Improved BLCD and Its Application in Gear Surface Defect Detection

Gears are important components in mechanical transmission, and monitoring their health is crucial for the safe operation of equipment. Since defects that occur during operation are mainly located on the gear surface and can be captured by industrial cameras, conditions are conducive to machine vision online inspection. Currently, research on vision-based online detection methods for gear surface defects is limited, and traditional image decomposition methods (such as Bidimensional Empirical Mode Decomposition, BEMD) are inefficient, which restricts the detection speed of the system. The Bidimensional Local Characteristics-Scale Decomposition (BLCD)proposed by Dongxu improves detection efficiency. However, the issues of boundary effect and mode mixing still exist. In response to the boundary effect and mode mixing issues that arise in the bidimensional image decomposition process using the BLCD method, corresponding improvements are proposed. First, based on the principle of boundary effects, an adaptive image extension method based on the probability density of edge extremum points is proposed. Then, referring to methods that solve mode mixing in the EMD approach, three techniques are proposed: Bidimensional Ensemble Local Characteristic-scale Decomposition (BELCD), Bidimensional Complementary Ensemble Local Characteristic-scale Decomposition (BCELCD), and Bidimensional Complete Ensemble Local Characteristic-scale Decomposition with Adaptive Noise (BCELCDAN). BELCD uses multiple white noises with a mean of 0 to mask the interference present in the signal, obtaining a more accurate envelope. BCElCD uses dual complementary noise (such as two sets of perfectly anti-correlated positive and negative noise sequences) instead of single noise. Through the symmetry of the noise, precise cancellation of the noise is achieved during the ensemble averaging process after multiple decompositions.And after BCElCDAN decomposes a first-order IMF component, it immediately performs an averaging cancelation of complementary noise on that component, and then decomposes the next order based on the residual signal, preventing noise from transferring between different order modes and improving the purity of each IMF component. Denoising and detection effectiveness comparison experiments are conducted on gear surface defects. Experimental results show that the improved BLCD method is more practical in terms of denoising and detection.