研究目的
To develop a non-contact and high-speed damage detection technology for concrete structures using a laser Doppler vibrometer and machine learning.
研究成果
Machine learning using the convolutional neural network (CNN) was carried out using the frequency spectrum as an input quantity. By optimizing the structure of the neural network and preprocessing of the data, it was possible to discriminate between the cracked part and the intact part of the concrete with a high accuracy of over 90%.
研究不足
The difference between the defective part and the intact part in the frequency spectrum is not large, and it is difficult to manually determine a unified standard for discrimination. The experimental results did not show the predicted difference clearly in the low frequency region for the cracked part.
1：Experimental Design and Method Selection:
A laser Doppler vibrometer was used to obtain the vibrations of a concrete structure at a high signal-to-noise ratio. The observed vibration data were transformed into frequency spectra by Fourier transform. Machine learning using a convolutional neural network was employed to classify the spectra.
2：Sample Selection and Data Sources:
The object of the inspection was the concrete pillar of a building which had been cracked by an earthquake.
3：List of Experimental Equipment and Materials:
A Polytec laser vibrometer system with a CLV-700 Head, CLV-1000 Controller, and a Vib-Z-010 data acquisition unit was used. Polytec’s VibSoft software (version 3.3) was used for hardware control, acquisition, and postprocessing.
4：3) was used for hardware control, acquisition, and postprocessing.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: The distance from the laser Doppler vibrometer to the concrete surface was fixed to be 30 cm. A place about 5 cm away from the irradiation point of the laser was excited using an inspection hammer with a 100 g head weight. The sampling frequency was 25.6 kHz, and a time range of up to 1.28 s (32768 points) was measured.
5：6 kHz, and a time range of up to 28 s (32768 points) was measured.
Data Analysis Methods:
5. Data Analysis Methods: The acquired data was Fourier transformed by VibSoft and output as a spectrum of 0–10 kHz with a frequency resolution of 0.78125 Hz (12800 points). Machine learning using a convolutional neural network was carried out to classify the spectra.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
