Research on Anomaly Detection Method of Non-Standard Machinery Production Process Based on Semi-Supervised Learning

Abstract

The non-standard machinery refers to customized machinery produced to meet specific customer demands. The mainstream research direction in data stream anomaly detection focuses on deep learning, which involves learning data distribution through a large amount of training data. However, non-standard machinery equipment has the characteristics of a small production scale and sparse samples, making it difficult to obtain sufficient annotated training sets. This inadequacy in training data results in the model not learning enough, thereby rendering it unable to effectively detect abnormal events. In this paper, we propose a semi-supervised learning (SSL) based anomaly detection method. We employ a hybrid C-LSTM network based on the self-attention mechanism as an abnormality prediction model, where the convolutional neural network (CNN) and long short-term memory network (LSTM) extract spatiotemporal features of industrial data streams. The self-attention mechanism calculates the relationship weights between different positions in the input data, capturing long-term dependencies in time series data to fully learn data distribution. To improve the training effectiveness of the prediction model, we use an updating algorithm based on weighted fuzzy rough set (WFDA) to update the prediction model in a reverse manner. This algorithm can classify data streams in real-time, compare the classification results of the prediction model, and retrain unreliable data. The experimental results show that our proposed method achieves an F1 score of 0.955 and a recall value of 0.957 on a real-world data set, which is a 4.1% improvement in F1 score and a 6.4%improvement in recall compared to similar anomaly detection algorithms that do not use our proposed method.