基于滑动概率神经网络的早期故障诊断

Incipient Fault Diagnosis Based on Moving Probabilistic Neural Network

针对机电设备早期故障难以识别的问题,提出了一种动态的概率密度估计方法———滑动概率神经网络,用以跟踪分析测量信号的概率密度变化过程,及时发现早期故障.该网络以固定不变的抽样集作为第一层,动态滑动的测量信号作为样本层,通过求和层得到抽样集的条件概率密度估计,将样本层内测量信号的概率密度动态地投影到统一的抽样集上.将网络分解成以测量值为中心的子网络,来实现网络的递归运算,并且利用高斯函数的快速衰减特性或使用分段线性函数近似高斯函数,从而提高了网络的计算实时性.通过压缩机喘振过程数据的应用实例,表明该方法能够有效识别故障的早期征兆.

A new probability density tracing analysis method named moving probability neural networks （MPNN） is proposed to identify the incipient fault of electromechanical equipment. The MPNN is a three-layer network： the first layer consists of invariable sample set, dynamic signal is moved into the second layer, and condition probability density estimation of sample set is the output of the third layer. So the probability density of signal is continuously projected to the uniform sample set. The recursive algorithm is achieved by dividing the network into subnets. Using the attenuation characteristic of Gaussian function or the piecewise-linear approximation for Gaussian function, the computational load of MPNN is reduced further. Finally, the surge process data of centrifugal compressor is analyzed via this network to verify the effectiveness for diagnosis of the incipient faults.