Unsupervised clustering for the anomaly diagnosis of plunger lift operations

Abstract

Plunger lift is effective for dealing with liquid accumulation in gas wells. With the development of digital plunger control systems and large databases, machine learning has demonstrated significant advantages in plunger lift parameter optimization and pattern recognition. Generating high-quality labbels for massive field data, which is crucial for training machine learning models, is still a challenge due to the lack of a cost-effective labeling method. This paper presents an unsupervised clustering method to distinguish patterns for the plunger lift data. A model based on the transformer encoder is trained to identify periodic points in the continuous data. The periodic points are then used to partition the data into periodic sub-datasets with four distinct labels, including anomaly labels and three distinct operational conditions. Comparison of the effectiveness of data dimension reduction techniques is conducted, including Principal component analysis, Multidimensional scaling, statistical methods, and Autoencoder, in improving the clustering quality. Results show that the deep-neural-networks-based autoencoder achieves the highest clustering accuracy due to its ability to learn more compact representations through joint optimization of clustering loss and reconstruction loss. Moreover, the cyclic-feature-based algorithm is found to outperform the sliding window to obtain the input data for clustering.