An anomaly detection method for rotating machinery monitoring based on the most representative data

With the development of concepts of industry 4.0, condition monitoring techniques are changing. Large amounts of generated data require diagnostic procedures to be automated, which drives the need for new and better methods of autonomous interpretations of vibration condition monitoring data. However, if new methods are to be operational, they need to be verified under real industrial conditions and compared with well-established expert-based diagnostic techniques. This article introduces the novel algorithm of data preprocessing for the nearest-neighbor-based anomaly detection. This approach is validated on real industrial machinery in a series of case studies. The population of over-hung centrifugal fans, employed in the same industrial process, were monitored continuously according to the proposed methodology for an extended time period. Piezoceramic accelerometers were used to register time-domain vibration data. The data were processed to extract several signal features to serve as inputs to the anomaly detection algorithm. The novel solution is compared to the well-established condition monitoring approach. Presented data include not only the intact state of machinery but also a machine breakdown case and serious deterioration of the machine condition. The influence of maintenance work is also presented in the article. Authors show the data-driven approach to condition monitoring, which can be used as one of many predictive maintenance techniques.