Zero-shot pipeline fault detection using percussion method and multi-attribute learning model

Abstract

In recent years, the machine learning (ML)-based percussion method has gained considerable attention as a cost-effective and user-friendly non-destructive testing (NDT) technique. However, traditional ML classification methods fail to identify previously unseen fault levels that are not included in the training dataset, thereby limiting their practical applicability. This paper proposes a zero-shot pipeline fault detection method based on a multi-attribute learning model to identify unseen fault classes without requiring their direct signal samples during training. In this method, each fault category is represented by a six-dimensional attribute vector that characterizes its unique properties. During the attribute learning phase, a multi-attribute learning model is constructed by integrating a one-dimensional convolutional neural network (1D-CNN) with a bidirectional long short-term memory network (BiLSTM) to predict the fault attributes. Fault recognition is subsequently achieved using a Euclidean distance-based classifier, which categorizes the predicted attribute vectors based on their similarity to predefined attribute representations. The results demonstrate that when the test set originates from previously unseen pipelines, the proposed method significantly outperforms other approaches in terms of classification performance, exhibiting superior adaptability and robustness. Importantly, it effectively identifies unseen fault severity, overcoming the limitations of traditional methods. In conclusion, the proposed method offers an innovative solution to the problem of data scarcity in fault diagnosis, with promising potential for application in complex industrial environments.