Abstract: Rigid guide is a key component of vertical shaft hoisting systems. Due to shaft deformation and steel corrosion, faults are likely to occur, affecting the normal operation of the cage. At present, vibration detection methods are mostly used for rigid guide fault diagnosis, but the diagnostic accuracy is easily affected by operating conditions such as cage load and running speed. To address this problem, eddy current sensors are used to collect fault signals of the rigid guides, ensuring that the signal features are not affected by the cage's operating environment. To improve the accuracy of rigid guide fault identification, Residual Attention One-Dimensional Convolutional Neural Network (RA1DCNN) was proposed. The network extracted multi-scale features through parallel multi-scale convolutions, enhancing its ability to perceive signal features at different scales. Channel attention and spatial attention modules were introduced and combined with the residual learning mechanism to design a residual attention module, which simultaneously captured channel and spatial feature information, extracting more discriminative features. An experimental platform for rigid guide faults was established to simulate different types and severities of rigid guide faults. Ablation experiments and comparative experiments were conducted on RA1DCNN. The results showed that the RA1DCNN achieved 100% accuracy in identifying rigid guide fault categories, an average accuracy of 99.7% in identifying the severity of clearance faults, and an average accuracy of 97.68% in identifying the severity of misalignment faults, verifying the effectiveness of the multi-scale convolution layers and residual attention module. The overall fault identification accuracy reached 98.05%, outperforming comparative models including one-dimensional convolutional neural networks.