# Recurrent neural network long short term memory model to detect the pile toe using raw data of pile integrity test

## TL;DR

This paper introduces an AI system using RNN-LSTM to automatically detect pile toe locations from pile integrity test data, reducing reliance on human interpretation.

## Contribution

A novel RNN-LSTM model is proposed for generating accurate velocity reflectograms in pile integrity testing.

## Key findings

- The six-layer, 32-neuron LSTM model achieved training and validation R2 scores of 0.9126 and 0.8778.
- 84% of validation set predictions for toe location were classified as 'Good', with low 'Bad' prediction rates.
- The model demonstrated high accuracy and low mis-adoption risk compared to human-generated reflectograms.

## Abstract

This article proposes a novel approach to automatically generate velocity reflectogram of Pile Integrity Testing using a Recurrent Neural Network with Long Short-Term Memory (RNN-LSTM) model. Conventional Low-Strain Integrity Testing (LSIT) accuracy relies significantly on expert interpretation of reflected wave signals and entails subjectivity as well as efficiency limitations. The purpose of this study is to develop an artificial intelligence system capable of learning wave propagation behavior from acceleration inputs and generating reflectogram that capture pile toe locations correctly, thereby reducing dependence on human experience. The proposed technique eliminates human error and increases both the reliability and efficiency of the model. The strategy involved the collection of LSIT data from several of Egypt’s driven piles projects, followed by systematic preprocessing which converted raw acceleration signals into digitized velocity–time series. Several RNN-LSTM networks with various hidden layers and neurons were trained and optimized against performance including measures the coefficient of determination (R2), computational expense, and visual examination of reflectogram. The proposed six-layer, 32-neuron LSTM model achieved an optimum balance between accuracy and computational expense and yielded training and validation R2 of 0.9126 and 0.8778, respectively, and demonstrated satisfactory predictive generalization. Visual examinations also guaranteed the validity of the model, where “Good” predictions for toe location were up to 84% for the validation set and 89.5% for the training set, while “Fair” and “Bad” predictions had an average of only 10% and 5%, respectively. The experiments demonstrate that the RNN-LSTM model effectively mimics human-generated reflectogram with high accuracy and low mis-adoption risk. Lastly, this research describes how deep learning, namely RNN-LSTM, presents an excellent alternative to the conventional generated reflectogram, greater reliability, and reduced reliance on human experience.