基于信息压缩的四层前向网络模型体系结构及其应用

FOUR LAYERED FORWARD ANN HIERARCHY FOR HUGE PERCEPTUAL PROBLEM BASED ON INFORMATION COMPRESSION

关键词:  人工神经元网络  信息压缩＋  网络结构  液体推进剂火箭发动机  故障检测

Key words:  Artificial neural networks  Information compression +  Littice structure  Liquid propellant rocket engine  Fault detection

摘要:

提出了一种基于信息压缩原理的四层前向网络体系结构，用以为液体火箭发动机部分参数建立网络模型。利用所选参数的时间序列对这些参数实现平滑、滤波与预测等处理，提高了基于试车数据驱动的故障监控系统对量测噪声、传感器故障与系统过程噪声的鲁棒性。该四层结构大大缩小了网络权值规模，网络模型的训练采用演化策略。

Abstract:

Due to the shortcomings of three layered forward nearal network,a four layered forward neural network model was constructed based on the information compression principle,and this network hierarchy was used to build a start up plus main stage model for parts of the parameters of liquid rocket engine system.The time series of the input parameters were used to smooth and filter as well as predict themselves.which were required in the health monitoring system for rocket engine.Thus the robustness of the health minitoring system to the measurement and process noises as well as the sensor faults were greatly improved.Evolution Strategies were used to train the weight and coefficients of the activation functions.A procedural training strategy was applied in the optimization of the huge 4 layered network model in which the network was initially trained globally in a whole,then broke down into layers for locally training.In addition,as this network hierarchy has less weight and hidden nodes than the conventional 3 layered forward network model,it is a classical network hierarchy for huge forward network model for function approximation.