TF-Mamba: Intelligent fault diagnosis model for wind turbine bearings based on multimodal selective gating Mamba

To address the non-stationarity of wind-turbine bearing vibration signals, the need for collaborative modeling of one-dimensional sequential features and two-dimensional visual features, and the computational cost of sequence modeling, this study proposes an intelligent fault diagnosis model based on multimodal selective-gating Mamba, termed TF-Mamba. A progressive multimodal fusion framework of “feature extraction–CoSSM–explicit selection–global decision-making” is constructed. Specifically, CoSSM enables cross-modal feature interaction through shared-parameter Mamba encoding and reciprocal gating. Bi-SG performs adaptive explicit selection of fused features along the channel and spatial dimensions, and EnSSM models global sequential dependencies for final decision-making. Experiments are conducted on CWRU and the State Grid Wind-turbine Dataset (SGWD), with BMTM-Net and WCamba included as recent baselines. The main comparison yields accuracies of 99.07% and 96.85%, respectively. Across ten repeated runs, TF-Mamba achieves mean accuracies of 99.02 ± 0.05 % on CWRU and 96.75 ± 0.12 % on SGWD. Paired permutation tests against BMTM-Net give Holm-adjusted p = 0.0156 for Accuracy and Macro-F1 on both datasets. In four leave-one-load-out CWRU evaluations, its mean accuracy across the held-out loads is 97.75%, with a mean Macro-F1 of 0.9757. A separate gate-substitution comparison gives Bi-SG an SGWD mean-accuracy advantage of 0.54 percentage points over CBAM. The network contains 2.81 million parameters and has a reported single-sample inference time of 4.68 ms on an RTX 4090. These results support the diagnostic value of progressive time–frequency collaboration and selective fusion under the evaluated dataset and load conditions.