Fighting fraudulent insurance claims is a vital task for insurance companies as it costs them billions of dollars each year. Fraudulent insurance claims happen in all areas of insurance, with auto insurance claims being the most widely reported and prominent type of fraud. Traditional methods for identifying fraudulent claims, such as statistical techniques for predictive modeling, can be both costly and inaccurate. In this research, we propose a new way to detect fraudulent insurance claims using a data-driven approach. We clean and augment the data using analysis-based techniques to deal with an imbalanced dataset. Three pre-trained Convolutional Neural Network (CNN) models, AlexNet, InceptionV3 and Resnet101, are selected and minimized by reducing the redundant blocks of layers. These CNN models are stacked in parallel with a proposed 1D CNN model using Bagged Ensemble Learning, where an SVM classifier is used to extract the results separately for the CNN models, which is later combined using the majority polling technique. The proposed method was tested on a public dataset and produced an accuracy of 98%, with a 2% Brier score loss. The numerical experiments demonstrate that the proposed approach achieves promising results for detecting fake accident claims.