太阳集团网tyc9728(百度)有限公司

During oil and gas drilling, the accuracy of intelligent early kick recognition models relies on their detection of abnormal variation trends in monitoring parameters. These variation trends are contained in the data sequences of monitoring parameters corresponding to early kick samples. Therefore, accurate identification of the abnormal variation trends of monitoring parameters in these samples is of great importance for early kick monitoring. This understanding can serve as a prior knowledge to guide the learning process of intelligent models. Since intelligent models use loss values calculated from their loss functions to determine which samples should be focused on during the model training process, the prior knowledge can be used to modify their loss functions. First, a knowledge-guided attention allocation network is proposed, which allocates attention weights to monitoring parameters and enables the intelligent model to pay greater attention to these critical ones. Then, a knowledge-guided bidirectional long short-term memory (KG-Bi-LSTM) model was established by embedding prior knowledge into its binary cross-entropy loss function. Firstly, in order to reduce the model's attention on the most easily identifiable samples corresponding to normal drilling, the square of the model's prediction error for each normal sample is used as an attenuation factor for loss values corresponding to normal samples. In this way, the model can learn more from kick samples. Secondly, to increase the model's focus on learning the variation trends of monitoring parameters corresponding to the early kicks, a monotonically decreasing exponential function is employed to weight the loss term for kick samples. This approach assigns greater weights to samples from the early kicks, thereby preventing underfitting by emphasizing key trend recognition. Subsequently, to prevent the model from overfitting due to the weighting of loss values, the sum of squares of the network weights is added to the loss function as a constraint term. Thus, a balance between model overfitting and underfitting is achieved through knowledge guidance and weight constraint. Thereafter, a knowledge-guided loss function is obtained by combining the above weighted loss terms and the constraint term. Finally, this loss function is applied to the Bi-LSTM-based kick recognition model to produce a KG-Bi-LSTM version. Experiments were conducted using 4,599 samples from field data, the results show that compared with the prevailing Bi-LSTM model, the proposed KG-Bi-LSTM has a 5.28% increase in recognition accuracy and an earlier alarm time of nearly 2 min.