Traditional control charts and machine learning methods have been used to detect equipment failures in semiconductor processes. However, detecting failures and identifying their root causes can be challenging because of the complexity of the process and structural characteristics of the equipment. Moreover, the anomaly sections for each part can be imbalanced, which can hinder classification performance. In this study, we propose a method to detect failures using the correlation of variables and data augmentation. The proposed method consists of five steps: (1) conversion of the multivariate time series data of the equipment into signature matrices, (2) detection of anomalies using a convolutional autoencoder, (3) augmenting the number of residual matrices of minority classes (4) learning convolutional neural network that use the residual matrixes of definite abnormal sections, and (5) application of Grad-CAM to interpret the convolutional neural network. We demonstrate the effectiveness and applicability of the proposed method using real-world multivariate time series data obtained from ashing process equipment in semiconductor manufacturing.