A Novelty Detection Approach to Compressor Valve Fault Detection and Localization

Reciprocating gas compressors represent one of the most widely used industrial compressors today. Their popularity is largely due to their ability to operate over a wide range of temperatures, pressures, and gas types. The main drawback to this type of compressor tends to be the cost of maintenance, which often eclipses the cost to maintain a similarly sized centrifugal compressor. Much of the explanation in maintenance cost can be directly attributed to valve manifold wear and failure, which accounts for 36% of compressor shut-downs and 50% of overall repair cost. Continuous condition monitoring analysis can provide insight into valve health without the need to shut down and inspect the compressor. Moreover, these solutions can be provided to customers as a retrofit of potentially decades old units already in service in the field. The challenge in developing data-driven methodologies is that key measurements (beyond pressure and temperature) are often not available as the failure is developing, and certainly not in enough units to train a classification methodology. Furthermore, compressor valve failure can occur in a number of ways that make testing in the lab difficult to seed. For this work a novelty detection technique, when applied to reciprocating compressor vibration signals, can provide a way of monitoring the health of these valve manifolds without the need for unhealthy data for classification.

The experimental test unit used in for this work is a Dresser-Rand ESH-1 reciprocating compressor located at the Rochester Institute of Technology's (RIT) Compression Test Cell. The single stage, dual acting compressor was donated by Dresser-Rand and installed at RIT in 2010. One of their smaller industrial compressor, the ESH-1 is driven by a 10hp electric motor and is commonly used in the petrochemical industry.  It is an intermittent flow, positive displacement air compressor with a single piston which pressurizes cylinders on both sides of the piston head, and can be operated under full load, half-load, or unloaded. Each cylinder has a set of inlet suction valves which allow air to be drawn in at atmospheric pressure, and a set of discharge valves which allow compressed air to be discharged into a receiver tank. Each valve assembly includes 16 individual valves, each with a poppet and spring to keep the valves closed until a pressure imbalance is achieved. Condition monitoring of these valve-spring manifold assemblies is the focus of this research.

Presented in this paper is the development of a vibration-based novelty detection algorithm for locating and identifying valve wear within industrial reciprocating compressors through the combined use of time-frequency analysis, image-based pattern recognition, and one-class support vector machines. A commonly reported cause of valve wear related machine down time is wear in the valve seat, causing a change in flow profile into and out of the compression chamber. Seeded faults were introduced into the valve manifolds of the ESH-1 industrial compressor and vibration data collected and separated into individual crank cycles before being analyzed using a Short-time Fourier Transform. The resulting time-frequency data was processed as an image, and features used for classification were extracted using 1st and 2nd order images statistics and shape factors. A one-class support vector machine learning algorithm was then trained using data collected during healthy operation and then used to both detect and locate anomalous valve behavior with a greater than 70% success rate.