一种LSTM神经网络和卡尔曼滤波相结合的复合材料承载预测方法

复合材料是基于多种材料组分的不同组合方式经由相应工艺加工而成的新型材料，近年来凭借其优良的综合性能被广泛应用于交通、建筑等领域，其试验分析结果有时候与经验分析存在较大误差，从而建立可信的分析方法对复合材料承载性能进行验证具有十分重要的理论意义。基于长短期记忆（long short term memory，LSTM）深度学习网络模型的预测精度受数据序列长度影响，提出一种LSTM神经网络和Kalman滤波相结合的复合材料承载预测方法，既可以克服训练数据序列长度对传统LSTM神经网络的影响，又使得Kalman滤波可以从输入数据中学习。仿真结果表明，该方法可以获得优良的预测性能：LSTM-KF模型的承载预测误差将LSTM模型的预测误差从0.033 0 kN减小到0.016 0 kN，降幅为51.52%。

A load forecast method of composite materials based on LSTM network and Kalman filtering

Composite material is a new type of material based on different combinations of various material components. It has been widely used in transportation, construction and other fields because of its excellent comprehensive properties during recent years. Despite of the large errors between experimental analysis and empirical analysis result sometimes,it is of great theoretical significance to establish a credible theoretical analysis to verify the bearing properties of composite materials.Considering that the forecast accuracy is affected by the length of the data sequence when using the memory characteristics of LSTM to predict the load, a load forecast method of composite materials combined with LSTM network and Kalman filtering is proposed. The model can learn from the data avoiding the dependence of traditional Kalman filtering on the dynamic model, at the same time the influence of training data sequence length on traditional LSTM can be overcome to some extent.The results show that the method proposed in this paper can obtain more excellent predictive performance:(1)The performance of LSTM-KF is better than the independent LSTM, and the prediction curve of LSTM-KF is closer to the actual load value;(2)The prediction error of LSTM-KF reduces that of LSTM from 0.033 0 kN to 0.016 0 kN, a decrease of 51.52%.