US 12,468,955 B1
Meta-learning strategy based method for early warning of attenuation of bearing capacity of soft soil pile foundation
Wei Qin, Wenzhou (CN); Xin Ye, Wenzhou (CN); Yuqi Pan, Wenzhou (CN); Guoliang Dai, Wenzhou (CN); and Dong Wang, Wenzhou (CN)
Filed by Wenzhou University, Wenzhou (CN)
Filed on Jun. 12, 2025, as Appl. No. 19/236,772.
Claims priority of application No. 202411833181.6 (CN), filed on Dec. 13, 2024.
Int. Cl. G06F 30/20 (2020.01); G01N 3/08 (2006.01); G06N 3/044 (2023.01); G06N 3/0985 (2023.01)
CPC G06N 3/0985 (2023.01) [G01N 3/08 (2013.01); G06N 3/044 (2023.01)] 8 Claims
OG exemplary drawing
 
1. A meta-learning strategy based method for early warning of attenuation of a bearing capacity of a soft soil pile foundation, comprising:
S1, building a loop loading test device,
performing a loop loading test, and obtaining a horizontal bearing capacity signal at the top of the model pile under a loop horizontal load, wherein a prototype pile is equivalent to a scaled model pile based on a second similarity theory in the loop loading test and then is subjected to the loop loading test;
S2, setting up a data set, wherein the data set comprises a training data set, a target data set, and a pre-training data set for a base model;
S3, building and training the base model, and obtaining the base model by training a bidirectional echo state network;
S4, introducing a meta-learning strategy into the base model obtained in S3, inputting the pre-training data set for pre-training, and obtaining a pre-trained model, wherein the meta-learning strategy is a Reptile meta-learning algorithm, inner loop training and outer loop training are performed in Reptile meta learning, the pre-trained model is optimized, and parameters of the pre-trained model are updated; and
the pre-training data set is used for pre-training as follows:
S41, performing copying to obtain two identical models ϕ and ϕ′;
S42, randomly sampling one pre-training set in S2, making the model ϕ′ reconstruct a sequence by using the bidirectional echo state network for the pre-training set, and completing one time of inner loop training of meta learning;
S43, updating parameters of the model ϕ according to a formula ϕ=ϕ+β*(ϕ′−ϕ) on the basis of a training result of ϕ′, and completing one outer loop of the meta learning, wherein β is a meta-learning rate; and
S44, repeating S42 and S43, sequentially selecting one pre-training set for pre-training, and completing pre-training of the model after set inner loops and outer loops are completed; and
S5, inputting the target data set into the pre-trained model for training, and obtaining a model prediction result.