Since the data are often polluted by numerous measured noise or outliers, traditional subspace discriminant analysis is difficult to extract optimal diagnostic information. To alleviate the impact of the problem, a robust principal subspace discriminant analysis algorithm for fault diagnosis is designed. On the premise of decreasing the impact of redundant information, the optimal latent features can be calculated. Specifically, in the algorithm, dual constraints of the weighted principal subspace center and $l_{2,1}$-norm are introduced into the objective function to suppress outliers and noise. Besides, considering that the current changes of the data in a dynamic process rely on past observations, merely analyzing the current data may lead to an incorrect interpretation of the mechanism model, especially in the presence of similar variable data under the two different conditions. Therefore, based on the robust principal subspace discriminant analysis, we further develop its dynamic enhanced version. The dynamic enhanced method utilizes the dynamic augmented matrix to enhance the latent features of historical data into current shifted features, so as to enlarge the difference between similar modes. Finally, the experimental results arranged on the Tennessee Eastman process and a commercial multi-phase flow process demonstrate that the proposed method has advanced diagnostic performance and satisfactory convergence speed.

由于数据经常受到大量测量噪声或异常值的污染，传统的子空间判别分析难以提取最佳诊断信息。为了减轻该问题的影响，设计了一种用于故障诊断的鲁棒主子空间判别分析算法。在减少冗余信息影响的前提下，可以计算出最优的潜在特征。具体来说，在算法中，加权主子空间中心和${升}_{2,1}$-norm 被引入目标函数以抑制异常值和噪声。此外，考虑到动态过程中数据的当前变化依赖于过去的观察，仅分析当前数据可能会导致对机理模型的错误解释，尤其是在两种不同条件下存在相似的变量数据的情况下。因此，基于稳健的主子空间判别分析，我们进一步开发了其动态增强版本。动态增强方法利用动态增强矩阵将历史数据的潜在特征增强为当前的移位特征，从而扩大相似模式之间的差异。最后，