基于概率密度函数的车轮多边形幅值检测算法OA

Polygonal wear fault diagnosis of railway wheels primarily focuses on the identification of harmonic orders and wear amplitudes.Signal analysis methods enable the separation and identification of polygonal wear orders.However,due to the superposition of multivariate vibrations during train operation,accurately identifying wear amplitudes remains challenging.Current quantitative analyses of wheel polygonal wear predominantly rely on big data-driven approaches and achieve acceptable detection accuracy.Nevertheless,the performance of these methods under limited-sample conditions has not been rigorously validated.This paper proposes a feature extraction algorithm based on the probability density functions of axle-box vibration acceleration.A generalized regression neural network(GRNN)is employed to recognize the amplitudes of polygonal wear under limited-sample conditions.The proposed method reconstructs the feature space by enlarging inter-class distances while minimizing intra-class distances according to the morphological variations of the probability density functions.Ablation experiments demonstrate that,compared with models without the proposed feature extraction algorithm,the root mean square errors(RMSEs)of the GRNN,residual neural network(ResNet),and model-agnostic meta-learning(MAML)decrease by 68％,33.8％,and 25.3％,respectively.Additionally,the required training iterations for ResNet and MAML are reduced by 80％and 60％,respectively.GRNN achieves a balanced advantage in training sample efficiency,prediction accuracy,and convergence speed,making it particularly suitable for fault diagnosis under limited-sample conditions.