| Drilling engineering is a type of complex geological projects which includes exploration,drilling,well cementing,and other processes.The drilling process is one of the most important processes of drilling engineering,which requires many sensors to collect and monitor on the entire process.During the drilling process,these data reflect key information such as the workload of the drilling machine and rock structures,which can help operators control the on-site situation,judge the safety of the underground conditions and assist in prediction,localization and quantitative diagnosis of downhole faults.Therefore,sensor data is crucial in the drilling process and plays an irreplaceable role in the drilling process.With the development of sensor technology,various machine learning and data analysis methods have also been widely used in this field.However,the period of the drilling process is large,and the on-site situation is complex.Sensor failures,signal interference,environmental factors,data transmission interruptions,and improper human operations may result in the missing of sensor data.Moreover,the drilling process requires real-time monitoring and feedback,and the missing parts in data needs to be analyzed and demonstrated by professionals,which requires a lot of time and effort.However,few studies currently focus on the impact of missing data on drilling process analysis and prediction.The research is needed because missing data may lead to a decrease in model accuracy,thereby affecting the efficiency and safety of the drilling process.To address the problem above,we propose advanced missing data imputation algorithms for the drilling process,which consist two modules: the missing time series value imputation module and the multi-task learning module.The disappeared time series value imputation module includes two respective submodules for aggregation of temporal and spatial features.The past temporal feature state is learned by the temporal aggregation submodule.In contrast,the dynamic aggregation submodule integrates the current time point input and the past prediction results to aggregate the current temporal representation.The multi-task learning module describes the common method for classification and prediction tasks.In the time series value imputation model,the imputation loss of the missing time series value imputation module and the classification and prediction losses of the multi-task learning module are combined as the total loss of,which learns a universal hidden features representation by the end-to-end learning process of the drilling time series.During the forward propagation,well-learned hidden representation in missing imputation module can improve the accuracy of the downstream classification and prediction tasks.In the backpropagation process,the classification and prediction losses will be passed to the process of the missing imputation module.Instead of being limited to the accuracy of a single task,the multiple loss common optimization model actively explores the space of other tasks and learns a broader knowledge of the timing,so more efficient and robust results are performed on the other data sets.Two time series feature datasets in actual drilling process are introduced to verify the effectiveness of the algorithm.The proposed algorithm performs better than the existing methods on the two tasks.The effect on downstream tasks shows that the imputation module improves the representation ability of features.At the same time,the introduction of the multi-task learning module makes the model more generalized.The experiments detail the effects of different missing rates on the models and analyzes the specific impact of actual missing data.Parameter analysis experiments show the results under different parameters.Furthermore,this paper conducts ablation and visualization experiments on the dataset,and the results of these experiments verify the effectiveness and superiority of the model. |
