Domain constrained cascadic multireceptive learning networks for machine health monitoring in complex manufacturing systems

Precise condition monitoring of manufacturing systems is crucial for maintaining efficient industrial production. In practical manufacturing applications, typical components of manufacturing system such as gearboxes and bearings mainly operate under fluctuating conditions, resulting in obvious nonlinear characteristics in the monitored vibration signals. Nonetheless, numerous extant algorithms are crafted based on the stationary presumption that the signal's amplitude and frequency remain constant, failing to reflect the real-world scenarios prevalent in industrial environments. In this research, we propose a domain constrained cascadic multirepetive learning network as a response to this challenge. Initially, we leverage cascadic multireceptive learning modules, multiscale feature aggregation modules, and an adaptive filtering module to establish the feature extractor for acquiring multireceptive and multilevel features from monitored signals. Next, a conditional label regulation loss is devised as the loss function to enhance the model's robustness in complex scenarios. Finally, a domain constrained label adjuster is designed to align the actual labels based on the input data, thereby guiding the feature extractor in learning the domain-invariant feature. Three case studies demonstrate that the DC-CMLN model outperforms seven state-of-the-art algorithms, particularly when applied to mechanical datasets collected under nonstationary conditions.