A Predictive Fault Diagnosis Method for Fire Pump Sliding Bearings Based on the Vision Transformer with Local Mean Decomposition

As urbanization progresses and the building inventory increases rapidly, the reliability of fire protection systems in buildings has become critical for ensuring the safety of people and property. There is an urgent need for the automated detection of fire protection system equipment. To address this challenge, a fault diagnosis method for fire pump systems based on the Vision Transformer with Local Mean Decomposition (LMD-ViT) is proposed. An attention mechanism is applied, specifically targeting sliding bearings, a key underwater component in fire pump systems. In the proposed method, the vibration signal is first smoothed, and then LMD is used to enhance the quality of frequency-domain signal analysis. Subsequently, leveraging the attention mechanism, self-attention and cross-attention mechanisms, along with a weight-sharing mechanism, are introduced to further extract fault feature information from the signal decomposition diagram. This supports the accurate identification of wear faults in sliding bearings, which is 98.8%. Moreover, the relationship between the wear of sliding bearings and the head performance of fire pumps is investigated. By classifying the severity of wear, corresponding head performance metrics can be derived. This approach enables the predictive diagnosis of fire pumps, which is also a promising solution in various related applications.