Research on NO_x concentration filling method for remote OBM data of heavy-duty vehicles based on temporal neural network

The nitrogen oxides (NO_x) concentration data collected by remote On-Board Monitoring (OBM) systems for heavy-duty diesel vehicles often suffer from severe missing issues, which hinders accurate emission assessment. Based on real-world OBM data from 50 heavy-duty trucks of different tonnage classes, this study developed a deep learning model that incorporated vehicle weight stratification and integrated a Temporal Convolutional Network (TCN) with a Bidirectional Long Short-Term Memory (BiLSTM) network to impute the missing NO_x concentration values. Results indicated that OBM feature variables during vehicle operation exhibited significant differences in their relationships with NO_x concentrations across different tonnage categories. The stratified modeling approach demonstrated that the proposed TCN-BiLSTM model outperformed conventional machine learning models in prediction accuracy and error control for small- and medium- tonnage vehicles, particularly under long-duration data gaps. For heavy-duty trucks in the 20 to 40-ton range, the TCN-BiLSTM model demonstrated particularly strong performance under long-duration missing scenarios (over 300 seconds). Compared with the structurally similar TCN-LSTM model, the RMSE of TCN-BiLSTM was reduced by 29.97% and 24.60% at the 300-second and 400-second missing intervals, respectively. However, for heavy-duty trucks with a gross weight over 40 tons, the limited sample size hindered the model's ability to fully capture their operational characteristics, resulting in relatively inferior prediction performance compared to small- and medium- tonnage vehicles. This study provides a methodological basis for improving the utilization of OBM data and the accuracy of emission monitoring for heavy-duty vehicles. Future work may focus on enhancing model adaptability to large- tonnage vehicles by expanding the dataset or incorporating transfer learning and few-shot learning techniques.