Abstract
Pipe network sludge is a complex pollutant aggregate deposited during long-term operation of urban sewage pipelines, and a key target for pollution control in environmental monitoring systems. Accurate source classification is critical for treatment optimization, pollution tracing, and resource recovery. Traditional methods have drawbacks like long processing time and low efficiency. Near-infrared spectroscopy (NIR) offers a new approach but faces spectral redundancy, limited samples, and biased features. This paper proposes CABNas-nir, a deep neural network under the neural architecture search (NAS) framework, integrating competitive adaptive reweighted sampling (CARS), baseline drift augmentation, and active learning (AL). It selects key spectral features via CARS to remove redundancy, uses baseline drift to generate augmented samples for small-sample issues, employs AL with K-means to select high-value samples, and constructs an optimal convolutional neural network(CNN)+long short-term memory(LSTM) model via NAS. Experiments show 92.86% accuracy, 14.29% higher than support vector machine (SVM,78.57%) and 35.72% higher than that of extreme gradient boosting (XGBoost,57.14%). SHapley Additive exPlanations (SHAP) analysis shows high-contribution spectra in 1400-1700 nm, with 1600-1700 nm significant. This algorithm significantly enhances the robustness of identifying the sources of pipe network sludge, laying a research foundation for the rapid and accurate identification of pipe network sludge.