传统的工业清洗方式多种多样，主要包含物理清洗方式和化学药剂清洗方式，但是传统工业清洗方式存在着环境污染、清洗效率低等难以解决的问题。在工业科技日益进步、经济快速发展的今天，在国家环境保护法律法规越来越严格的大背景下，传统工业清洗方式逐渐转向一种新型清洗方式，即激光清洗方式。与传统清洗方式相比，激光清洗技术具有绿色、清洗效果佳、应用范围广、精度高、非接触式和可达性好等突出优势。随着激光清洗技术的不断进步，激光清洗方式将会逐渐替代传统工业清洗方式，并广泛应用于工业清洗领域。但由于激光清洗的过程非常复杂且高度非线性，涉及到各种物理化学变化，激光清洗过程的研究始终是一个具有挑战性的问题。所以激光清洗技术相关问题的研究具有重要意义。铁锈作为一种常见的污染物，采用激光清洗是一种常用清洗方法。本文基于深度学习方法，通过激光清洗样本的图像来对激光清洗过程进行建模，建立包含清洗前、激光参数、清洗后效果的神经网络模型，实现激光清洗效果、激光参数的预测。

      论文的研究工作和主要贡献包括：

（1）建立一套完整的激光清洗数据集。目前没有公开的关于激光清洗数据集，而用深度学习方法来研究激光清洗技术，则需要大量数据的支撑。本文建立了一套完整的激光清洗数据集用于研究工作。

（2）提出一种激光清洁度量化指标。监督学习需要训练数据的清洁度标签。因此，本文提出了一种定量的激光清洁度指标用于标记清洁度，清洁度可以通过计算清洗后图像和标准图像的色差值来定量的表示出来。

（3）使用深度神经网络来预测激光清洗清洁度。本文使用深度神经网络对激光清洗过程进行建模。根据清洗前图像、激光参数，应用神经网络预测激光清洗后的样本清洁度，实验结果预测准确度达 82.23%。

（4）使用深度神经网络预测激光参数。最后，本文研究了对于给定的待清洁图像和预期清洁度，应用深度神经网络来预测一套激光参数，对激光清洗过程或工艺的优化起到良好且关键的辅助作用。

Traditional industrial cleaning methods are diverse, mainly including physical cleaning methods and chemical agent cleaning methods, but traditional industrial cleaning methods have problems , such as environmental pollution and low cleaning efficiency. With the increasing development of industrial science and technology and rapid economic development, under the background of increasingly strict national environmental protection laws and regulations, traditional industrial cleaning methods have gradually turned to a new type of cleaning method, that is, laser cleaning. Compared with traditional cleaning methods, laser cleaning technology has outstanding advantages such as green color, good cleaning effect, wide application range, high precision, non-contact type and good accessibility. With the continuous advancement of laser cleaning technology, the laser cleaning method will gradually replace the traditional industrial cleaning methods, and it will be widely used in the field of the industrial cleaning. However, because the process of laser cleaning is very complex and highly nonlinear, and it involves various physical and chemical changes, the research of laser cleaning process is always a challenging problem. Therefore, the research of laser cleaning technology is of great significance. As a common pollutant, rust is usually cleaned by laser cleaning technology. In this paper, based on the deep learning method, the laser cleaning process is modeled by the image of the laser cleaning samples, and a neural network model including pre-cleaning, laser parameters and the cleanliness is established to predict the laser cleaning cleanliness and the laser parameters.

The main work and contents of this thesis are:

Firstly, building a laser cleaning data set. Currently, there is no public data set about the laser cleaning, the use of deep learning requires a large amount of data. In this paper I have built a laser cleaning data set for research work.

Secondly, a method of cleanliness evaluation index is proposed. Supervised learning is a functional machine learning task inferred from the labeled training data. Therefore, in this paper I explore a quantitative laser cleanliness index. Cleanliness can be quantitatively expressed by calculating the color difference between the cleaned image and the standard image.

Thirdly, using deep neural networks to predict laser cleaning cleanliness. In this paper, I use deep neural networks to model the laser cleaning process. Based on the image before cleaning and the corresponding laser parameters, a neural network is used to predict the cleanliness of the laser cleaning. The prediction accuracy of the experimental result reaches 82.23%.

Finally, using deep neural network to predict laser parameters. I study the application of deep neural networks to predict a set of laser parameters for a given image to be cleaned and expected cleanliness. This experiment can effectively optimize the laser parameters and play a key auxiliary role in the optimization of laser cleaning process.