The goal of any company is to employ assets to maximize production and revenue. So it’s crucial to manage expenses by minimizing routine or unscheduled maintenance. It all boils down to increasing the useful life of expensive capital equipment. Historically, these goals have been pursued using condition-based monitoring (CBM) solutions or OEM-provided asset management tools. However, in a world where equipment is increasingly reliable, and failure data scarce, traditional approaches are far less effective than they could be.
Normal behavior modeling: a new AI tool for an old problem
Artificial intelligence-enabled Normal Behavior Modeling (NBM) in predictive maintenance of complex systems, simultaneously automates performance data analysis while minimizing alert fatigue and false positives. It allows the monitoring system to adapt, evolving its understanding of the normal state of an asset as it ages, providing alerts based on complex interactions between the many components and parameters of the system. NBM can be applied to complex systems or single pieces of equipment, such as a hydraulic pump or wind turbine. Once trained to understand the system’s normal operating state, the model continuously evaluates the incoming stream of sensor data and generates alerts whenever anomalies are detected.
NBM offers several advantages, the most significant being its ability to base alerts on a holistic understanding of the system throughout its lifetime. Traditional approaches quickly become stale with time, failing to recognize the inherently interconnected nature of complex systems.
NBM models are unconcerned with the type of equipment being monitored. Whether wind turbine or nuclear reactor, the model simply evaluates a data stream, develops its understanding of normality, and generates alerts when it perceives that normality has been violated.
Making the most of performance data with AI
NBM is based on algorithms known as autoencoders, whose input layers ingest a continuous stream of quantitative data from equipment sensors. This data is then compressed. Numerical weights are applied to each node, with the goal of reproducing the input values at the output layer. Achieving this typically requires thousands of iterations, with weights tweaked slightly after each iteration. Once input/output parity is achieved, the model is said to have learned the normal state necessary to yield subsequent actionable alerts.
One of the primary benefits of NBM is that it reduces thousands of time-series data points to a single metric. This is known as a risk score. Determining whether maintenance action should be taken based on the risk score depends on the sensitivity of anomaly detection. In basic versions of NBM, there’s a min/max threshold value for each risk score. In more advanced approaches, the number and duration of threshold-exceeding instances is used to improve model sensitivity. NBM systems can evolve to take into account equipment aging, maintenance/repairs, and externalities like availability of time, people, and money.
National Grid’s Grain LNG terminal benefits from AI solutions
National Grid (NG), headquartered in the UK, is a multinational provider of electricity and gas products and services. The company operates numerous liquid natural gas (LNG) terminals, including a large facility in Kent, the Grain terminal. Beginning in 2021, NG partnered with Avathon to operate a proof-of-concept pilot of NBM technology for several cryogenic pumps and compressors.
Over the first two phases of the pilot, 17 cryogenic pumps were modeled using five months of historical performance data and one month of validation for phase one, followed by 2.5 years of historical data and six months of validation for phase two. Once this work was complete, four selected compressors were then modeled using 14 months of training data, followed by six months of validation. Across all 21 assets, three AI models were created running over three years of performance data.
Following model training and validation, the models successfully identified 75% of compressor events, with an average lead time of eight days, and a false alert rate of just one alert per 1-3 months of operational time. The solution detected 75% – 94% of compressor events, resulting in a potential operational savings of $1.46 million per year.
In addition to reducing compressor priority 1 and 2 events from around 18 to somewhere near 4 per year (saving nearly $65k/event), the solution also improved the expected lifetime of monitored assets by around 8 years. By providing more than a week of lead time, the system enabled technicians to schedule preventive maintenance in the most efficient manner possible; this minimized service disruptions and outages. The near elimination of false alerts saved the organization significant money and staff time.
NBM uses the power of AI to optimize performance and extend the lifetimes of complex industrial assets like those used by National Grid. By continuously and holistically assessing the operation of these systems, companies reduce capital expenditures, minimize operating expenses, and enhance the safety of their operations.
To learn more about Normal Behavior Modeling and the industries Avathon serves, click HERE.
