As a key sensor for real-time monitoring of coal-rock mass stress, borehole stress gauges are persistently affected by complex factors such as strong electromagnetic interference and mining-induced disturbances in underground environments. The reliability of their performance and data directly impacts the accuracy of disaster risk identification. To address the challenge of identifying sensor failure states, a failure detection method for borehole stress gauges was proposed based on multi-criteria fusion and attribution analysis. A five-dimensional criterion system was established, incorporating the proportion of abnormal data, prolonged consecutive identical values, overall coefficient of variation, segmented coefficient of variation anomalies, and leakage trend ratios, to systematically characterize diverse failure modes. A logistic regression model was further integrated to construct a quantitative failure risk assessment function. This method was implemented as a failure detection module within a coal mine disaster intelligent early-warning platform. Validation using field data from 582 monitoring points at Zhaozhuang Coal Mine demonstrate that the proposed method achieves an accuracy rate of 91.89%, precision rate of 96.97%, and an AUC value of 0.945 2. Compared to deep learning models based on raw data, the method significantly improves discrimination accuracy, effectively overcoming the challenge of capturing failure patterns under complex mining-induced stress conditions. SHAP-based attribution analysis quantified the marginal contributions of each criterion to failure detection, with feature importance distributions aligning closely with the physical mechanisms of sensor failure. This validated the model’s interpretability and field adaptability. The research outcomes provide critical technical support for the intelligent upgrading of coal mine safety monitoring systems and can be directly applied to on-site maintenance practices.