An Industrial Signal Generation Method Based on Lightweight Conditional Diffusion Model
摘要
深度学习模型为缓解数据不均衡问题提供了有效途径,然而,现有方法难以兼顾生成质量与计算效率,例如生成对抗网络存在训练不稳定问题、扩散模型的计算复杂度较高.为提高工业信号的可用性,提出基于轻量化条件扩散模型的工业信号生成方法,在保持扩散模型生成优势的同时显著提高训练效率,并降低模型推理计算量.该方法主要分为自编码器信号降维与扩散模型样本生成两个阶段.阶段一基于自编码器建立原始信号的隐变量分布结果,构建轻量化的低维空间,以提高整体模型的计算效率;阶段二通过条件扩散模型的马尔科夫链噪声推理过程,对阶段一的输出结果进行重建,以获取高质量的工业信号.针对轻量化空间取值有限导致的类别混叠问题,以聚合同类并区分异类为目标,设计对比学习损失来提高空间中的类别区分度.考虑到故障幅值分布特征表现为周期性衰减冲击,提出分解损失函数来挖掘信号隐藏特征并提升生成样本质量.在两种数据集上的实验结果表明,该方法能够为工业系统提供可靠的数据支持.
Deep learning models provide an effective approach to alleviating the problem of data imbalance.However,generation quality and computational efficiency are difficult to balance in existing methods.For example,generative adversarial networks suffer from unstable training,while diffusion models require high computational complexity.To improve the usability of industrial signals,an industrial signal generation method based on a light-weight conditional diffusion model is proposed.The generation advantages of diffusion models are retained.Mean-while,training efficiency is significantly improved,and the computational cost of model inference is reduced.The proposed method mainly consists of two stages:Signal dimensionality reduction using an autoencoder and sample generation using a diffusion model.In the first stage,the latent variable distribution results of the original signal are learned by an autoencoder.A lightweight low-dimensional space is then constructed to improve the computational efficiency of the overall model.In the second stage,the output result of the first stage is reconstructed through the Markov-chain-based noise inference process of a conditional diffusion model.High-quality industrial signals are thereby generated.The limited range of values in the lightweight space may cause class confusion.To address this problem,a contrastive learning loss is designed to aggregate samples from the same class and separate samples from different classes.Consequently,class discriminability in the lightweight space is improved.Fault amplitude distribu-tions are characterized by periodic decaying impulses.Therefore,a decomposition loss function is proposed to cap-ture hidden signal characteristics and improve the quality of generated samples.Experimental results on two data-sets demonstrate that the proposed method can provide reliable data support for industrial systems.