PSI - Issue 64

7

Nikhil Holsamudrkar et al. / Procedia Structural Integrity 64 (2024) 580–587 Holsamudrkar Nikhil et al./ Structural Integrity Procedia 00 (2019) 000 – 000

586

Fig. 6 – (b) describes the Receiver Operating Characteristic (ROC) curve, indicating the diagnostic ability of a binary classifier system as its discrimination threshold is varied. The Area Under the ROC Curve (AUC-ROC) is a single scalar value that summarizes the classifier's performance over all possible thresholds. However, the current study involves a multiclass system one vs all method for generating an ROC curve. Further micro-averages of all classes are taken to arrive at a final ROC curve that indicates overall performance. An AUC-ROC of 0.72 indicates that the classifier has a moderate ability to distinguish between the positive and negative classes. It's better than random guessing (AUC-ROC = 0.5) but still has room for improvement. The author's work in this direction is in progress and includes hybrid models to mitigate this problem. Fig. 7 contains various evaluation metrics used to classify the data. These metrics encompass different aspects of model performance, including predictive accuracy, precision, recall, and the balance between them. They include metrics like Informedness (BM), which assesses the model's ability to predict positive and negative cases. Positive Predictive Value (PPV) and Negative Predictive Value (NPV) quantify the accuracy of positive and negative predictions, respectively. Meanwhile, the F1 Score provides a balanced measure of precision and recall. Additionally, metrics like the Matthews Correlation Coefficient (MCC) and the Diagnostic Odds Ratio (DOR) offer comprehensive assessments of classification performance by considering both true positive and false positive rates. These metrics collectively provide a thorough understanding of the model's classification capabilities.

Fig. 7 . Heatmap for different metrics

5. Conclusions The present study develops a Deep CNN model to classify damage mechanisms in a flexurally strengthened RC beam with FRCM composites. It includes fiber rupture, fiber-matrix debonding, concrete/Mortar cracking, and steel yielding. The data collection includes the waveform data in the frequency range of 20-1000 kHz using an Acoustic emission system. The waveform data was converted to DWT images, considering the waveform's time, frequency, and amplitude features. The model's training accuracy is about 95%, whereas the test accuracy is 87%. The low test accuracy is due to the introduction of additional noise in the test data to simulate real-world settings. Further work in this direction includes expanding the dataset for other failure modes (shear) and improvising the model results using advanced techniques.