Top

Improved ConvNeXtV2 CBAM Model for Formwork Collapse Risk Prediction Using Load Sensing and Construction Progress Data

Authors

Zulfikar Osman
Faculty of Data Science and Artificial Intelligence, Sunway University, Petaling Jaya, 47500, Malaysia
Priyanka Gandhi
Faculty of Data Science and Artificial Intelligence, Sunway University, Petaling Jaya, 47500, Malaysia
Syafiq Abdullah
Faculty of Data Science and Artificial Intelligence, Sunway University, Petaling Jaya, 47500, Malaysia

Files

Abstract

Formwork collapse-risk forecasting requires alignment of changing support loads with construction progress and an explicit future capacity-reference target. This study evaluates a compact progress-aware ConvNeXtV2-CBAM encoder on a numerical construction-sequence reference. Six support-load measurements, their increments, cumulative progress and recent progress increments form a four-channel, six-support, sixteen-interval historical map at each prediction endpoint. A ConvNeXtV2-style convolutional representation with global response normalization is refined by sequential channel and spatial attention. Shared features generate binary probabilities for capacity-reference crossings within the next two, four and six intervals, alongside auxiliary estimates of maximum future utilization. The reference contains 180 independent construction episodes and 3,960 historical windows; complete episodes are assigned to training, model selection or testing. The combined familiar-rate and withheld-rate test partitions contain 1,848 windows. Positive-class F1 reaches 0.8993, 0.9010 and 0.8808 at the respective horizons, with corresponding accuracies of 96.86%, 96.16% and 94.48%. Under a six- interval warning region, threshold 0.50 and two-endpoint persistence, the model detects 50 of 61 eligible reference crossings, or 81.97%, with a conditional mean advance of 4.34 intervals among detected events. The median network- forward time is 1.075 ms for 8,441 trainable parameters. Component comparisons show that an eight-interval history obtains six-interval F1 0.9011 and progress-channel removal obtains 0.8861, both above the full configuration's 0.8808, whereas removal of auxiliary utilization supervision yields 0.8604. These results characterize forecast, warning and efficiency trade-offs for the declared numerical load–capacity process and motivate validation on independently monitored formwork operations.

Details