The early detection and accurate classification of mechanical faults in rotating machinery is a critical challenge in industrial predictive maintenance. This paper presents a novel hybrid architecture that synergistically combines the expressive logical reasoning capabilities of Logic Tensor Networks (LTNs) with the powerful sequence modeling features of Transformer architectures. Our approach targets fault classification tasks on two publicly available benchmark datasets: the SpectraSimulator gearbox fault diagnosis data provided by the Open Energy Data Initiative (OEDI), and the gear fault datasets from the University of Connecticut (UoC). The Transformer component serves as a high-fidelity feature extractor, leveraging attention mechanisms to learn temporal and spectral dependencies in vibration signals. These latent representations are then integrated into an LTN framework that enforces domain-specific logical constraints through soft first-order logic, providing a structured and interpretable reasoning layer. Importantly, we introduce a dynamic rule generation module that formulates symbolic constraints based on evolving patterns in the data, domain knowledge, and clustering-driven fault signatures. These rules are encoded in the LTN to enhance generalization and reduce overfitting on limited or noisy annotations. Experimental results demonstrate that our hybrid model significantly outperforms conventional deep learning classifiers and standalone logic-based systems, particularly in settings with scarce labeled data or ambiguous fault conditions. The integration of dynamically induced logical rules with learned attention-based features allows the system to retain high accuracy while ensuring semantic interpretability and robustness. This work highlights the potential of neuro-symbolic fusion for intelligent fault diagnosis in complex mechanical systems.