Objective Wearable inertial measurement unit (IMU) technology and random forest (RF) algorithm were used to detect the fatigue level of long-distance walking with backpack load, and the feasibility of fatigue detection of load-bearing walking and optimal IMU combination scheme were explored. Methods Thirty healthy male college students were recruited to carry out long-distance backpack walking. Xsens MVN Link inertial motion capture system and Borg-RPE fatigue scale were used to collect kinematic data and subjective fatigue values of load-bearing walking, and fatigue was divided into 3 levels: without fatigue, moderate fatigue and severe fatigue. The original data were extracted; gait segmentation, data screening and feature extraction were carried out; and RF model was used for machine learning of sample features. Finally, the accuracy rate, precision, confusion matrix and AUC (area under the curve) were calculated to evaluate the detection effect of different IMU combinations. Results The accuracy of one right femur IMU was 82.55%, and the accuracy of five IMU combinations was 87.94%. In a combination of IMUs, at least one upper body IMU was included, and the left limb had more IMUs than the right limb. The RF model had a higher level of fatigue detection for load-bearing walking; when four IMUs were used, the AUC of 3-level fatigue was 0.99, 0.97 and 0.99, respectively. Conclusions IMU technology and RF algorithm have high accuracy and classification ability in the 3-level fatigue detection task of walking with backpack load. In practical application, it is recommended to use 1–5 IMUs, and the combination of upper body IMU and lower limb IMU configuration scheme is preferred.