Improved stochastic configuration network for bridge damage and anomaly detection using long-term monitoring data

The Stochastic Configuration Network (SCN) is a powerful incremental learning algorithm that dynamically generates network structures during training. However, as a fully connected neural network, it is not adept at capturing the internal dynamic changes of monitoring data and suffers from node redundancy. To address the inadequacy of SCN in handling multi-sensor monitoring data, this paper proposes a feature extraction method called Mean of Positive Values (MPV) to randomly extract the intrinsic features of monitoring data, thereby reconfiguring the original SCN. This improved SCN based on random convolution is named SCN based on Improved Random Convolution (IRC-SCN). Furthermore, to enhance the efficiency of SCN, this study introduces a Random Node Removal based on Importance Ranking (RNR-IR) algorithm. The proposed methods are evaluated on two bridge monitoring datasets for damage identification and anomaly detection, demonstrating their effectiveness. The model based on MPV achieves an accuracy increase of approximately 1% compared to the comparative methods on the test set. Unlike traditional node deletion algorithms, RNR-IR can improve the performance of model by approximately 2% with the removal of around 10% of neurons.