Contact :

One of the tools for preventive or predictive maintenance in centrifugal compressors is a routine of performance tests on them.

To carry out these tests there are codes such as: ASME PTC 10 Performance Test Code on Compressors and Exhausters.

Beyond the importance of these tests to evaluate the performance of the centrifugal compressor in the face of surge and stonewall, one may wonder why it is important to plan a routine for performance tests, since there are online measurements of the operating parameters (flow, speed, temperatures and suction and discharge pressures), even more so to the values ​​measured online, machine learning techniques can be applied.

One answer is related to metrology and has to do with carrying out the tests under the same method and under the same conditions.

Conceptually, the method comprises meeting requierements before , during the test and at some point after it.

Prior to the tests, the quality of the measurements should be evaluated, maintenance diagnostic tasks are carried out such as the use of thermographic cameras to detect the leakage of by-pass or recycle valves, parameters of vibrations, temperatures of bearing and parameters related to the lubrication circuit, especially the presence of metallic particles in the analysis of oil samples.

Furthermore, some details such as the location of the temperature measurement thermowells and their quantity (in the case of compressor discharge) and the gas flow profile in the suction.

During the test, where samples of the operating values of interest are taken, the operation must be stable, the differences between the recorded values must be less than the specified tolerances.

Finally, carrying out the tests under the same method and under the same conditions, allows us to compare the values ​​between tests, correctly.

As results of the tests, two values ​​are important: the head and the thermal performance of the compressor. These values are contrasted with the manufacturer’s design conditions, locating the points on the respective maps:

The uncertainty in the determination of both values could be around 2%.

A machine learning algorithm to predict the polytropic head and the thermal performance of the centrifugal compressor, may not make sense, since it can be calculated with greater accuracy, but if you can provide some information the feature importance, this is a measure of how important is a variable in the prediction of a value, with respect to the rest of the variables. We will analyze the case of thermal performance.

The mechanisms of thermal performance loss that we can cite are: Skin friction loss, Vaneless diffuser loss, Vaned diffuser loss, Leakage loss, Disk friction loss, Clearance loss, Recirculation loss, Incidence loss, Blade loading loss, and others.

For each equation, the variables have a coefficient of sensitivity different from the rest, for example, in the case of the thermal efficiency equation, a variation of one degree centigrade in the discharge temperature of the centrifugal compressor implies a greater variation in the thermal efficiency than a variation of one bar of pressure, this is due to the sensitivity coefficients in the enthalpy differences.

The situation is different for each equation of the performance loss mechanisms mentioned above, where the pressure takes importance.

We can illustrate this with the following examples, where the loss models described above are mathematically applied, which shows the feature importance values for a prediction of thermal efficiency natural gas centrifugal compressor with good thermal performance (estimated within 2% of the value specified by the manufacturer) and another with poor thermal performance (estimated below 5% of the value specified by the manufacturer). Machine model for prediction used is gradient boost.Remember that we are not interested in the predicted value, but in the importance of the variable in the prediction.

Feature importance for good thermal performance:

Feature importance for poor thermal performance:

Where:

Conclusions:

This example tries to illustrate, the application of one of the pillars in data science, which is the expertise in the field in which it is applied, in this case recognizing the importance of the traditional method and in what aspect the application of machine learning could help

Source of the picture below: https://www.quora.com/What-fields-within-Data-Science-are-the-most-relevant