|
|
|
| Fault prediction model of centrifugal pump based on external magnetic field |
| LUO Yin*, CHEN Yinwei, QIN Xuecong, CHEN Yunfei |
| National Research Center of Pumps, Jiangsu University, Zhenjiang, Jiangsu 212013, China |
|
|
|
|
Abstract Aiming at the lack of universality and timeliness of traditional fault malfunction diagnosis method for centrifugal pumps in the complex and changeable engineering environment, a fault malfunction prediction model of centrifugal pumps based on improved KNN algorithm by Mahalanobis distance was proposed. At first, the external magnetic field signals under fault malfunction condition were processed. Accordingly, the corresponding working condition indexes were obtained. Afterwards, the weight analysis was implemented by ReliefF algorithm to extract features, providing a database of malfunction prediction and classification for KNN algorithm. During the process, original Euclidean distance in KNN algorithm was replaced by Mahalanobis distance to eliminate the dimensional influence between feature indexes, so as to improve the accuracy of the prediction results. The key value K of KNN algorithm was sifted by the 10-fold cross validation method. Consequently, the operation result of the prediction model was proved to be the best when the value of K was 120. When the centrifugal pump was running off-design point, the malfunction prediction model established by the improved KNN algorithm could accurately predict the possible failures or faults according to the external magnetic field signals, effectively solving the serious hysteresis of traditional monitoring methods. The model training data contained the external magnetic field signals under various working conditions of 0.2Qd~1.2Qd full flow. The test results show that the malfunction prediction accuracy of the fault prediction model is 0.831 5 at 0.4Qd, 0.799 9 at 0.8Qd, 0.852 7 at 1.0Qd and 0.874 1 at 1.2Qd working condition, respectively, which basically realizes the accurate prediction of centrifugal pump malfunction.
|
|
Received: 29 December 2021
|
|
|
|
|
|
|
