重庆大学 大数据与软件学院
TP399???????
国家重点研发计划
School of Big Data Software Engineering,Chongqing University
National Key Research and Development and Program of China
针对起重机械设备健康状态多时间单位步长预测中出现的监测数据时间跨度小、数据量密集、特征多维、没有标签的问题，提出了一种结合卷积神经网络（CNN）和编码解码器双向长短期循环神经网络（ED-BLSTM）的起重机械设备健康预测方法。对监测数据进行时序排列，在保证相同输入输出时间步长尺寸的情况下对数据集切分重组；将处理后数据集输入到卷积神经网络，提取出主要特征，得到多维矩阵；采用基于编码解码器的双向长短期循环神经网络对多维矩阵进行训练，建立起重机械多时间单位步长的目标预测模型；从而达到长期预测起重机械设备健康状态的目的。对比实验表明，所提方法的验证损失最多降低0.474%，最少降低0.097%；和预测损失最多降低1.411%，最少降低1.230%相比于目前的主流技术均有降低，其实际预测性能有较大提高，对工业起重机械健康预测技术的发展有积极意义。
Aiming at the problem of multi time-step-length prediction of lifting machinery, a method for health prediction of lifting machinery equipment combining convolutional neural network (CNN) and bidirectional long short-term memory with encoder-decoder(ED-BLSTM) is proposed. Firstly, sorting the monitoring data in chronological order, and then segmenting and reorganizing the data set when ensuring the same input and outputting time step size. Then inputting the processed data set to the convolutional neural network, and extracting its main features to obtain Multi-dimensional matrix. This matrix is used as input to train the bidirectional long and short term cyclic neural network based on the encoder-decoder, and establishing the target prediction model of multi-time unit step of lifting machinery. Comparative experiments show that the verification loss and prediction loss of the proposed method are reduced compared with the current mainstream technology, and its actual prediction performance is greatly improved, which has positive significance for the development of industrial equipment health prediction technology.
