<p>This article proposes a fault diagnosis method that combines fractional order attention entropy feature extraction and machine learning to solve the problems of large noise interference in the vibration signal of plunger pump bearings and multiple parameter settings in fault diagnosis algorithms. Firstly, expands the attention entropy with the ability of signal morphology representation in fractional order and gives the method to determine the optimal parameters. Secondly, upgrades the fractional-order attention entropy using the fine composite multi-scale analysis algorithm, the improved hierarchical decomposition algorithm and the multi-channel data analysis method, so that it can extract the high-dimensional features with stronger representation ability. And then the Sparse Softmax Feature Selection (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\({S}^{2}FS\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mrow> <mi>S</mi> </mrow> <mn>2</mn> </msup> <mi>F</mi> <mi>S</mi> </mrow> </math></EquationSource> </InlineEquation>) algorithm is used to screen the most valuable combination of eigenvalues to improve the efficiency of fault diagnosis. Finally, the distributed weighted adaptive nearest neighbor classifier (DWANN) proposed in this paper is used for fault pattern recognition, in which the parameter K value is selected based on the early stop rule and is adjusted adaptively by data driven. The experimental design in this paper is based on the plunger pump full loaded roller bearing, which is rare in vibration signal measurement. When the sample length is 2048, the accuracy of fault type recognition under normal signal and strong noise interference(SNR=-1dB) can reach 100% and 91.67% respectively.</p>