Abstract:

To address the problems that vibration signal of flip-chip defect detection in avionics equipment is easily affected by noise， and the defect characteristics are not obvious， a Convolutional Sparse Coding （CSC） method based on locally reweighted is proposed to reconstruct and de-noise. In this study， CSC model is used to globally represent the vibration signal， so as to avoid the problem of high dictionary dimensions and effectively reduce the computational complexity of training dictionary and sparse decomposition. Secondly， to solve the problem of different sparsity of flip-chip vibration signals， a reweighted CSC model is proposed. In order to suppress local noise， a local reweighted CSC model is constructed. In the iterative process， the energy entropy is redistributed in the way of weights， and applied to the weighted strategy， which can better match the local block CSC structure. In addition， an effective acceleration strategy of Stochastic Gradient Descent （SGD） is proposed， which uses Anderson acceleration （AA） extrapolation method to accelerate the SGD algorithm. This strategy linearly combines the historical iteration information of the convolution dictionary to accelerate the learning of the convolution dictionary and improve the accuracy of dictionary solution. The results of simulation and actual flip chip vibration signal experiments show that the proposed CSC method can effectively remove noise in flip chip vibration signal， and is more competitive and superior to the existing popular CSC denoise algorithms.

Key words: flip chips, vibration signal, convolutional sparse coding, reweighting, dictionary learning, signal denoising