Abstract:
Rapid development of machine learning algorithms provides important analysis methods for safety prediction and early warning of deep foundation pits. In this study, the concept of safety levels of deep foundations is proposed to indicate the safety condition of foundation systems. Different machine learning methods, including BP (Back Propagation) neural network, GA-BP (Genetic Algorithm-Back Propagation) neural network, NARX (Nonlinear Autoregressive Exogenous Model) dynamic neural network, and Elman neural network method, were used to train and establish the mathematical prediction models of the monitored axial force, settlement, and support structure displacement of deep foundation pits within the past year. A comparative study shows that the average maximum error was less than 5.5%, and the obtained error of the GA-BP based neural network was the smallest among all prediction methods. The data difference between measured data and the related predictions was further analyzed and found to satisfy a normal distribution. This data difference can be considered a continuous random variable in a continuous normal distribution. The level of significance of the data difference between the measured and predicted data of three key parameters—including axial force, settlement, and support structure displacement—was analyzed in terms of both normal distribution and Euclidean distance methods. A new method was proposed to evaluate the safety level of deep foundation pits and was further verified using field data of a real foundation pit. This method is characterized by the advantages of ease of use for engineers, real-time prediction, optimized prediction methods, high accuracy prediction, and hence offers a critical method for the evaluation of safety levels of foundation pits.