Acoustic Emission Analysis in Ultrasound Welding

So far in this research, a novel way of carrying out non-destructive testing on ultrasound welds has been explored. For the non-destructive aspect, acoustic emissions emitted during ultrasound welding are utilized. However, for a machine learning task, the raw data was extremely computationally expensive, and the dataset was also very small. To address this, the raw data was processed and spectrograms are generated from them, which are an accurate representation of acoustic emissions in the time-frequency domain. The generated spectrograms are used as images for an image classification task. Here, the issue of the dataset being very small still exists, and there is a heavy class imbalance where the number of successful welds outnumbered the number of failed welds by at least 2:1. To address this, data augmentation techniques are employed and a lot more spectrograms are generated from the original data. Data is also generated in a way that addresses the class imbalance problem. At the end, a dataset with almost 6,500 images is obtained, with a 70-15-15 split for training, validation, and testing. For prediction, three state-of-the-art machine learning models are leveraged: ResNet- 18, EfficientNet_B0, and Vision Transformer (ViT). These models are all selected based on their performance on the ImageNet dataset, making them highly suitable for image classification tasks. These models were modified for the use-case by changing their final classification layers and carrying out other parameter tuning. The models were all evaluated on an unseen test set for fairness, and the best model was selected based on various factors such as accuracy, precision, recall, F1 score, and training stability. Just comparing false negative and false positive rates is not enough to determine which model is the best, so a statistical significance test is performed to determine if the difference in accuracies is significant. McNemar’s test is employed to compare the performances of all three models. It is found that the performance of ViT is statistically equivalent to EfficientNet_B0, and both of them are significantly better than ResNet-18. These results are discussed and the best model is recommended for industrial deployment. Evidence is provided that EfficientNet_B0 is the best model to use for industrial deployment, as it is more computationally efficient, has a good balance of accuracy (97%) and precision (0.97), and is more stable during training. The stability in training is 57 critical for industrial applications, as it ensures that the model will perform consistently across different environments and data distributions. Therefore, for industrial deployment, EfficientNet_B0 is the best model to use. In conclusion, it has been shown that a machine learning model can be trained to classify ultrasound welds based on acoustic emissions. It has also been shown that EfficientNet_B0 is the best model to use for industrial deployment, as it is more computationally efficient, has a good balance of accuracy and precision, and is more stable during training.