We introduce a novel approach to enhance gearbox fault diagnosis by integrating Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs) for vibrational data analysis. Our method aims to improve fault detection accuracy, particularly in identifying subtle anomalies like broken teeth. However, real-world data often contains noise, which can hinder the effectiveness of such models. To address this challenge, we incorporate Singular Value Decomposition (SVD) pooling layers within the model. Our methodology starts with continuous wavelet transform (CWT), applied to the vibrational data to reveal crucial frequency-domain features. Concurrently, a CNN, using the Inception architecture, extracts spatial features. Simultaneously, LSTM networks capture temporal patterns. The unique feature representations from the CNN and LSTM branches are fused, creating a holistic feature set incorporating spatial, material, and frequency-domain information. This integrated feature set is then classified using a fully connected neural network. Our method's effectiveness is rigorously validated through comprehensive experiments on a diverse dataset. The results demonstrate exceptional accuracy in identifying gearbox faults, even in the early stages. This research advances predictive maintenance, offering a precise and comprehensive approach to gearbox fault diagnosis. In conclusion, the fusion of LSTM and CNN architectures for vibrational data analysis holds promise for gearbox fault diagnosis, benefiting industries reliant on machinery reliability and operational efficiency.