applications, what these companies share in common
is a need to keep a fleet of engines running while controlling the costs associated with maintenance, downtime and replacement.
Where We're Going
This post kicks off a series of posts taking a look at the engine health monitoring literature. The rest of this post provides an overview of the literature and builds the structure we'll use to explore decades worth of literature. We're going to start by binning EHM systems into three different regimes: anomaly detection, model based estimators of engine health, and model free estimators of engine health. After we've discussed what distinguishes these three types of EHM systems, we'll follow up by discussing each kind of system in further detail. The in-depth discussions will be relegated to their own posts.
Regimes of Engine Health Monitoring
An initial reading of literature on engine health monitoring and prediction caused me to divide the systems up into three different regimes: Anomaly detection, Model Based Prediction, and Model Free Prediction. The systems have the same goal writ large: increase effective life of engines and reduce the overall cost of maintaining them. What separates the systems into the different groups is how they intend to achieve that overarching goal, what they try to divine from divine from engine performance data, and how that divination occurs.
Descriptive vs Prescriptive
The first differentiator between EHMSs is how they attempt to achieve their overall goals. Some systems provide suggested maintenance actions directly. Others attempt to help humans make those suggests by providing insight into the current performance and likely state of the engines. I'll refer to these as prescriptive and descriptive systems respectively. Anomaly detection systems tend to be descriptive, model based systems tend to be prescriptive, and model free systems frequently operate as either prescriptive or descriptive systems.
This difference is most obvious in what the systems produce in service of monitoring engine health. Anomaly detection systems detect anomalies and thereby group engine performance into two different groups: "That looks fine" and "Hey, that was weird". Model free systems tend to produce similar groupings, as in "The performance of this engine is similar to the performance of these other engines and I hope you know what to do with that information." That said, we often see parameter estimations come out of model free systems, as in "This engine is 0.83 out of 1.0 healthy".
Model based systems are, frequently, different than that in their output. They do often estimate states or values associated with the health of an engine. However, by virtue of having an underlying model, they tend to include a suggested action along with the state.
Model based systems tend to have output along the lines of "This engine is operating nominally, and further operation will cause degradation at the normal rate" and "This part of the engine shakes too much during operation, further operation will cause degradation at a higher than average rate, and this kind of maintenance is recommended."
Model Free vs Model Based
The other axis along which EHMSs differentiate themselves is in how much information about the underlying engine they consume. This axis is almost continuous. Model based systems rely on a model of the engines in question. Model free systems don't, as the name might suggest. That isn't to say that they incorporate no prior knowledge of the system. For example, a model free system might know about a handful of different failure modes of the engine and try to label an engine as being in one of those states. States and actions causing a transition between states are integral to a model-based system. If you don't have both, I believe you are working in the model free
Although model free systems don't model the outcome of actions in the system itself, that doesn't mean that they can't suggest actions. Revisit the notion of a system that classifies engines into several states, some of which are failing. It is natural to associate recommended courses of action with failing states. In fact, you could refine the partitioning of engine states down to the point that every classification had but one recommended action to take. In effect, your model free system would be prescribing a course of action.
What Was, What Is, What Shall Be
Similar to the distinction between descriptive and prescriptive models, EHMS differ on where in the stream of events they attempt to make a divination. Some only attempt to attribute state or value to recorded data points. As such, they only deal with the past, unless data is streamed to the EHMS in real time. Some systems take all of the data up until now and try to make a prediction as to the current state of affairs. Finally, some systems attempt to predict what will happen at some, or any, future point.
Engine health management systems measure, predict, and in some instances act to improve the health of engine systems. There's been a lot of research into EHMS because engines are expensive and failures are also expensive. A key component of modern EHMS is using historic data on engine performance and behavior to predict future behavior and potentially provide maintenance ahead of a failure or to reduce
overall maintenance costs with timely maintenance.
In the next blog post in this series, we'll be looking at anomaly detection for engine health monitoring. Generally, the idea is that the engine is usually ok until something weird happens, and then the engine is probably not ok. Anomaly detection is the means by which we automatically notice that something weird has happened.