【作者】 李响；

【导师】 周洪超；

【作者基本信息】 山东大学 ， 电子科学与技术， 2019， 硕士

【摘要】 随着信息化时代的到来,电子设备以及成为人们生活中必不可少的一部分。作为一项独立的工业门类,电子设备的生产已经成为带动国家经济发展的引擎。人们对电子设备的数量与质量都有着飞速上涨的需求。电子设备生产环节中需要大量的元器件焊接工作,焊接的质量直接影响了产品的质量、性能和成本。因此焊点的质量检测工作直接影响到了产品的质量与生产效率。但目前为止出现的焊点质量检测方法都依赖大量的人力,不但检测效率低下,而且需要企业承担巨额劳务开支。近年来互联网企业的快速发展使得传统的实体产业面临更为严峻的生存环境,作为劳动密集型企业的典型,电子生产厂商之间的竞争更是愈加激烈。工厂不得不降低生产成本、提升市场竞争力,不断向着自动化、智能化方向发展。因此企业迫切希望一种全新的廉价高效的焊点质量检测方法。人工智能技术的发展给这一问题提供了全新的解决思路,尤其近年来伴随着计算机计算能力的指数增长,人工智能在许多领域都已超过人类,尤其在大量的简单重复性工作方面。使用人工智能目标检测算法替代人力完成焊点的质量检测工作无疑是一种高效并且可行的方法。因此本文致力于研究一种基于深度学习目标算法的焊点质量检测方案,其应具有高效率、高准确率、低成本、使用方便、无须专门对工人进行培训等优势。为此,本文首先联系电子生产厂家得到焊点数据来源,构建了焊点质量检测数据集,然后针对焊点质量检测任务的要求与指标,使用自己构建的数据集改进了Faster R-CNN的训练过程,使其能够达到工业焊点质量检测的标准,随后又对算法实地部署过程中出现的问题进行分析,与生产厂家一同研究讨论,最终设计出一种完美贴合现有生产线的算法部署方案。为利用工业生产过程中无穷尽的数据来优化算法,本文提出了算法的增量学习方法,利用源源不断的新数据不断完善数据集使其更加贴合真实的数据分布,从而不断提升检测算法的准确率。更多还原

【Abstract】 With the advent of the information age,electronic devices have become an indispensable part of people’s lives.As an independent industrial category,the production of electronic equipment has become the engine that drives the country’s economic development.There is a rapidly rising demand for the quantity and quality of electronic devices.A large number of components are required for welding in the production of electronic equipment.The quality of the welding directly affects the quality,performance and cost of the product.Therefore,the quality inspection of solder joints directly affects the quality and production efficiency of the products.However,the solder joint quality inspection methods that have appeared so far rely on a large amount of manpower,which not only has low detection efficiency,but also requires enterprises to afford huge labor expenses.In recent years,the rapid development of Internet companies has made the traditional physical industry face a more severe living environment.As a typical example of labor-intensive enterprises,the competition among electronic manufacturers is even more intense.The factory has to reduce production costs,improve market competitiveness,and continue to develop toward automation and intelligence.Therefore,the enterprise is eager for a new cheap and efficient solderjoint quality inspection method.The development of artificial intelligence technology has provided a new solution to this problem.Especially in recent years.with the exponential growth of computing power of computer,artificial intelligence has surpassed humans in many fields,especially in terms of a lot of simple repetitive work.It is undoubtedly an efficient and feasible method to use artificial intelligence target detection algorithm instead of manpower to complete the quality detection of solder joints.Therefore,this paper is devoted to the study of a solder joint quality inspection method based on the deep learning target algorithm,which should have the advantages of high efficiency,high accuracy,low cost,easy to use and no special training for workers.To this end,this paper contacted the electronics manufacturer to obtain the solder joint data source,built the solder joint quality inspection dataset,and then improved the Faster R-CNN training process using our own dataset to achieve the solderjoint quality inspection task requirements and indicators,and then analyzes the problems that arise during the field deployment of the algorithm,and studies with the manufacturer to finally design an algorithm deployment solution that perfectly fits the existing production line.In order to optimize the algorithm with the infinite data in the industrial production process,this paper proposes an incremental learning method for the algorithm,which uses continuous new data to continuously expansion our dataset to make it more consistent with real data distribution,thus improving the accuracy of the detection algorithm rate continuously.更多还原