A Bladed Disk Fault Diagnosis Method Based on Multi-Channel Spatio-Temporal Representation of Blade Tip Timing Signals and an Attention-Enhanced Residual Network

Turboshaft engine bladed disks are prone to typical faults such as tenon-root cracks and rub-impact during operation. Blade tip timing (BTT) signals inherently contain both blade arrival-time information and blade spatial-distribution information, which are coupled with each other. Conventional data representation methods lack the capability to deeply mine such complex coupling relationships. To address this issue, a diagnostic method integrating a physically coupled multi-channel two-dimensional spatiotemporal representation with an attention-enhanced residual network is proposed. First, the Error Compensation Coefficient (ECC) method is employed to correct the raw blade spacing data, eliminating static deviations introduced by manufacturing and assembly. Second, the corrected data are organized into a two-dimensional field indexed by blade spacing and revolution—the physical variables of the measurement—and three physically complementary channels are constructed, namely the ECC baseline value, the first-order temporal difference (ΔS) and the spatial coefficient of variation (SCV), forming an ECC–ΔS–SCV three-channel two-dimensional spatiotemporal feature map that represents the magnitude, the temporal dynamics and the spatial cooperativity of the blade group, respectively. Finally, the feature map is decoded by a deliberately standard ResNet18 backbone embedded with the Convolutional Block Attention Module (CBAM, with stage-adapted spatial kernels), so that the contribution of the representation can be isolated from that of the network. Experimental data were collected from a self-built mortise–tenon bladed-disk test rig under three conditions—healthy, tenon-root crack and rub-impact—at rotational speeds of 500, 1000 and 1500 rpm. Each configuration is trained ten times using a fixed data split with different random seeds, and all results are reported as mean ± standard deviation. Under identical data splits and training configurations, the proposed method achieves mean classification accuracies of 95.07%, 92.62% and 91.25% at the three speeds (92.98% on average), outperforming five mainstream deep learning benchmark models. Ablation studies confirm that each feature channel is necessary and the attention module as a whole is beneficial, and that the improvement is attributable to the representation rather than to model scale; a leave-one-speed-out evaluation further quantifies the performance degradation when the test speed is excluded from training. This study provides a physically coupled representation and an effective diagnostic framework for BTT-based fault diagnosis of rotating machinery.