(137b) Adaptive Anomaly Detection and Its Application to Online Asset Monitoring

Asset failure modes are often complex. Even with multiple examples of the same failure type there can be significant differences between the process data for each case. This, coupled with the limited number of examples of each failure mode available in an industrial setting can make data-driven modelling of asset failure using typical machine learning techniques very difficult.

To overcome these issues, we have developed a method of data-driven anomaly detection that can effectively identify anomalous time series behaviour based on a small portion of past ‘normal’ operation. This removes the reliance on large datasets associated with typical supervised or unsupervised machine learning solutions.

This anomaly detection system was used as the basis of an online monitoring solution for a two stage hyper-compressor. The specific problem being addressed related to the identification and tracking of packing ring failures for the compressor to eliminate the corresponding unplanned down-time.

Adaptive Anomaly Detection

This method consists of several separate and independent algorithms that monitor time series data in real-time, identifying any points in the data that deviate from normal behaviour. The specific definition of normal and abnormal is different in each algorithm, meaning that a wide range of behaviour can be covered by the system.

A range of statistical methods are used to build the underlying algorithms including Bayesian change point detection, adaptive windowing, k-means clustering, and statistical variance analysis. Each algorithm continually updates its model of normal behaviour allowing it to adapt to new operational modes automatically. This allows the anomaly detection system to run continuously without any intervention, even when conditions change.

In certain circumstances it may be appropriate to use a single anomaly detection algorithm. However, it has been found that combining the outputs of several algorithms minimises the number of false positives detected by the system. To combine the results of the algorithms a quorum voting system is employed. This means that an anomaly is only reported if the majority of the algorithms agree on the result. To account for the difference in techniques between the algorithms the results do not have to align exactly in time, but within an automatically determined time window.

This method of anomaly detection can be used to identify asset failure without needing a specific model of the failure mode, and its adaptive nature allows it to be used in any environment, even those that regularly change operating conditions.

Asset Monitoring Case Study

A two-stage hyper compressor was identified as a key asset in a customer’s process. This hyper-compressor has 20 cylinders, each cylinder has a packing box containing 5 packing rings that are responsible for the sealing of the system. The packing boxes must be replaced periodically as over time the rings degrade and fail. If the packing boxes are not changed in time and all the rings fail, an automatic plant shutdown is triggered. This unplanned down-time lasts for approximately 2 days and costs in the region of £500,000 each time.

The packing boxes themselves are sealed and cannot be inspected, but there is a limited amount of process data available for the asset: total leak gas for the whole compressor, plunger temperature for each cylinder, and x-runout and y-runout (measures of displacement) for each cylinder. The challenge was to create a system that could monitor this process data to determine when a ring had failed, and in which of the 20 packing boxes the failure had occurred. With this information it would be possible to track the status of each packing box and schedule replacements accordingly, avoiding both premature replacement and unplanned outages due to total packing box failures.

To achieve this the anomaly detection system presented above was applied using a two-stage approach. It was found that the total leak gas measurement showed a clear spiking signature when a packing ring failed. This signature was accurately identified by the anomaly detection system with a very low false positive rate.

To identify the specific packing box containing the failed ring the anomaly detection system was then run over the cylinder-specific measurements for each cylinder (60 time series in total). Correlation analysis was then used to find any anomalous behaviour in these signals that corresponds to the anomaly found in the total leak gas. This successfully identified the packing box containing the failed ring.

This compressor monitoring approach was found to be a highly effective method for identifying failed packing rings. As well as successfully identifying all the historical ring failures including the affected packing box, it has also found several confirmed packing ring failures in the live environment since it has been deployed to the customer site. It is now used to determine the packing box replacement schedule.

Summary

An adaptive anomaly detection system has been presented that allows for the real-time monitoring of process data. This statistical method is extremely versatile, and does not require any pre-processed or labelled training data. It can be applied to a wide range of assets for fault detection and automatically adapts to changing plant conditions.

As a case study for the anomaly detection methodology the results of a hyper-compressor monitoring system are presented. This uses anomaly detection applied to over 60 time series to identify packing ring failures. The results show that the anomaly detection system can be successfully deployed as an online asset monitoring tool.