Structural damage identification based on a novel statistics-informed neural network and piezoelectric current signals

Identifying simultaneous damage-related changes in multiple structural regions requires informative signal representations and reliable regional regression. This study proposes a statistics-informed multi-input multi-output convolutional neural network (SIMIMO-CNN) for structural damage identification using lead zirconate titanate (PZT) current time histories. Multichannel convolutional features are fused with channel-wise root-mean-square (RMS) descriptors to combine temporal patterns with explicit signal-amplitude information. A region-weighted mixed mean squared error–Huber loss applies separate penalties to damaged and undamaged regions, with an adaptive Huber threshold. The network maps the sensor signals directly to a continuous regional prediction vector. Experiments use an 18-region concrete slab and a five-region scaled bridge model. For per structure, average damaged-region relative error decreases from 0.151 to 0.047 for the slab compared with a conventional CNN. The corresponding bridge error decreases from 0.243 to 0.131. Average cosine similarity reaches 0.9908 and 0.9810, respectively. These results confirmed the effectiveness of the proposed method.