Share

Semi-supervised learning algorithm for running-in analysis on compressors.

Number: 1445

Author(s) : MACHADO J. L., THALER G., FLESCH R. C. C., MACHADO J. P. Z.

Summary

This work presents a semi-supervised machine learning algorithm for identifying the end of the running-in process in reciprocating hermetic compressors. This identification is important for guaranteeing that the compressor reaches its tribological steady state before it is tested. Testing a not run-in compressor can result in erroneous measurements, particularly in energy performance tests. The most widely used procedure to avoid such problems is to operate the compressor under specific conditions for an empirically determined number of hours, but this method does not guarantee the end of the running-in process. The objective of this work is to present a fully automatic method for classification of the compressor state into running-in or tribological steady state. A test rig was built for imposing particular operating conditions for the compressor under test and for acquiring experimental data, such as pressures and temperatures at specific points of the refrigeration circuit, as well as the electric current. Four compressors which have never been turned on before were tested using the proposed rig to acquire data of the running-in phenomenon. The same compressors were tested twice more after operating for several hours to acquire data of compressors which are known to be in a tribological steady state. Processing based on a delayed space sliding window was used in the test time series to subsequence the root mean squared values of the electric current dataset. In addition, a semi-supervised machine learning method named self-training was used, considering k-nearest neighbors (KNN), random forest, and support vector machine methods as classifiers. The only two pieces of information assumed for the proposed method are that at the beginning of the first test of each compressor unit it was not run in yet, and that every other test the compressor was in its tribological steady state. The best classifier for identifying the end of the running-in process was obtained using the KNN method with less than 60 neighbors and a small number of features (4 or less). The results are consistent with comparative analyses and the running-in literature.

Available documents

Format PDF

Pages: 9

Available

Free

Download

Details

  • Original title: Semi-supervised learning algorithm for running-in analysis on compressors.
  • Record ID : 30033618
  • Languages: English
  • Subject: Technology
  • Source: 2024 Purdue Conferences. 27th International Compressor Engineering Conference at Purdue.
  • Publication date: 2024/07/18

Links


See other articles from the proceedings (80)
See the conference proceedings

Indexing