Prior knowledge-augmented wavelet spatial-temporal graph network and its application in anomaly detection for electric submersible pumps

Prior knowledge-augmented wavelet spatial-temporal graph network and its application in anomaly detection for electric submersible pumps. Real-time anomaly detection for electric submersible pumps is essential for ensuring the long-term safe production of petroleum. However, existing process modeling approaches do not fully capture the multidimensional spatial-temporal interactions in the data or adequately leverage prior knowledge, leading to suboptimal performance. To address these limitations, this paper introduces a prior knowledge-augmented wavelet spatial-temporal graph network (PKWST-GN). The proposed method initially develops a fusion strategy that integrates features derived from prior knowledge with those obtained from the wavelet domain via a discrete wavelet transform. A multivariate time-series correlation learning module is then employed to effectively reveal spatial interdependencies among variables. Subsequently, a temporal module constructed with dilated convolutions is utilized to extract long-term temporal features from the fused representations, while a cross-channel graph convolution based on spatial relationships encodes the spatial features of the electric submersible pumps, thereby achieving synergistic spatial-temporal learning and deep integration. Moreover, an adaptive threshold anomaly detection mechanism is incorporated into PKWST-GN to mitigate the high false alarm rates induced by perturbations. Finally, experimental validation using production data from a real oil field demonstrates that the proposed method significantly enhances anomaly detection performance.