The increase of penetration of renewable energies has posed inevitable challenges to the stability and safety of power system operations, especially in large-scale multi-machine power systems. Emergency control is thereby crucial to avoid catastrophic accidents, and identifying coherent generators is the basis of wide-area control of a multi-machine power system. However, existing approaches are rule-based or rely on shallow machine learning, lacking effectiveness and robustness due to their insufficient ability of pattern mining from system monitoring indicators. To fill the gap, this paper proposes a novel end-to-end generator coherency identification framework, leveraging an improved auto-encoder to comprehensively exploit information of phasor measurement units (PMUs) obtained from wide-area measuring systems (WAMS). The framework jointly trains the feature extraction module and the clustering module to fully explore the shared knowledge and obtain cluster-specific representations. In addition, a visualization component is equipped with the process-agnostic framework for interpretability. Simulated and practical case studies validate the effectiveness of the proposed approach as it outperforms both deep learning baselines and state-of-the-art methods on all datasets under various situations, including observation window size changes, noisy data, or data missing at random.