Normal behavior modeling is the state of the art in predictive maintenance of complex systems and equipment. It simultaneously automates the complicated performance data analysis process while minimizing alert fatigue and false positives. It allows the monitoring system to adapt and evolve its understanding of the normal state of an asset as it ages, enabling alerts to be based on the complex and frequently unclear interactions between the many components and parameters within (and sometimes outside of) the system. For a deeper dive into NBM technology, check out our white paper Predictive Maintenance Using Normal Behavior Modeling. The following is an abbreviated version of how NBM works and how it can transform your business.
Modern industrial equipment routinely costs millions of dollars to purchase and operate. The goal of any industrial concern—whether manufacturing, power generation, or aviation—is, therefore, not only to employ that equipment to maximize production (and hence revenue) but also to manage ongoing expenses by minimizing routine or unscheduled maintenance and increasing the useful life of these expensive capital assets.
Historically, these goals have been pursued using condition-based monitoring (CBM) solutions or OEM-provided asset management tools incorporating physics-based models. However, in a world where equipment is increasingly reliable, and failure data is scarce, traditional approaches are often far less effective than they could be. New techniques are thus required to achieve these goals, i.e., minimizing operational costs while maximizing revenue generation and equipment lifetime more proactively and effectively.
Normal behavior modeling (NBM)—typically used in the field of predictive maintenance—is an anomaly detection technique that can be applied to a complex system or a single piece of equipment, such as a hydraulic pump or wind turbine. Once trained to understand a system’s ‘normal’ operating state, the model continues to evaluate the incoming data stream of sensor data and generates an alert when an anomalous condition is detected.
However, long before the term NBM came into existence, equipment and process operators were tasked with keeping things running for as long and as continuously as possible. Historically, this challenge has been addressed with approaches ranging from repair or replacement of an entirely failed device (run-to-failure maintenance) to following an OEM-provided replacement schedule (preventive maintenance) to attempting to foresee impending failures (predictive maintenance) with statistical process control (SPC), condition-based maintenance (CBM), and physics-based models (PBM). However, all of these techniques suffer from one or more common challenges:
- Focus on a single device or piece of equipment
- Indication of a malfunction only after the problem is already occurring
- The need for mathematical knowledge of an ideal machine and how the real machine deviates from that ideal
With the turn of the 21st century, the ubiquity of sensors, and the arrival of big data, high-bandwidth data transmission, and petaflop computing capabilities, a new system-wide performance assessment strategy can now be achieved proactively and cost-effectively. Normal behavior modeling satisfies this need with increasingly great precision and effectiveness.
Holistic system understanding
NBM offers several advantages, the most significant of which is its ability to continuously monitor and provide proactive alerts of impending problems on a complex system, basing these alerts on a holistic understanding of the system throughout the asset’s life. By contrast, traditional approaches to system monitoring quickly become stale over time as they fail to recognize the inherently interconnected nature of complex systems, not only to other components in the system but also to the external environment.
An NBM model is not concerned with what kind of equipment it is monitoring. Whether it’s keeping tabs on a wind turbine, fuel pump, or nuclear reactor is irrelevant since the model is simply evaluating a data stream, developing its understanding of normality, and generating alerts whenever it perceives that normality has been violated.
Besides holistic system understanding and data agnosticism, NBM also provides:
- More effective monitoring (internally or externally) of dynamic environments.
- Alerts on anomalies long BEFORE failures occur, saving money and time.
- Ideal balance between alert fatigue from too many false positives or missed problems due to too many false negatives.
Central to the NBM concept is an algorithm known as an autoencoder, shown in Figure 1. Over time, the autoencoder’s input layer ingests a continuous stream of quantitative data from equipment sensors (temperature, pressure, etc.). This data is then fed to a hidden layer (of which there are typically several), where it gets compressed. Numerical weights (a value between 0 and 1) are then applied to each node, with the goal of eventually reproducing the input values at the output layer.
Achieving this outcome typically requires many thousands of iterations, with the weights tweaked slightly after each iteration until the outputs (X’n) have finally achieved parity (or as close to it as possible) with the original inputs (Xn). At this point, the model is said to have learned the normal state necessary to yield subsequent actionable alerts.
Scoring and alerts
One of the primary benefits of the NBM approach is that it reduces many thousands of individual time-series data points to a single metric that can be acted upon. This single metric is known as a risk score, a statistical summary of the extent to which the system’s overall performance deviates from the normal state of the system.
Determining whether or not maintenance action should be taken based on the risk score is a function of how sensitive we want the anomaly detection to be. In the most basic version of the NBM technique, we would simply assign a min/max threshold value for each risk score. In more advanced approaches, the number and duration of threshold-exceeding instances can be used to improve the model’s sensitivity.
Retraining and evolving normals
As opposed to more static traditional approaches, NBM systems can evolve to take into account equipment aging, maintenance/repairs, regulatory changes, and externalities like availability of time, people, and money. NBM enables this flexibility by retraining from time to time to reflect the latest reality. This process occurs in the same manner as initial training, except that results will be improved in subsequent iterations due to the availability of larger and more comprehensive data sets.
The principal purpose of NBM is to define the normal state of a complex system and then proactively identify instances where the system is operating outside of normal with sufficient advance warning to allow maintenance or repair actions to take place to avoid revenue loss, repair costs, and safety compromises that typically come with such failures.
There are many examples of complex systems to which NBM techniques can be applied:
- Production equipment on oil platforms—By modeling normal equipment temperatures, pressures, and rotation and flow rates, incipient problems can be identified early, saving upstream operators millions of dollars and significant regulatory exposure.
- Manufacturing plants—Out-of-normal operations in manufacturing plants can result in safety hazards, environmental violations, and inferior quality in the products being produced.
- Commercial and military aviation—Jet engines and other complex airborne hardware are routinely subject to enormous operational stresses. Minor problems can quickly cascade into expensive and dangerous situations, risking lives and the loss of enormous capital investments.
- Electric power generation, transmission, and distribution—Electric generation assets, whether renewable or traditional fossil-fuel-powered, require extremely high-reliability performance, making these assets ideal candidates for predictive maintenance methodologies like NBM.
To learn more about how SparkCognition uses NBM techniques to enable predictive maintenance, read our white paper Predictive Maintenance Using Normal Behavior Modeling.