A technique for on-line quality detection of ultrasonic wire bonding is developed. The electrical signals from the ultrasonic generator supply, namely, voltage and current, are picked up by a measuring circuit and transformed into digital signals by a data acquisition system. A new feature extraction method is presented to characterize the transient property of the electrical signals and further evaluate the bond quality. The method includes three steps. First, the captured voltage and current are filtered by digital bandpass filter banks to obtain the corresponding subband signals such as fundamental signal, second harmonic, and third harmonic. Second, each subband envelope is obtained using the Hilbert transform for further feature extraction. Third, the subband envelopes are, respectively, separated into three phases, namely, envelope rising, stable, and damping phases, to extract the tiny waveform changes. The different waveform features are extracted from each phase of these subband envelopes. The principal components analysis method is used for the feature selection in order to remove the relevant information and reduce the dimension of original feature variables. Using the selected features as inputs, an artificial neural network is constructed to identify the complex bond fault pattern. By analyzing experimental data with the proposed feature extraction method and neural network, the results demonstrate the advantages of the proposed feature extraction method and the constructed artificial neural network in detecting and identifying bond quality.
A Stacked Autoencoder Neural Network based Automated Feature Extraction Method for Anomaly detection in On-line Condition Monitoring
