Tightness evaluation of bolted joint structures using dynamic characteristic based on machine learning

Abstract

Due to the rapid development of deep learning technology, it is used for vibration-based structural condition monitoring. When vibration is used to extract features for system diagnostics, it is important to correlate the measured signal with the current state of the structure. The measured vibration response shows a large deviation in the spectral and transient characteristics of the system to be monitored. As a result, diagnostics using vibration requires a full understanding of the extracted features in order to eliminate the influence or unnecessary changes of the surrounding environment. In-depth learning-based algorithms are expected to increase their application in these complex problems due to their flexibility and robustness. This dissertation provides a summary of the study applying machine learning algorithms for defect monitoring. The study was categorized using vibration coefficients. A simple interpretation of deep neural networks is provided to guide further applications of structural vibration analysis. This dissertation presents a new method of measuring the clamping force using the vibration of a bolt. The resonance frequency of the bolt increases with the clamping force during the tightening process. These characteristics were measured and utilized in the k-means clustering algorithm. The bolt specimen was fixed to the load cell using a nut runner for verification of the proposed method. The clamping force measured accurately was displayed. Using the labeled data, the clamping force was estimated from the vibration response. To use the proposed method when assembling real parts, an accelerometer was attached to the nut runner for vibration measurement. This allowed the clamping force to be continuously monitored without affecting the part. The estimated clamping force was compared to the clamping force of the torque method. When the vibration of the bolt was transmitted through the nut runner, there was a loss of high frequency vibration energy. The clamping force was determined by maintaining the vibration of the resonance frequency band of the bolt. Components of the low-band were excluded using a band-pass filter. The clamping force of the bolts used for the vehicle's lower arms and links was also accurately evaluated. By using the proposed method, changes in the clamping force during the manufacturing process can be continuously monitored. Clamping forces were determined using acoustic signals as well as vibrations that occur when the bolts are tightened. Characteristic changes are utilized as features analyzed by a convolutional neural network. The clamping force is measured using a load cell and then used during labeling for sorting. To measure radiated noise, a microphone is installed near the fixed part. In order to apply this to deep learning classification, a signal processing method was proposed and data was reinforced. The convolutional neural network architecture was modeled and the clamping force was determined by the classification method. The estimates were compared to actual load cell measurements. In the case of fastened bolts, the contact strength depends on the difference in the clamping force. To detect this, an excitation type measuring device was devised and attached to the bolt head. When the excitation force of various sizes is applied to the measuring equipment, the bolt and the structure are separated, which acts as an attenuation of the vibration signal. The clamping force was predicted using a non-linear separation phenomenon that appeared to be small at a large clamping force. Coefficients of exponential decay were derived using curve fit for each section, and the difference between their mean values was measured.