The Concept of Using LSTM to Detect Moisture in Brick Walls by Means of Electrical Impedance Tomography

Abstract

This paper refers to an original concept of tomographic measurement of brick wall humidity using an algorithm based on long short-term memory (LSTM) neural networks. The measurement vector was treated as a data sequence with a single time step in the presented study. This approach enabled the use of an algorithm utilising a recurrent deep neural network of the LSTM type as a system for converting the measurement vector into output images. A prototype electrical impedance tomograph was used in the research. The LSTM network, which is often employed for time series classification, was used to tackle the inverse problem. The task of the LSTM network was to convert 448 voltage measurements into spatial images of a selected section of a historical building’s brick wall. The 3D tomographic image mesh consisted of 11,297 finite elements. A novelty is using the measurement vector as a single time step sequence consisting of 448 features (channels). Through the appropriate selection of network parameters and the training algorithm, it was possible to obtain an LSTM network that reconstructs images of damp brick walls with high accuracy. Additionally, the reconstruction times are very short.

Application of Convolutional Neural Networks in Wall Moisture Identification by Eit Method

“…In order to verify the effectiveness of CNN in the problem of identifying moisture inside the walls of buildings using electri-cal impedance tomography (EIT) [3,4], a set of simulation cases was generated. The data set included 40,000 measurement vectors (inputs) and pattern images (outputs).…”