Anomalies in a business setting aren’t usually good. But sometimes we can put anomalies to work and use them to our advantage. We can use information about anomalies to:
- determine process exceptions
- detect fraud and theft
- detect security intrusions
- predict future service disruptions, outages, and failures
In our post, Confident Anomaly Detection: Overcoming Data Issues, we highlighted common obstacles to detecting anomalies in your assets’ or operations’ data. We also described some of the paths around these challenges that we’ve recently deployed for customers.
How do we go about putting anomalies to work for us? (Hint: we build and deploy a rock-solid analytical model.) Let’s focus on predicting future service disruptions from failure of our industrial assets. Industrial assets pose specific problems:
- typically have some sort of mechanical or degradative parts that require condition or state monitoring
- require expensive maintenance to keep in fine working order
- are expensive to buy, repair, and replace
- are core to our business and productivity, so outages and failures can be anywhere from costly to catastrophic
The main challenge is finding anomalies in industrial assets’ condition, performance, or output to avoid the catastrophic – and the costly. We want a reliable analytical model that will help us do just that. Our predictive model must be built on four essential pillars.
Learn continuously. Not all variations in operational performance will be defined during the model’s training period. New anomalies will occur, and the model must be capable of learning that these variants are also anomalies and not part of the definition of ‘normal’. This is fundamental to and a key strength of an unsupervised machine learning approach. Machine learning models can learn in real time from streaming data or from periodic ingestion of operational data from assets.
Predict events individually. A condition-based maintenance (CBM) approach, the improvement that industrial organizations often use to replace scheduled maintenance, suffers from the disadvantage of using a single, stationary distribution model to predict the possibility of asset failure. This model is based on historical averages about how the asset model has degraded in the past, but lacks the ability to confidently predict a future event for a specific asset. SpaceTime’s machine learning approach will calculate the probability of failure for each specific asset.
Offer extensibility. The model shouldn’t be specific to one asset model; it should be extensible across the asset class. If we want to predict the energy storage capacity of a utility’s battery storage system, the model predicting the energy storage degradation rate should transfer across manufacturer and model.
Deliver confidence. Ultimately the goal is to make better data-driven decisions. Better implies decisions can be made with the highest confidence in what our analytics tell us or recommend. While all predictions have a level of uncertainty, machine learning anomaly detection can deliver the model with the highest level statistical confidence.
In the world of industrial-grade assets, nothing lasts forever. But if we listen closely to what our assets are saying, we can find clues to future disruption or failure events. By modeling the data, we can apply those clues to predict the lifecycle of each of our assets.