Equipment Anomaly Detection (EAD) is a neural network based application for detecting equipment anomalies. EAD automatically learns relationships between hundreds of equipment parameters. EAD models are easily updated as new normal operating conditions are encountered and can provide contextual data for easy action planning or other disposition. EAD can be an ideal option in certain operational conditions, such as when limited domain expertise is available for initial model build, model training, and general day-to-day use by system engineers.

The EAD system will:

• Help identify early indications of equipment failures
• Improve equipment reliability
• Allow preventative actions to mitigate failure
• Detect anomalies and deviations from the standard operating parameters of the equipment
• Indicate the beginning of an equipment failure

EAD uses a specific type of artificial recurrent neural network (RNN) architecture composed of long short-term memory (LSTM) autoencoder units. This architecture has demonstrated success in producing expected values for its collection of modeled sensors based not only on the current dataset, but also upon the data from a configured number of time steps into the past.

As such, EAD can effectively monitor transient behavior like startups, shutdowns, and rapid operating changes that are more challenging for similarity-based models like PdP.

In addition, RNNs can evaluate a large number of sensors, including vast combinations of digital inputs. Combined, these features provide anomaly detection functionality that is sensitive to degradation with few false-positive indications.

EAD models are constructed and trained with historical data using the same Windows based client as PdP—the Smart Platform Architect.

Trained models are deployed to the Smart Platform server to produce runtime predictions.

Real-time alarms and diagnostic tools are provided to end users on the Smart Platform integrated web-based GUI alongside PdP and StressWave analytics.

Using historical data, EAD can produce robust, efficient models that can be trained with little involvement from subject matter experts yet still provide early warning of equipment failures with few nuisance alarms. EAD also includes limited alarm functionality based on the deviation of individual sensor behaviors from predicted values (residual) for any given snapshot of data.

In addition, the roadmap for future releases includes an integration layer for retrieving contextual data (such as CMMS and logs), updating them according to findings, and utilizing information derived in anomaly disposition.