<p>Servo steel struts are an emerging support technique for deep excavations, capable of adjusting axial forces to control excavation-induced wall deflections during construction. Their effectiveness depends on the appropriate adjustment of axial force values. This paper proposes an active axial force adjustment method for deep excavations supported by multi-layer servo steel struts. The method first enhances deflection predictions by updating soil parameters using stage-by-stage monitoring data, and then adaptively adjusts axial forces based on the refined deflection predictions. The Bayesian updating of soil parameters and the adjustment of servo axial forces are performed in a staged manner, resulting in axial force schemes that align with the actual field responses at each excavation stage. To improve computational efficiency in Bayesian updating, Bidirectional Long Short-Term Memory (BiLSTM) neural networks are developed as surrogate forward models, capturing the relationships among soil parameters, axial forces of multi-layer servo steel struts, and deflections along depth. The effectiveness of the axial force schemes is evaluated using the failure probability of the serviceability limit state in subsequent excavation stages. A numerical case study and a real-world deep excavation project are used to demonstrate the proposed method. The results indicate that the proposed method can accurately predict the excavation-induced wall deflections and provide axial force settings of servo steel struts that maintain deflections within target limits. For excavations supported by multi-layer servo struts, the axial forces near the excavation face should be carefully determined.</p>