基于大规模训练神经网络的微小故障在线检测

神经网络已经广泛应用于系统建模和模式识别领域。但为了逼近未知的参数或者系统动态,需要大量的神经元达到足够的逼近精度,因此导致了计算负荷的增大。运算量制约着大规模神经网络计算,无法使其应用到实际的在线系统中。CPU处理无法保证在线数据的同步运算,需要借助图形处理单元GPU(Graphic Processing Unit)来解决实时性同步运算问题。首先,利用RBF神经网络的持续激励PE(Persistent Excitation)特性对系统输入进行分析,减少神经元的数目且优化设计算法,从而提高逼近精度。其次,基于LabVIEW平台,利用LabVIEW的GPU高性能分析工具包实现神经网络算法和并行计算。最后,在一台航空低速轴流压气机中开发基于大规模训练神经网络的LabVIEW系统。实验结果表明,提出的方法可以实现对系统的在线实时运行,满足航空失速检测的要求。

Online Detection of Incipient Fault Based on Large-scale Neural Networks

Neural networks have been widely used for the system modeling and pattern recognition.However,in order to approximate the unknown parameters or system dynamics,it needs enough neurons to achieve sufficiently accurate approximation,which leads to increase of the computational cost.The computation would restrict the online application of the large-scale neural networks.Because CPU processing cannot keep pace with online data capture,the commonly available graphics processors are used for the bulk of data processing in online systems.First,the input of the system was analyzed by persistent excitation characteristics of RBF neural network,reducing the number of neurons and optimizing design optimization algorithm to improve the approximation error.Secondly,LabVIEW and LabVIEW GPU analysis toolkit were used to achieve algorithm implementation and parallel computing.Finally,online experiment of stall detection was conducted in a low speed axial compressor based on LabVIEW.Experimental results show that the proposed method can meet compressors stall detection of online operating system.