Abstract: Rock bolt support is one of the most widely used methods for reinforcing surrounding rock in mines, tunnels, and underground engineering. Accurate monitoring of bolt force is critical for evaluating rock mass stability and optimizing support parameters. Existing monitoring techniques typically rely on external or embedded sensors, which suffer from limitations such as poor adaptability, high cost, susceptibility to interference, and complex installation. These constraints hinder their application in full-section, large-scale, and multi-bolt synchronous measurements in underground roadways. To address these challenges, this study proposes a non-contact and non-destructive method for measuring the forces of rock bolt groups based on 3D laser scanning and deep learning. A mathematical relationship between the 3D deformation field of the bearing plate and the axial force of the bolt is first derived. High-precision point cloud data under various working conditions are acquired through pull-out tests combined with 3D laser scanning. A normalization framework for computing the deformation field is established. A total of 36 representative working conditions are designed, yielding over 1 000 datasets of deformation fields and corresponding axial forces. A convolutional neural network (CNN) model is then trained to map the deformation field to bolt axial force, achieving an average prediction error within ±19.25 kN, demonstrating high accuracy and strong generalization capability. Based on this method, two engineering application systems are developed: A portable single-bolt force measurement system that uses a handheld laser scanner to obtain point cloud data of the bearing plate and quickly predicts the bolt force using the pre-trained model; a tripod mounted full-roadway bolt group force measurement system that acquires full cross-sectional point cloud data with a high resolution laser scanner, automatically identifies and segments the bearing plates using an improved PointNet++ algorithm, and then estimates the axial forces of multiple bolts in batch using the trained deep learning model. The study demonstrates that the integration of 3D laser scanning and deep learning overcomes the limitations of traditional point-based measurement methods, enabling large-scale, multi-target, and high-precision non-contact monitoring. This technology holds significant potential for evaluating bolt installation quality and surrounding rock stability, offering a new approach toward intelligent monitoring of underground support systems.