To address the dual challenges of obtaining substantial labelled fault data under a single operating condition for the same equipment and distribution shifts across multi-source domains, a multi-source domain generalisation method was proposed based on a time-frequency interaction strategy incorporating a time-frequency contrast loss. A multi-source domain adaptation strategy was employed to reduce global distribution differences across all source domains, combined with a time-frequency contrast loss to preserve categorical relationships between samples, thereby enhancing the model’s generalisation capability under unknown operating conditions. To validate the method’s efficacy, transfer diagnostics experiments were conducted on a variable-condition bearing dataset. Results demonstrate superior diagnostic capability and generalisation performance under unknown conditions compared to existing multi-domain generalisation approaches, providing a solution for zero-shot transfer diagnostics in industrial equipment.