【摘要】 针对轴承套圈磨削过程中砂轮磨损状态识别精度低的问题，提出一种基于多尺度卷积神经网络(MSCNN)、黑翅鸢优化算法(BKA)和最小二乘支持向量机(LSSVM)的砂轮磨损状态识别模型。采集不同磨削工况下的砂轮全寿命周期声发射信号，使用部分集成局部特征尺度分解(PELCD)对声发射信号进行降噪处理，选取方差贡献率大于5%的本征尺度分量对信号进行重构；使用MSCNN提取信号特征，同时构建MSCNN全连接层结果特征数据集；最后，将特征集划分为训练集、验证集和测试集，使用BKA算法优化LSSVM的惩罚因子与核参数，以提升模型分类性能，并基于优化后的BKA-LSSVM实现磨损状态的识别。结果表明：经PELCD降噪后，MSCNN-BKA-LSSVM模型对砂轮初期、中期和严重磨损状态的识别准确率分别达到97.613%、96.322%和95.802%;消融实验中，在不同磨削工况下，模型的平均识别准确率达到97.309%,仅使用LSSVM的基准模型准确率为81.502%,加入BKA优化后的BKA-LSSVM模型准确率提升至88.195%。所建模型对砂轮磨损状态具有更好的泛化性能和识别效果。更多还原

【Abstract】 To address the issue of low accuracy in identifying grinding wheel wear states during bearing ring grinding, a wear state identification model based on multi-scale convolutional neural network(MSCNN),black-winged kite optimization algorithm(BKA),and least squares support vector machine(LSSVM) was proposed.Acoustic emission signals over the full life cycle of the grinding wheel were collected under different grinding conditions, and the acoustic emission signals were denoised using partial ensemble local characteristic-scale decomposition(PELCD),and intrinsic scale components with a variance contribution rate greater than 5% were selected to reconstruct the signals.MSCNN was employed to extract signal features, and a feature dataset was constructed from the results of the fully connected layers of MSCNN.Finally, the feature set was divided into training, validation, and testing sets.The penalty factor and kernel parameters of the LSSVM were optimized using the BKA algorithm to enhance the model′s classification performance, and the identification of wear states was achieved based on the optimized BKA-LSSVM model.The results show that after denoising with PELCD,the identification accuracies of the MSCNN-BKA-LSSVM model for initial, intermediate, and severe wear states of the grinding wheel reach 97.613%,96.322%,and 95.802%,respectively.In ablation experiments, under different grinding conditions, the average identification accuracy of the model is 97.309%,while the accuracy of the baseline model using only LSSVM is 81.502%.The accuracy of the BKA-LSSVM model, optimized with BKA,is improved to 88.195%.The proposed model is demonstrated to have better generalization performance and identification effectiveness for grinding wheel wear states.更多还原