Abstract
In industrial equipment fault diagnosis, different components of various equipment exhibit diverse motion forms and intensities, leading to significant differences in the variation patterns of vibration signal parameters across components. Under interference, both the trend information and the transient information of signals are affected by noise. This results in signal blurring and overlapping, making it challenging to characterize the multimodal nature of the signals. Based on this, the paper proposes a decoupled learning with a reduced convergence domain (DL-RCD) method. First, an innovative Haar wavelet attention operator is introduced, which decouples different frequency features in complex pattern signals, effectively narrowing the convergence domain of network operation parameters. Second, the proposed multifeature extraction operator constructs a process of horizontal and vertical functional interaction mapping. This ensures that when extracting coupled features, the model’s long-term operational trends are not disrupted by short-term transient features. Third, the proposed joint measurement model can analyze the correlation between the macro characteristics and detailed information of the signals. In addition, the constructed scale correction fusion module explores the joint encoding of channel and feature scale dimensions, reducing the negative impact of interference information on decision-making. Since DL-RCD enforces the model to focus on specific frequency bands where fault features reside through frequency domain attention, it reduces the influence of irrelevant information and noise, thus maintaining efficiency even with small and diverse datasets. This general model offers greater advantages in industrial scenarios by simplifying maintenance processes and reducing the number of models that need to be managed and updated, especially in situations where edge computing device resources are limited.
Get full access to this article
View all access options for this article.