As modern printing enterprises continuously improve production efficiency, bearings, as crucial rotating core components in printing equipment, are prone to varying degrees of damage during long-term high-speed operation. Failure to promptly detect bearing faults can result in issues ranging from disruptions in the normal operation of the printing equipment, affecting print quality at best, to equipment damage and economic losses, or even personal safety accidents at worst. Therefore, effectively diagnosing bearing faults in printing equipment is crucial to ensuring long-term, high-speed, and reliable operation.
Traditional methods for bearing fault diagnosis primarily involve analyzing collected vibration signals using signal processing techniques. These methods aim to eliminate redundant components and extract information that reflects fault characteristics, thereby achieving bearing fault diagnosis. While these methods initially showed promising results in fault diagnosis research, the increasing demand for production efficiency has led to the collection of massive amounts of signals reflecting equipment operating states. This situation has rendered traditional fault diagnosis methods less effective.
In the actual working conditions of printing equipment, interactions among components within bearings lead to low signal-to-noise ratios in the vibration signals collected by sensors. This reduces the accuracy of feature extraction by deep learning algorithms, thereby limiting the application scope of these algorithms in mechanical fault diagnosis. To address this issue, Jill Group innovatively preprocesses the raw vibration signals of printing equipment bearings using Morlet wavelets, which offer excellent resolution in both time and frequency domains. This preprocessing transforms the signals into two-dimensional time-frequency maps, enhancing the feature information of the original signals. Subsequently, convolutional neural networks are employed to learn the fault features from the time-frequency spectra. This network introduces the concept of decomposed convolution within traditional convolutional neural network models, combining high computational performance with sparse filter-level characteristics. It adaptsively learns useful information reflecting fault features from the time-frequency maps of bearing vibration signals, establishes nonlinear mapping relationships between
