<p>Robot imaging-based pipeline monitoring, such as that used in the oil and gas industry, plays a crucial role in object detection and localization for pipeline inspection. It ensures safety and efficiency by transmitting real-time visuals to Supervisory Control and Data Acquisition (SCADA) systems. Deep learning-based methods have recently been utilized due to their superior feature extraction and pattern recognition capabilities. However, existing studies often overlook interpretability and uncertainty, making it difficult to align results with real-world scenarios. This paper proposes a Convolutional Neural Network (CNN) optimized with Bayesian hyperparameter tuning, incorporating an Accuracy, Uncertainty, and Interpretability (AUI)-based objective. The approach begins with data processing using the You Only Look Once (YOLO) labeling format and a Gaussian filter for denoising and quality enhancement. Uncertainty is quantified via confidence intervals, while interpretability is embedded during training through a mapping-based reconstruction process that links outputs back to inputs. Unlike conventional post-hoc interpretability methods, our approach integrates interpretability directly into the optimization loop. Experimental results on a realistic imaging dataset of vehicular robots show that AUI-CNN achieves a mean classification performance of 0.8826 during training and 0.8323 during testing, improving 4.18% over CNN and 0.42% over U-CNN. Embedding interpretability and uncertainty within training enhances reliability and transparency, which are critical factors in safety-sensitive domains like oil and gas. These results demonstrate a robust, interpretable, and real-time solution for pipeline monitoring.</p>