Electromechanical Impedance Data-Driven Metal Structural Tensile Stress Identification Using Generative Adversarial...

Area of Science:

• Materials Science
• Mechanical Engineering
• Artificial Intelligence

Background:

• Deep learning networks are crucial for automated structural stress identification using electromechanical impedance/admittance (EMI/EMA) data from piezoelectric ceramic (PZT) transducers.
• Insufficient data quantity and quality typically hinder the performance of data-driven deep learning models in this domain.

Purpose of the Study:

• To propose an innovative data enhancement method, the EMA generative adversarial network (EMAGAN), to address data inefficiency and deficiency for deep learning-based stress identification.
• To improve the accuracy and efficiency of metal structural stress identification using limited measurement data.

Main Methods:

• Developed an original data enhancement method using the EMA generative adversarial network (EMAGAN).
• Tuned a novel data-normalization algorithm to facilitate EMAGAN-based dataset generation.
• Integrated synthetic and original datasets and fed them into an adaptively established one-dimensional convolutional neural network (1DCNN) for stress prediction.
• Validated the method on an aluminum beam specimen under uniaxial tensile load monitored by PZT transducers.

Main Results:

• The EMAGAN generated high-accuracy EMA data, exceeding the normal collection method by over 380 times.
• The synthetic datasets significantly improved the performance of the 1DCNN model for stress identification.
• Statistical error analysis confirmed the efficacy of the generated EMA datasets compared to raw data.

Conclusions:

• The EMAGAN paired with 1DCNN offers a promising approach for data-driven metal structural stress identification.
• This method enhances efficiency, intelligence, and accuracy in identifying structural stress.
• The study overcomes limitations of traditional EMA techniques in data acquisition for deep learning applications.