一种基于神经网络的轮胎慢漏气时间序列预测方法

A Neural Network-based Method for Predicting Tire Slow Leak Time Series

轮胎胎压不足的状况一旦发生,易导致行车过程中车辆的失控并带来不可逆的交通伤亡事故。而轮胎的慢漏气故障是一种常见的交通事故诱因,且该诱因不易察觉。因此,为了及时预测到轮胎的慢漏气故障,本研究以某型纯电动车的轮胎胎压时间序列变化数据为基准数据,实现改进的长短期记忆网络,建构基于神经网络的轮胎慢漏气时间序列预测模型。首先,分别进行原始时间序列的正常胎压变化趋势与异常胎压变化趋势的数据筛选与清洗等工作;其次,分别基于变分模态分解及自适应差分进化算法,实现长短期记忆网络的模型训练;最终,分别基于慢漏气时间序列校验集,进行消融实验的预测结果评估与可视化样例的对比分析。消融实验结果相较于基础的长短期记忆网络提升了15%左右的性能,可视化样例中大部分的慢漏气时间预测差值波动范围在6h内。综合实验结果可验证本研究所实现的基于神经网络的轮胎慢漏气时间序列预测模型的优越性。

Once a tire pressure deficiency occurs,it will easily lead to loss of control of the vehicle during driving and bring irreversible traffic injuries and fatalities. The slow air leakage failure of tires is a common cause of traffic accidents,and this cause is not easily detectable. Therefore,in order to predict the slow air leakage failure of tires in time,a time series of tire pressure variation data of a pure electric vehicle is used as the benchmark data in this study. An improved long and short-term memory network is implemented to construct a neural network-based time series prediction model of tire slow air leakage. Firstly, the data filtering and cleaning of the normal and abnormal tire pressure trends of the original time series are carried out respectively;secondly,the model training of the long and short-term memory network is realized based on the variational modal decomposition and adaptive differential evolution algorithm respectively;finally,the evaluation of the prediction results of the ablation experiment and the comparison analysis of the visualized samples based on the calibration set of the slow air leakage time series are performed respectively. The results of the ablation experiment improved the performance by about 15% compared with the basic long-and short-term memory network;most of the slow air leakage time prediction differences in the visualization samples fluctuated within 6 hours. The comprehensive experimental results can verify the superiority of the neural network-based tire slow air leak time series prediction model implemented in this study.