基于级联回归模型的电力设备异常状态图像检测算法研究

This application proposes a ‘cascading learning strategy combining deep learning with decision trees’ to achieve image defects detection of power equipment. In order to solve the problems of few samples and weak model generalization ability in Power Domain, we introduce cascade regression model, which has the advantages of low parameter adjustment dependency, flexibility, easy expansion and theoretical analysis in contrast to traditional ways. Firstly,the attention mechanism of inspection tour is simulated from the perspective of computer, and the effects of different features and feature combinations on visual attention are analyzed by several feature selection methods. Then,based on the regression model of deep neural network, starting with improving the learning ability of feature expression, reducing the computational complexity and reducing the dependence of parameter adjustment, a cascade regression model based on probability function and decision forest is proposed, which is applied to the task of small sample image data set. Combining the advantages of the two methods, cascading learning network is proposed to reduce the number of hyper-parameters significantly, and a defects detection system of power equipment is constructed. At last, adaptive to the characteristics of complex substation scenarios, numerous equipment and constantly updating, a model generation method that automatically performs feature extraction and train learning is proposed by this project. It sets up a neural architecture search with reinforcement learning applying for power-specific hardware deployment environments.

电力设备运行状态的异常检测是实现电力智慧运维的关键核心技术之一，本项目针对异常图像样本数量少、模型泛化能力弱等问题，研究深度学习和决策森林相融合的级联学习策略，以实现电力设备运行状态的异常检测。主要研究内容包括：结合电力巡检业务流程，模拟运维人员视觉注意力机制，研究不同层次有效特征的提取机制；研究决策函数与深度神经网络相融合的特征学习方法，大幅减少超参数数量，建立针对小样本数据集的级联回归模型；研究自动进行特征提取、学习训练的模型自寻优技术，建立多任务、多节点自学习框架和架构搜索方法，提高模型的泛化能力；面向电力特定硬件部署环境，研究针对训练数据的自动扩充与采样搜索方法，以满足特定硬件平台对模型的低延迟、低功耗要求。项目研究将为电力系统的安全、稳定运行提供技术支撑。

电力设备运行状态的异常检测是实现电力智慧运维的关键核心技术之一。本项目针对电力智慧运维中异常图像样本数量少、模型泛化能力弱等多场景共性问题，研究了深度学习和决策森林相融合的级联学习策略；结合电力巡检业务流程，模拟了运维人员视觉注意力机制，建立了针对小样本数据集的级联回归模型，提出了基于级联回归模型的电力设备异常状态图像检测算法；研究自动进行特征提取、学习训练的模型自寻优技术，建立多任务、多节点自学习框架和架构搜索方法，提高模型的泛化能力；此外，面向电力特定硬件部署环境，设计针对训练数据的自动扩充与采样搜索方法，以满足特定硬件平台对模型的低延迟、低功耗要求，为电力系统的安全、稳定运行提供理论和数据支撑。与此同时，项目的相关成果在电力、公共安全等领域进行了初步转化应用，并取得了良好成效。