Software fault prediction plays a key role by identifying defect-prone modules as early as possible in the development lifecycle to reduce maintenance costs and increase software reliability. The emergence of deep learning methodologies over the past years as useful tools for automatic fault detection has to do with its capacity to learn complex relationships from high-dimensional software metrics. In this paper, we compare the performance of three different Deep Learning Models i.e. CNN, RNN and a Cascade model consisting of a combination of CNN and LSTM layers on software failure prediction. Extensive experimental evaluation is conducted with a labeled dataset of a variety of software metrics. Normalization of dataset & reshaping the data into sequences for learning. Each model is trained and validated using standard performance metrics: Accuracy, F1 Score, Precision, Recall, and ROC AUC. The CNN and the RNN models had similar performance, with accuracies of 81.4% and 81.1%, respectively. The proposed Cascade model outperforms both alternatives by far, achieving 86.5% accuracy, 84.5 F1 Score, 84.6 Precision, 86.5 Recall and 83.7 ROC AUC. Results suggested that the Cascade model leverages spatial and sequential relationships in the input data effectively, leading to improved fault prediction accuracy and reliability. Overall, this hybrid scheme outperforms all stand-alone models in all evaluated measurements, making it a viable candidate for practical defect detection systems. The results suggest that combining the feature extraction capabilities of CNNs with the temporal modeling capabilities of LSTMs can significantly enhance predictive performance. Future work will include extending the model to support cross-project failure prediction and incorporating explainable AI approaches for better interpretability.