两种基于神经网络的故障诊断方法

COMPARISON BETWEEN TWO FAULT DIAGNOSIS METHODS BASED ON NEURAL NETWORK

复杂设备或系统的故障诊断中常采用神经网络构建故障映射关系,针对实际应用中神经网络存在收敛速度慢、学习记忆不稳定等不足,以一电站锅炉送风机为诊断对象研究了基于带有偏差单元的递归神经网络故障诊断方法。对状态检测系统采集的信号进行逻辑处理,分离出8个故障特征参数。以8种常见故障模式作为BP网络和递归神经网络的训练样本,对训练过程和仿真结果作了对比分析,结果表明该诊断方法在收敛速度、精度和稳定性能等方面均有良好改善,满足了系统在线故障诊断的需求。

Neural network （NN） is often used to construct a fault mapping for the fault diagnosis of a complex equipment or system. Because there are some problems existing in the normal neural networks such as the long training time and dilemma of stability, a fault diagnosis method based on recurrent neural network （RNN） with deviation error units is presented by taking the forced draught fan as research object. 8 fault characteristic parameters are extracted after the logic processing of the signals from the detection system. 8 common fault patterns are used as fault samples to train the BP NN and the RNN, much comparison and analysis is carried out between the two networks in simulation result and training process. It is shown that the RNN diagnosis approach is better than BP NN method in converging speed, accuracy and stability. It can satisfy the practical need of an on-line fault diagnosis system.