A novel method of detection of noisy samples in high-dimensional sequential data considering both attribute and label noise

• The log-likelihood of samples is modeled considering time dependencies in input data. • The proposed method can simultaneously detect both attribute and label noise. • The robustness of the detection method is enhanced by an iterative process. • The effectiveness of the proposed method is verified using two case studies. The high-dimensional sequential data available across various industrial scenarios may be contaminated with both attribute and label noise, hindering the establishment of accurate deep learning-based prediction models. The existing noise detection methods can only detect one type of noise. Conversely, in this article, a novel noisy samples detection method is proposed to detect both types of noise simultaneously through generative learning. An enhanced variational recurrent prediction model (EVRPM) is proposed to model the log-likelihood of samples, which incorporates a label predictor and an auxiliary task into the variational recurrent neural network. Moreover, an iterative detection process is adopted to refine EVRPM training and enhance noisy sample detection, which is particularly beneficial for low-quality datasets. A prediction model with higher prediction accuracy can be obtained using the refined dataset. The effectiveness and superiority of the proposed method are verified using both public and real experimental datasets.