关键词: 盾构隧道, 接头抗弯刚度, 接缝面影响系数, 深度神经网络, 足尺试验

Abstract: Existing methods for calculating segment-joint bending stiffness fail to account for the effects of complex joint-interface configurations, hindering quantitative evaluation and resulting in experience-based design practices. To address this limitation, an intelligent prediction method for quantifying joint stiffness that explicitly incorporates the influence of complex joint-interface parameters was proposed. First, a refined numerical model capturing complex interface configurations was established, and the effects of different interface parameters on joint stiffness were systematically investigated. Subsequently, the underlying influence mechanisms were analyzed alongside other joint parameters, establishing a stiffness-calculation method based on an interface influence coefficient. Furthermore, the effects of different parameters on this coefficient were discussed, and a deep learning method incorporating an attention block was developed to predict the coefficient across parameter spaces. Finally, full-scale mechanical tests on segment joints with three representative thicknesses (550, 600, and 650 mm) were conducted to validate the framework. The results indicate the following: (1) Interface parameters substantially influence joint stiffness, with maximum variations reaching 30%－40% across different interface configurations. (2) Within the ranges of other joint-parameter combinations standard to practical underground engineering, the effects of interface parameters and general structural variables on joint stiffness are independent, allowing for the simplification of interface effects as a standalone quantitative coefficient in joint-stiffness calculations. (3) Joint axial force is identified as the most influential factor governing the interface influence coefficient; the developed deep neural network demonstrates good training and predictive performance, achieving a maximum prediction error of 5.6% on the validation set. (4) The proposed framework enables accurate predictions of joint stiffness under combined variations in joint axial force and interface parameters, with relative errors remaining within 9.0% when benchmarked against full-scale experimental results.

Key words: shield tunnel, joint bending stiffness, influence coefficient of joint interface, deep neural network, full-scale test