Interpretable multiple-sensor spectra fusion methodology for constructing dual-weight composite health indicator

Machine inevitably exhibits degradation characteristics during operation. Once machine degradation reaches a critical level, it may lead to a severe machine failure. To avoid any unexpected machine failure, Prognostics and Health Management (PHM) technology has been widely adopted, leveraging online sensor data for real-time performance monitoring and preventive maintenance. Performance degradation assessment serves as a foundation of the PHM; however, most existing studies primarily focus on single-sensor applications. The effectiveness of single-sensor modeling is highly dependent on the sensitivity of sensor data, and it is further influenced by factors such as sensor type, installation position, and environmental conditions. To overcome these limitations, this paper proposes a multiple-sensor spectra fusion methodology for constructing a dual-weight composite health indicator. The proposed model simultaneously incorporates frequency weights and sensor weights, optimizing both through a dual-weight optimization framework. A stepwise optimization strategy is introduced, breaking the optimization process into two manageable stages. Our proposed methodology was verified through two illustrative experiments using gearbox run-to-failure vibration data. Results demonstrate that optimized frequency and sensor weights align well with fault characteristics and sensor importance, confirming the physical interpretability of the optimization process.