Abstract: Spare parts business is an important part of the auto parts aftermarket.Aiming at the problems of incomplete information and diversity in the decision-making process of automobile spare parts,a regularized VIT-BiLSTM two-level spare parts decision-making model was proposed.The data was divided into two levels according to the type of auto parts to obtain its inner connection.Then,the Vision Transformer (VIT) model was used to extract key features from the accessory data.The bidirectional long-term dependencies between features were captured by the Bidirectional Long Short Term Memory (BiLSTM) recurrent neural network.Meanwhile,the group lasso regularization term was incorporated into each sequence unit of BiLSTM to further improve the accuracy of the model.The accessory data of the third-party cloud platform was applied to conduct an example analysis.The experimental results showed that the decision-making accuracy rates of the first-level and second-level models were as high as 99% and 97% respectively,the recall rates were 97.3% and 96.6% respectively,and F values were 0.977 and 0.964 respectively.It showed that the proposed model could provide real-time data reference for spare parts dealers to assist them in making spare parts decisions.

Key words: auto parts, deep network, two-level spare parts decision, vision transformer model, bidirectional long short term memory model, group lasso regularization

摘要： 备件业务是汽车配件售后市场重要组成部分,针对汽车备件决策过程中信息不完备与多样性的问题,提出一种正则化VIT-BiLSTM两级备件决策模型。首先,根据配件类型对数据进行两级划分,以获取其内在联系。然后,利用Vision Transformer(VIT)模型对配件数据进行关键特征的提取。随后,通过双向长短时记忆循环神经网络(BiLSTM)捕捉特征之间的双向长时依赖关系,并在每个序列单元中融入组套索正则化项,进一步提高模型准确率。最后,利用第三方云平台的配件数据进行算例分析。实验结果表明,模型一级与二级的决策准确率分别高达99%、97%,召回率分别为97.3%、96.6%,F值分别为0.977、0.964,说明本模型可以为配件代理商提供实时数据参考,辅助其进行备件决策。

关键词: 汽车配件, 深度网络, 两级备件决策, VIT模型, BiLSTM模型, 组套索正则化