Multi-Weighted Partial Domain Adaptation for Sucker Rod Pump Fault Diagnosis Using Motor Power Data

Abstract

Motor power curves (MPCs) have received great attention for use in diagnosing the working conditions of sucker rod pumping systems (SRPSs) because of their advantages in accessibility and real-time performance. However, existing MPC-based approaches mostly need a rigorous assumption that the MPC instances of different working conditions are sufficient, which does not hold in industrial scenarios. To this end, this paper proposes an unsupervised fault diagnosis methodology to leverage readily available dynamometer cards (DCs) to diagnose collected unlabeled MPCs. Firstly, a mathematical model of the SRPS is presented to convert actual DCs to MPCs. Secondly, a novel diagnostic methodology based on adversarial domain adaptation is proposed for the problem of data distribution discrepancy across the collected and converted MPCs. Specifically, the collected unlabeled MPCs may only cover a subset of the working conditions of the abundant DCs, which will easily cause negative transfer and lead to dramatic performance degradation. This proposed methodology employs class-level and distribution-level weighting strategies so as to guide the network to focus on the instances from shared categories and down-weight the outlier ones. Validation experiments are performed to evaluate the mathematical model and the diagnostic methodology with a set of actual MPCs collected by a self-developed device. The experimental result indicates that the accuracy of the proposed algorithm can reach 99.3% in diagnosing actual MPCs when only labeled DCs and unlabeled actual MPCs are used.