提出了一种基于并行重构堆叠自编码的故障诊断方法。该方法在常规堆叠自编码器基础上引入预设故障方向，采用梯度下降法在所有可能的预设故障方向上进行数据重构，通过比较重构后的平方预测误差（SPE）来确定最佳故障方向和故障幅值，从而抑制残差污染。针对大规模复杂系统的高维特点，进一步通过并行重构方法来提高数据重构效率，减少计算时间，满足在线诊断要求。采用数值算例和工程算例来验证所提方法的有效性。结果表明：该方法对于单参数简单故障和多参数复杂故障都有很好的诊断效果，与常规堆叠自编码方法相比，大大降低了故障误诊率，提高了诊断准确性。

This paper proposes a new fault diagnosis method of parallel reconstructionbased stacked autoencoder(PRBSAE). In order to suppress residual pollution, this method introduces preset fault directions on the basis of conventional stacked autoencoders. The gradient descent method is used to reconstruct data in all possible preset fault directions, and the optimal fault direction and fault amplitude are determined by comparing the square prediction error (SPE) after reconstruction. In view of the highdimensional characteristics of largescale complex systems, the parallel reconstruction method is further used to improve the data correction efficiency, reduce the calculation time and meet the requirements of online diagnosis.The effectiveness of the proposed method is evaluated on a numerical example and an industrial example. The results show that the method proposed in this paper has a good diagnostic affect both on simple singleparameter faults and complex multiparameter faults. Compared with the conventional stacked autoencoders method, this method greatly reduces the fault diagnosis rate and improves the diagnosis accuracy.

TP206.3

国家自然科学基金面上项目（51976031）