Objective To construct an improved bidirectional long short-term memory (BiLSTM) model incorporating a temporal attention mechanism for fault warning of multi-parameter patient monitor. Methods Quality control records and maintenance data from 96 patient monitors in the emergency department of Shanghai East Hospital, Tongji University, were collected. Functional channels were used as modeling units, and 5,674 channel-level time-series samples were generated using a sliding-window approach. The samples were divided into training and test sets at a ratio of 7 vs. 3. A temporal attention layer was added after the BiLSTM encoder, and long short-term memory (LSTM), gated recurrent unit (GRN), temporal convolutional network (TCN), and standard BiLSTM models were used as comparators. The models were systematically evaluated using the area under the curve (AUC) of receiver operator characteristic (ROC) curve, precision-recall (PR) curve, probability calibration curve, and threshold-dependent classification metrics to identify the model with the best overall performance. Results The improved BiLSTM achieved the best overall performance, with an AUC of 0.918, average precision (AP) of 0.553, sensitivity of 80.43%, and F1 score of 65.49%. The AUC differences between the improved BiLSTM and all comparator models were statistically significant (all P＜0.05). The Hosmer-Lemeshow test indicated good calibration of the improved model (P＝0.480), whereas the comparator models showed significant calibration bias (all P＜0.001). Conclusion The improved BiLSTM model can identify cross-cycle degradation trends in quality control parameters of multi-parameter patient monitor. It outperforms comparator models in both fault warning and probability calibration, and can provide auxiliary decisionmaking basis for preventive maintenance of medical equipment.