基于经验模式滤波与循环神经网络的水锤压力信号预测

为准确预测水锤信号变化规律,实现对水锤冲击强度和能量等特性的提前预判,针对水锤冲击信号提出了一种基于经验模式分解(empirical mode decomposition,EMD)和循环神经网络(recurrent neural network,RNN)的模型预测方法.首先,通过EMD获取具有不同频率的IMF分量,根据水锤信号的频域特性剔除高频噪声分量并重构信号以实现滤波,滤波后信号能量损失不足0.1%;进而,建立了基于RNN模型的时间序列预测模型,搭建试验平台获取水锤冲击信号,完成了RNN模型的训练和参数调节;随后,对不同流速下水锤冲击信号进行序列预测,在测试集与训练集流速不同条件下,得到了准确的预测结果,表现出一定泛化能力.对比分析预测水锤信号与实际信号,得到R²系数大于0.990 0,幅值和能量损失不足1%,验证了所提出方法的正确性,主要结论和建模方法可为各类输水系统的风险评估、管路监测和健康管理提供理论指导和技术手段.

Prediction for water hammer pressure signal based on empirical mode filter and recurrent neural network

To accurately predict the change trend of water hammer impact signal and get the characte-ristics of impact strength and energy in advance, a prediction model based on Empirical Mode Decomposition(EMD)and Recurrent Neural Network(RNN)was proposed for the pressure signals from water hammer impact. Firstly, a series of Intrinsic Mode Function(IMF)was obtained from EMD method, and then high frequency noise components were eliminated according frequency characteristics in water hammer to realize the filtering and reconstruction. The filtered signal energy loss is less than 0.1% and the smoothness is good. Secondly, a prediction model for time series based on RNN model is established and the test platform is built to obtain the water hammer impact signal, and the RNN model was trained with a few samples. Then, a sequence prediction for water hammer impact signals under different flow rates was realized. Although the flow rates are different between the training and testing sets, the result is accuracy and reliable according assessment based on energy loss, amplitude loss and R² coefficient. Comparing the predicted water hammer signal with the actual signal, the R² coefficient is greater than 0.9900, and the amplitude and energy loss are less than 1%, which verifies the correctness of the proposed method.