An innovative approach to detecting moisture in walls using electrical tomography and LSTM networks
A team of researchers from the Lublin University of Technology, the NETRIX S.A. Research and Development Center, the WSEI Academy in Lublin, the Wrocław University of Science and Technology, and Netrix S.A. have developed a groundbreaking method for imaging moisture in masonry structures by combining the technology of electrical impedance tomography (EIT) with advanced LSTM neural networks. The innovative approach, described in the journal Measurement, uses multi-sequence measurements to reconstruct the moisture distribution, significantly outperforming previous methods based on single data vectors.
Unlike classical solutions, the proposed method is based on the analysis of five independent measurement sequences, each of which was intentionally disturbed with noise of up to 30% to reflect realistic measurement conditions. In a real environment, such interferences – resulting from, among others, the presence of electromagnetic fields or equipment instability – constitute a serious challenge that the new technology effectively overcomes.
A multi-branch BiLSTM neural network with a 21-layer architecture was used to analyze the data. Its design allows for simultaneous processing of data in both time directions, which allows for capturing complex spatial-temporal relationships related to the moisture distribution. This allows for obtaining more precise reconstructions, even in the case of walls with a complex material structure.
The effectiveness of the method was confirmed by comparing the EIT results with data obtained using reference methods, such as dielectric measurements, gravimetric measurements and thermal imaging. The agreement between the reconstructed images and the actual moisture distribution was high, which is also confirmed by the reconstruction quality indicators. The analysis of metrics such as MSE, PSNR, SSIM and ICC showed a significant improvement in accuracy – the average MSE error was reduced from 3.657 to 2.863, and the SSIM structural similarity index increased from 0.835 to 0.868 when using five measurement vectors instead of one.
Moisture in walls is not only destructive to the structure of the building, but also dangerous to the health of residents. It causes degradation of building materials, promotes the growth of mold and fungi, reduces the thermal insulation properties of walls and leads to expensive repairs. Moisture diagnostics is of particular importance in the case of historic buildings, where invasive methods are often impossible to use. The new technology allows for precise and non-destructive moisture detection, which makes it a tool with great potential in the conservation of architectural heritage and in monitoring the technical condition of modern buildings.
The developed solution represents a significant step forward in the field of non-invasive building diagnostics, demonstrating that the integration of impedance tomography with deep learning models such as LSTM allows for high precision in complex and noisy measurement environments. Further research could pave the way for commercialization of this technology in the heritage protection, construction and civil engineering sectors.
The full publication is available at:
link to publication