Use of the double-stage LSTM network in electrical tomography for 3D wall moisture imaging

This paper deals with a new two-step algorithmic method to improve the accuracy of imaging moisture in building walls using electrical impedance tomography (EIT). The problem of assessing dampness in buildings is important both from the point of view of the national cultural heritage, economics, safety, and aesthetics of real estate and the health of people inside buildings. The main impediment to EIT development is solving the ill-posed and inverse problem that makes obtaining reliable images of high-resolution moisture distribution difficult. To resolve this problem, we propose the new double-stage neural network approach. First, the method's originality is determined by using the second neural network, which improves the images obtained thanks to the first network. Using the first network, training images are generated, constituting inputs for the second network, while both neural networks are trained on a similar set of pattern output images. The approach to input mea-surements and images is also original, which, thanks to the conversion into vectors (sequences), made it possible to use the LSTM (long short-term memory) network. The research proved the effectiveness of the new method.

Automated summary

This paper proposes a new two-step algorithmic method to improve the accuracy of moisture imaging in building walls using electrical impedance tomography (EIT). The problem of assessing dampness in buildings is important for various reasons, including cultural heritage, economics, safety, real estate aesthetics, and the health of occupants. The main challenge in EIT development is solving the ill-posed and inverse problem of obtaining reliable high-resolution moisture distribution images. To address this, the paper introduces a double-stage neural network approach that incorporates the use of a second network to enhance the images obtained from the first network. Both neural networks are trained on a similar set of pattern output images, with the first network generating training images for the second network. The use of LSTM network and a unique approach to input measurements and images further contribute to the effectiveness of the proposed method.