LSTM-driven dynamic load identification and weld fatigue life prediction for excavator working devices

To address the issue of insufficient accuracy in acquiring dynamic load spectra for weld fatigue life prediction in excavator working devices, this study proposes an AI-based load identification technology driven by "data-physics fusion." By collecting multi-condition strain data from the client side [ 1 ], a dynamic load identification model based on Long Short-Term Memory (LSTM) networks is constructed, breaking through the limitations of traditional linear assumptions and achieving spatiotemporal correlation modeling of multi-directional dynamic loads [ 2 ] at the hinge points of the boom and arm, reducing the inversion error from18.7 % to 6.3 %. The equivalent structural stress method is employed to predict weld fatigue life, and the results show that the predicted fatigue lives of the boom's rear top plate butt weld (10,250 h) and the arm's middle support weld (11,540 h) have errors of <27 % compared to after-sales failure cracking times (9200 h and 9117 h), validating the engineering effectiveness of the method. Additionally, through dynamic measurement point optimization and data-driven modeling, the study provides a high-precision, closed-loop verifiable intelligent solution for fatigue life assessment of engineering machinery structures [ 3 ].