Concrete pumping technology is widely applied in major construction projects. However, under long-distance or complex conditions, concrete transport often exhibits strongly non-steady characteristics due to the coupled factors like material rheology and pipeline topologies. This instability easily induces abnormal resistance surges, evolving into severe blockages. Conventional monitoring relies on post-event management or static pressure thresholds, triggering alarms only when pressure breaches a safety limit. Essentially, concrete blockage is a progressive dynamic accumulation of rheological resistance, leaving a critical "pressure accumulation" intermediate state. Static thresholds ignore this, causing severe pre-warning lag. AI methods show a potential, while intelligent perception faces three challenges, i.e., the binary classification defect overlooking the transitional state, the "majority class trap" from extreme class imbalance, and the contradiction between prediction accuracy and physical interpretability. To address these bottlenecks, this study was to propose a novel intelligent evaluation framework for non-steady concrete transport, deeply integrating rheological mechanisms with advanced machine learning architectures.

The research relied on the Zhenglong Hongshui River Bridge project, China. An embedded monitoring system with 600 Hz dynamic pressure sensors was seamlessly welded to the straight pipeline's inner wall near the pump outlet to acquire high-quality rheological resistance signals. During feature engineering, mechanical noises from valve reversing were eliminated, retaining only effective pressure-holding signals during main cylinder propulsion. Since deriving the exact friction coefficient via true flow velocity is a challenge in field engineering, macroscopic proxy variables were utilized. Assuming constant propulsion speed, the slope of the pressure curve was extracted to represent the transient growth rate of rheological friction. Standard deviation and range were extracted to quantify slip-layer stability and flow turbulence. To address extreme class imbalance in the 691-sample dataset (78.3% normal), the Synthetic Minority Over-sampling Technique (SMOTE) was applied to reconstruct the feature space, eliminating training bias. Subsequently, four ensemble algorithms (i.e., RF, XGBoost, LightGBM, GBDT) optimized by the Tree-structured Parzen Estimator (TPE) were used as base learners. Finally, a dynamic adaptive ensemble fusion architecture calculated sample-specific dynamic weights based on prediction confidence, replacing conventional static voting mechanisms.

The extracted time-domain features effectively map the dynamic evolution of in-tube rheological resistance. After SMOTE reconstruction and TPE optimization, the base models' potential to capture minority class abnormal conditions is maximized. During the comparative evaluation of the ensemble models, the study addresses the "accuracy paradox" inherent in highly imbalanced datasets. Conventional soft voting and hard voting models achieve an overall accuracy of approximately 0.90 simply because of the overwhelming majority class, dangerously masking high misdiagnosis rates for critical anomalies. The overall accuracy improves to 0.92 via implementing the dynamic adaptive fusion strategy. More importantly, the model substantially suppresses the omission rates for high-risk conditions. The recall rate for the "pressure accumulation" state reaches 91.43%, representing a 10.3% relative improvement over the soft voting method (i.e., 82.86%). For the most hazardous "blockage" state, the recall reaches 88.00%, outperforming the soft voting method (i.e., 76.00%) with a remarkable relative improvement of 15.8%. Furthermore, a SHapley Additive exPlanations (SHAP) visualization framework is introduced to verify the model's decision logic globally and locally. The global feature importance ranking confirms that the pressure curve slope is an absolute dominant feature triggering blockage predictions, followed by range and standard deviation. The SHAP beeswarm summary plots reveal that a sharp increase in these three features heavily biases the model toward a "blockage" decision. Conversely, smaller feature values drive the model toward a "normal" prediction. Local waterfall plots for individual samples further detail the positive and negative contributions of each feature to a specific prediction. This data-driven logic perfectly coincides with the physical reality that pipeline blockages instantly cause a sharp surge in rheological friction and severe dispersion in pressure values due to slip layer disruption. SHAP effectively explains the statistical feature attribution, and its strong alignment with engineering experience proves the model's physical rationality. Verifying strict causality requires a further extensive field validation across multiple complex scenarios.

Time-domain physical features validated the ability to characterize in-tube non-steady transport. The slope, standard deviation, and range extracted from high-frequency pressure sequences could serve as effective macroscopic proxies for concrete fluid resistance and slip-layer stability. The combination of SMOTE feature space reconstruction and TPE Bayesian hyperparameter tuning effectively overcome the small-sample fault warning dilemma under extreme class imbalance, eliminating training bias toward majority classes. The dynamic adaptive ensemble architecture demonstrated a core advantage in suppressing high-risk omission rates. It leveraged a minor overall precision increment to drastically reduce crucial fault omission risks via tolerating a minimal sacrifice in majority class accuracy, delivering an ample forward-intervention window. The SHAP attribution analysis achieved a white-box deconstruction of the ensemble decision logic. The alignment between the algorithm's reliance on slope surged and the true physical laws of rheological friction endowed the model with clear physical rationality, laying a solid theoretical foundation for the trustworthy application of AI technologies in intelligent civil engineering construction.