Abstract: Bearings are critical components in rotating machinery, and their condition directly determines the operational efficiency and safety of industrial equipment. Traditional bearing fault diagnosis methods rely heavily on manual feature extraction, making it difficult to deal with complex industrial scenarios such as strong noise, variable working conditions, and scarce fault samples. In recent years, deep neural networks have overcome the inherent limitations of traditional methods through robust adaptive feature learning, thereby establishing widespread utility in the field of bearing fault diagnosis. This review systematically summarizes the progress of neural network-based bearing fault diagnosis. First, the importance of bearing fault diagnosis and the shortcomings of traditional methods, including time-frequency analysis, empirical mode decomposition, variational mode decomposition, and entropy-based methods, are discussed, highlighting their strong dependence on expert experience, weak noise robustness, poor generalization under variable conditions, and difficulty in identifying compound faults. Subsequently, a comprehensive comparison of various neural network architectures is made. Convolutional neural networks (CNNs) are good at extracting local spatial features. A one-dimensional CNN is lightweight and functions in real time but has difficulty capturing long-term dependencies. A two-dimensional CNN combined with a time-frequency diagram has strong visual interpretability, but the computational cost is high. An improved CNN combined with multiscale convolution, graph convolution, attention mechanism, and residual connection could significantly enhance the robustness and generalizability of the model under strong noise and variable conditions; however, it also increases the complexity of the model. While recurrent neural networks (RNNs), particularly long short-term memory (LSTM) and bidirectional LSTM, effectively capture temporal dependencies in vibration signals, they are susceptible to vanishing or exploding gradients and incur high training costs. Hybrid models, such as CNN-RNN and CNN-Transformer, combine local feature extraction with global dependency modeling. CNN-RNN achieves balanced learning of spatiotemporal features at a moderate cost. CNN-Transformer demonstrates clear advantages in dealing with long sequences, strong noise, and compound faults, but requires substantial volumes of data and computing resources. This review also analyzes four major engineering challenges in bearing fault diagnosis: variable operating conditions, strong noise environments, small samples and data imbalance, and cross-machine or cross-domain diagnosis. Domain adaptation and domain generalization techniques for variable conditions are examined. For the industrial strong-noise environment, noise robust methods such as the deep residual shrinkage network and attention mechanism are summarized. The problem of data scarcity and imbalance is addressed through data enhancement strategies such as generative adversarial networks and digital twin technology, as well as few-shot and meta-learning methods. Cross-device diagnostic methods, such as federated learning and domain alignment strategies, are summarized. Although considerable progress has been made, several gaps remain in current research. First, the diagnostic accuracy of the model often decreases sharply under extreme dynamic conditions, such as continuous change of rotational speed, impact load, and temperature drift. Second, multisource heterogeneous data fusion faces challenges when the sensor sampling rates differ, and certain channel data are temporarily missing. Third, there is an inherent contradiction between the lightweight model design for edge deployment and the noise robustness under low signal-to-noise ratios. Fourth, unsupervised self-learning research is insufficient in the context of unlabeled data in actual industrial monitoring. To address these research gaps, this paper proposes future research directions, including domain generalization methods based on causal inference to decouple fault causal characteristics and working condition interference; cross-modal attention and graph neural network for intelligent multisource data fusion; a joint optimized lightweight neural network for resource-constrained edge devices that takes into account delay, power consumption, and noise robustness; and self-supervised learning for unlabeled monitoring data that automatically finds fault features. This review aims to provide a systematic reference for subsequent research and engineering implementation of intelligent bearing fault diagnosis, supporting the advancement of smart manufacturing.