A hybrid autoencoder framework integrating dual linear–nonlinear feature extraction for multimodal industrial process fault detection

Although autoencoders are adept at reducing data dimensionality and preserving important feature information, they exhibit insufficient extraction of the inherent linear and nonlinear features in multimodal industrial process data. Their data reconstruction ability is limited and highly dependent on the Gaussian distribution of the data. To address these issues, this paper proposes a hybrid autoencoder with dual feature extraction (HADFE), where the dual feature extraction refers to the joint learning of linear and nonlinear features. First, the mutual local neighbourhood standardization algorithm is employed to standardize the training data, transforming the multimodal data into unimodal data while retaining the important features of the original data. Second, a linear feature extraction mechanism from principal component analysis is incorporated into the autoencoder. This mechanism extracts linear features from the data, and based on these extracted features, the encoding weights of the encoder and the decoding weights of the decoder are constructed. An orthogonal penalty term is introduced to ensure the orthogonality of the weights, thereby enhancing the autoencoder's ability to extract linear features from the data. Then, both an regularization term and the nonlinear activation function LeakyReLU are introduced into the autoencoder's loss function to prevent model overfitting and improve the autoencoder's learning ability for nonlinear features. Finally, two modes of Tennessee Eastman (TE) process data are used as experimental data to validate the performance of HADFE. The experimental results show that the HADFE model significantly outperforms traditional models in fault detection.