DAS noise reduction method based on noise type perception reinforcement learning

Abstract

A DAS noise reduction method based on noise type perception reinforcement learning belongs to the technical field of machine learning and geophysical exploration signal processing, and constructs a reinforcement learning agent taking a full convolutional network as a core, and matches optimal noise reduction action for each region by analyzing noisy data and generating a strategy graph, thereby improving the noise type perception reinforcement learning accuracy. The action set is composed of a group of pre-training lightweight noise reduction networks special for specific types of noise. According to the method, a semi-supervised cooperative training strategy is introduced, and supervised rewards based on mean square errors are adopted for labeled synthetic data to ensure signal fidelity; a non-reference quality evaluator DAS-BRISQUE (DBQ) is designed for real data without labels so as to generate non-supervision rewards, and the intelligent agent optimization is guided by combining two reward signals, so that the field gap between the real data and the field gap between the real data can be effectively bridged, the generalization ability of the model in a complex real scene is remarkably improved, and diversification is realized.

Images