<p>Cyberbullying in Bangla social media is an increasingly critical issue, posing severe risks to mental health and online safety. Despite the growing need for automated detection systems, challenges such as limited linguistic resources and the nuanced characteristics of the Bangla language hinder model performance. In this study, we propose TriB-FNN, a novel and explainable three-stage fusion neural network, designed to effectively capture cyberbully in Bangla social media texts. The model integrates Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (Bi-LSTM), and Gated Recurrent Units (GRUs) in a hierarchical pipeline. To optimize training, we employ a Composite Adaptive Momentum (CAM) Optimizer, and compare its performance against traditional optimizers such as Adam, Adagrad, and SGD with momentum, demonstrating superior convergence and stability. The model was trained and evaluated on three publicly available datasets: an initial set of 44,001 Bangla social media comments which contains cyberbully, along with two supplementary datasets containing similar hate speech and bully text. This setup ensures robustness, diverse linguistic representation, and rich contextual variation. Extensive preprocessing and multiple text data augmentation strategies were applied to enhance model generalization under low-resource conditions. Cross-entropy loss is used for optimization, and the total number of trainable parameters are 15.3 million. The proposed model achieves 99.06% accuracy, with precision, recall, and F1-score all exceeding 99%, significantly outperforming baseline models. Furthermore, explainability technique such as LIME is employed to provide transparent and interpretable predictions, revealing the decision pathways of the model. These results highlight the efficacy and robustness of TriB-FNN for real-world cyberbully identification in Bangla text, setting a new benchmark for low-resource language classification tasks.</p>