Strain is a highly effective damage-sensitive feature for Structural Health Monitoring (SHM) of masonry structures. However, Environmental and Operational Variations (EOVs), such as air temperature changes, can induce drifts in strain time series that mask damage-related patterns, requiring compensatory techniques for reliable novelty detection. This paper proposes a cointegration-driven, auto-adaptive neural network strategy for strain-based SHM in masonry structures, a field where effective approaches remain limited. The methodology integrates nonlinear cointegration principles with a customized Multilayer Perceptron (MLP) regressor at the sensor level. Unlike standard implementations, the MLP captures time dependency through a modified input matrix that incorporates historical information, enabling damage-sensitive features cleansed of EOV effects. Novelty analysis then combines stationarity assessment with statistically robust control charts for damage detection. A Recurrent Neural Network (RNN) validates the proposed time-dependent MLP model, and a Convolutional Autoencoder (CAE) is adopted as a benchmark to assess novelty detection against an established deep-learning paradigm. A case study consisting of a full-scale masonry building monitored through smart bricks and subjected to varying damage scenarios under real environmental conditions demonstrates the effectiveness of the proposed strategy. Results show that the time-dependent MLP achieves higher computational efficiency than the RNN while maintaining accuracy, that the nonlinear cointegration approach overcomes key limitations of linear methods in EOV compensation and damage detection, and that the proposed framework matches the CAE in novelty detection at a markedly lower computational cost. Overall, this supports a broader and more effective application of strain-based SHM in masonry structures.
Mattiacci et al. (Mon,) studied this question.