The infiltration of chloride ions corrodes steel reinforcement and causes concrete to crack, reducing the overall bearing capacity and compromising the structure's service life and safety. Timely acquisition of chloride ion distribution is essential for accurately assessing the performance of concrete structures. We present a method for detecting the distribution of chloride ion penetration based on hyperspectral images and long short-term memory (LSTM). First, hyperspectral images of the chloride-ion-permeated split surfaces of the concrete are obtained, and a total of 1000 average spectral data points across 5 categories are collected. The Savitzky-Golay (S-G) data preprocessing algorithm is then applied to reduce spectral noise. Then, we employed a joint Competitive Adaptive Re-weighted Sampling-Principal Component Analysis (CARS-PCA) dimensionality reduction method to identify spectral feature bands highly correlated with concrete chloride distribution. Finally, the LSTM method is utilized to achieve precise classification of the various components of concrete and to accurately identify the distribution of chloride ions. The CARS-PCA dimensionality reduction model, when integrated with an LSTM network, yields superior performance models with an equivalent number of features. This combination achieves a classification accuracy of 93%, reflecting an average improvement of 4.71% over other classification algorithms.
Read full abstract