Multi-fidelity fusion for soil classification via LSTM and multi-head self-attention CNN model

Abstract

An effective soil classification method is essential for soil stratigraphy interpretation. Traditional methods rely on classification systems in design codes or empirical formulae based on single source data, which may not be accurate or suitable for every site. Recent advancements in data fusion techniques and deep learning have shown promising results in a wide range of practical domains, which provides motivation to explore its potential application in soil classification. To that end, this paper introduces a novel machine learning architecture using boosting long-short term memory (LSTM) machine to map low fidelity cone penetration test (CPT) data to high fidelity laboratory test (LT) data which are then fused and fed into a multi-head self-attention convolutional neural network (MSCNN) for soil classification. A rigorous loss function is developed to train a LSTM-MSCNN model in an end-to-end workflow. The MSCNN classifier is benchmarked against traditional machine learning methods to demonstrate its superior outperformance, with ablation analyses exploring the related merits of the attention mechanism and the influence of attention heads on classification results. The results reveal a significant improvement in the prediction of true ground profiles using the proposed LSTM-MSCNN model over conventional methods.