钢轨打磨是一种利用高速旋转的砂轮去除钢轨表面疲劳层的技术。当打磨砂轮钝化后，一般会出现材料去除效率降低和磨削区域温度升高等问题。为避免上述问题对打磨作业的负面影响，需及时更换寿命到限的砂轮。该文提出一种基于神经网络的钢轨打磨砂轮磨损程度与寿命预测方法，以合理地确定砂轮更换时机。该方法原理如下：采集与砂轮连接的电机主轴轴向加速度信号，并基于该信号提取出描述砂轮磨损程度的特征参数。对特征参数进行Z-Score变换，去除特征参数量纲并提高各参数间的可比性。利用XGBoost算法依据各特征参数的平均增益大小进行筛选，选择与寿命强相关的特征参数作为分析对象进一步处理。以磨损时间与砂轮磨损量融合策略作为判断砂轮磨损程度与寿命的标准。搭建神经网络建立筛选后的特征参数与砂轮磨损量、砂轮厚度的映射关系。利用试验装置获得伴随砂轮钝化的试验数据，将数据划分为相互独立的训练集与验证集，分别对搭建的神经网络进行训练与验证。结果表明，该方法在训练集与验证集中的判断正确率与预测精度基本持平。在验证集中，对砂轮磨损程度的判断正确率为87.9%，判断错误样本主要集中在磨损程度变化的过渡区间；对砂轮寿命的预测精度为-5.3%。不同打磨工艺参数下，该方法具有一定的通用性。

Rail grinding is a technology that removes the top fatigue layer of rails by applying high speed rotating grinding wheels. When the grinding wheels become abrasion, it typically leads to a decrease in material removal efficiency and an increase in grinding zone temperature. To prevent these issues from negatively affecting grinding operations, timely replacement of worn-out grinding wheels is necessary. This paper proposed a neural network-based method for predicting the wear level and residual life of rail grinding wheels, enabling the optimal timing for wheel replacement. The principle of this method is as follows: the axial acceleration signal of the motor spindle connected to the grinding wheel was collected, and based on this signal, characteristic parameters that describe the wear level of the grinding wheel were extracted. These characteristic parameters were then transformed using the Z-score method, which removes the dimensionality of the parameters and improves their comparability. After that, the XGBoost algorithm was employed to filter the most relevant features that are strongly correlated with the wheel’s service life. A fusion strategy integrating wear time and wheel wear volume was adopted as the criterion for assessing the wear and service life. Constructing a neural network model that mapped the filtered feature parameters to both the wear volume and the grinding wheel thickness. The experimental data along with the grinding wheels abrasion was obtained by using the experimental device. The data was divided into mutually independent training and validation sets, which were used to train and validate the constructed neural network, respectively. The results show that the method achieves comparable accuracy in both the training and validation sets. In the validation set, the accuracy for determining the wear level of the grinding wheel is 87.9%, with misclassified samples mainly concentrated in the transition zone of wear progression. The prediction accuracy for the grinding wheel life is -5.3%. Additionally, the method demonstrates a certain level of generalizability across different grinding process parameters.