基于知识蒸馏的回归方法在地震层位追踪中的应用

Application of regression method based on knowledge distillation in seismic horizon tracking

近年来深度学习方法在地震层位追踪上得到了广泛应用,但其存在计算量大、占用内存多等缺点.针对上述问题,通常采取知识蒸馏方法进行优化,其主要利用学生模型模仿教师模型的输出,将复杂的教师模型压缩为简单的学生模型,降低了神经网络模型的参数量和计算量,并使模型保持较高的精度.然而,现有的知识蒸馏方法大多侧重于分类任务,并要求教师和学生模型体系结构相似.为了突破上述限制,本文构建了一种用于回归任务的知识蒸馏方法,该方法首先使用不同架构的学生模型和教师模型,其教师模型和学生模型分别采用长短期记忆(LSTM)循环神经网络和卷积神经网络(CNN)进行构建,然后通过LSTM和CNN来构建一个新的目标函数,该方法的关键是将CNN预测结果和LSTM预测结果之间的差异作为知识蒸馏信息,从而得到适合回归任务的地震层位追踪模型,最后利用该模型进行层位追踪.经合成数据和实际数据测试,结果表明,本文的方法比基于CNN的层位追踪精度明显提高,且达到LSTM相当的层位追踪效果,验证了基于知识蒸馏的地震层位追踪方法的可行性和有效性.

In recent years,deep learning methods have been widely used in seismic horizon tracking,but they have disadvantages such as large computation and memory consumption.To address these problems,knowledge distillation methods are usually adopted for optimization,which mainly use the student model to imitate the output of the teacher model,compress the complex teacher model into a simple student model,reduce the number of parameters and computation of the neural network model,and keep the model with high precision.However,most of the existing knowledge distillation methods focus on classification tasks and require similar architectures of teacher and student models.To break through the above limitations,this paper constructs a knowledge distillation method for regression tasks,which first uses student models and teacher models with different architectures,and its teacher models and student models are constructed using Long Short-Term Memory(LSTM)recurrent neural networks and Convolutional Neural Networks(CNN),respectively.is to use the difference between CNN prediction results and LSTM prediction results as knowledge distillation information to obtain a seismic horizon tracking model suitable for the regression task,and finally use the model for horizon tracking.After the synthetic data and actual data testing,the results show that the method in this paper has significantly improved the precision than the CNN-based horizon tracking and achieved the comparable horizon tracking effect of LSTM,which verifies the feasibility and effectiveness of the knowledge distillation-based seismic horizon tracking method.