An accurate and timely prediction of falls in a complex environment is vital for population groups such as workers, the elderly, and power-assisted exoskeleton wearers. Enhancing the universality of fall warning methods has been regarded as one of the primary challenges in the field of precise anomaly detection and fall prediction. To address this issue, a gait abnormality detection and fall warning method is proposed in this paper. First, a wearable data acquisition system integrated with inertial measurement units and capacitive plantar pressure sensors is used to obtain real data on feet. Second, a human musculoskeletal model is built in AnyBody software to obtain simulation data on feet. By comparison, the effectiveness of the simulation model is verified and the characteristics of abnormal gait are determined. Third, a backpropagation network(BP) is cleverly combined with the hidden Markov model(HMM). The cooperation of neural network and probabilistic model is employed to detect the abnormal gait sequence before falling and make a first-level fall warning. Then, a mapping model between the real and simulation plantar pressures is constructed using a multiple linear regression algorithm to weaken the difference of stability thresholds of different people and conduct second-level fall warning. Finally, two common fall patterns, tripping and slipping, are used to test the proposed fall waring method. The average sensitivity, specificity, and accuracy of the gait anomaly detection and stability judgment are used as evaluation metrics. The results indicate that the proposed method achieves average sensitivity, specificity, and accuracy of 97%, 100%, and 98.5%, and of 100%, 96%, and 98.25%, on tripping and slipping patterns, respectively. Moreover, the proposed method could assess pedestrian stability and provide fall warnings of more than 300ms before a fall occurs.