In response to the distribution differences under different operating conditions and the privacy protection constraints of signal data, a semi-supervised domain adversarial federated collaborative model (SAFCM) was designed to realize the remaining useful life prediction of mechanical equipment under cross-operating conditions and privacy preservation. A bidirectional attention temporal convolutional network was designed to extract domain-invariant features through two directions of channel attention mechanism and self-attention mechanism. By using this convolutional network as a feature extractor, a cross-operating-condition semi-supervised domain adversarial model (COCSAM) was constructed to achieve highly robust domain adaptation between imbalanced domains through domain adversarial learning and domain distance measurement. The COCSAM was integrated into the federated collaborative framework to build the SAFCM. Through differentiated distribution and dynamic weighted updates of the sub-models, asynchronous network updates were achieved under data privacy protection, promoting the full utilization of data and efficient domain adaptation. Two datasets from different application scenarios were utilized to conduct the interpretability analysis and robustness study, the results of which demonstrated the robustness and superiority of the proposed SAFCM in real industrial scenarios.