A combined RF–GRNN algorithm for monitoring complex damage of bolted joints with high-level robustness

Bolted joints in aerospace, rail transportation, civil engineering, and other applications are prone to various forms of damage, such as bolt loosening and crack, due to the harsh service environments. The difficulties arising from crosstalk and nonlinear evaluation between composite damage and single damage at bolted joints in practical monitoring environments are particularly pronounced. In response to this challenge, we propose an integrated algorithm, namely the random forest-generalized regression neural network (RF–GRNN) algorithm, characterized by high-level robustness. The proposed algorithm adheres to the core concept of “specificity” and forms a multi-task prediction framework of “classification before regression.” This approach enables efficient damage assessment and mitigates the nonlinear complexity of the prediction model. To evaluate the effectiveness of the RF–GRNN algorithm, its generalization, robustness, and prediction accuracy based on the piezoelectric ultrasonic guided wave principle under 0%–10% noise (the corresponding minimum signal-to-noise ratio was 10.4 dB) interference were investigated. Additionally, we clarify the impact of feature selection and its combination method on the algorithm’s prediction performance, deciphering the intrinsic link between prediction performance and feature distribution. The results demonstrate that the presented algorithm achieves the accurate classification and quantification of multiple bolted structural states, including no damage, crack, loosening, and crack-loosening composite damage. Therefore, the RF–GRNN algorithm is an important attempt to solve the bottleneck present in existing damage monitoring techniques which fail to balance the identification and quantification of multiple damage models.