Diagnosing rolling-element bearing faults under time-varying operating regimes remains challenging for predictive maintenance. We present an Explainable Causal Disentanglement Network (XCDN) that combines a Conv1D autoencoder with an explicit split of the latent space into causal (fault-related) and contextual (speed/RPM) factors, enforced via domain-adversarial training with a gradient-reversal layer. A grouped, stratified split prevents leakage; a small random search over hyperparameters selects the best setting, which is re-trained on train+validation and evaluated once on a held-out test set. On a variable-speed bearing dataset, XCDN attains test accuracy=0.9994 and macroF1=0.9994 (2 errors over 3,510 windows), with a confusion matrix separating Healthy, Inner Race Fault, and Outer Race Fault. Latent-space analysis indicates that the causal subspace z c forms three separable class clusters, whereas the domain subspace z d aligns smoothly with RPM and shows no class structure. Explainability audits corroborate these findings: Integrated Gradients and SHAP on z c identify a small subset of latent dimensions as the main drivers of each diagnosis, while input-level IG and a simple channel-ablation check indicate that vibration predominates in fault classification and RPM in the context regressor.The approach yields a compact, traceable representation that supports simple thresholding on a few latent nodes under time-varying conditions.