This paper represents the adaptive multiple subtraction problem as a blind signal separation problem using multi-traces convolutional signal blind separation model. By expressing the difference between the predicted and true multiples using a 2D convolutional kernel, the authors propose an adaptive multiple subtraction method based on the multi-traces convolutional signal blind separation technique, which adopts maximization of the non-Gaussianity of the recovered primaries as the objective function. To solve the above non-linear optimization problem, we transfer it to an iterative linear one, which is realized by the iterative least squares algorithm. Taking advantage of the multi-traces convolutional signal blind separation model, the proposed method is applicable to the situation that there are differences in the time-space domain between the predicted and true multiples.