Using Faster R-CNN for intelligent fault diagnosis and correction in additive manufacturing

这项研究建立了一个创新的基于运行时间的闭环系统，该系统使用深度卷积神经网络(CNN)技术来监督和纠正加法制造(AM)中的错误，从而提高生产率和质量。目前，AM机器采用熔融沉积建模(FDM)来打印3D对象，并提供经济高效和快速的原型制作和大规模生产。然而，由于无法控制打印路径偏差，操作员经历了不可预测的打印成功率和高故障率。这是由于打印机不准确和不稳定，会产生打印质量不一致且通常不可接受的项目。本研究提出采用先进、快速的R-CNN算法技术来开发一种新颖的目标检测方案来应对这一挑战。该算法将利用与检测网络模块具有联合卷积特性的区域建议网络(RPN)来提供几乎无代价的区域建议。RPN的使用将有助于确定图像中可能包含打印故障的感兴趣区域。之后，算法将继续确定区域中的故障类别，并具体修改区域的大小，以检测准确的位置。最后，自动控制系统将根据R-CNN预测的错误信息调整和校正打印参数，从而允许回收正在打印的3D对象。该算法在本研究工作中的充分发展包括参数初始化、网络权值和偏移量的调整以及打印状态的优化。最终，本研究的成果将显著减少灯丝浪费，提高FDM打印机的生产率。合作伙伴组织将通过改进销售、利润、收入和向客户提供的服务质量，从这个自我诊断系统中受益。这一成果还将有助于3D打印技术扩展到更广泛的工业和商业应用领域，对加拿大经济和商业环境产生积极影响。

This research establishes an innovative runtime-based closed-loop system that uses a deep convolution neural network (CNN) technology to supervise and correct errors in additive manufacturing (AM), boosting both productivity and quality. Currently, AM machines employ fused deposition modelling (FDM) to print 3D objects and provide cost-effective and rapid prototyping and mass production. However, operators have experienced unpredictable levels of printing success and high malfunction rates because of their inability to control printing path deviations. This is due to inaccurate and unstable printers that produce items of inconsistent and often unacceptable printing quality. This research proposes to adopt advanced Faster R-CNN algorithm technology to develop a novel object detection program to address this challenge. This algorithm will use the region proposal network (RPN), which has joint convolutional features with the detection network module, to provide nearly costless region proposals. The use of the RPN will help to determine the interesting areas in the image that may contain printing malfunctions. After that, the algorithm will continue to determine the failure category in the area and specifically modify the size of areas to detect the accurate location. Finally, the automatic control system will adjust and correct printing parameters based on the error information the Faster R-CNN predicted, allowing for the 3D objects being printed to be salvaged. The full development of this algorithm in this research work includes parameters initialization, adjusting the weights and bias of the network and optimizations of printing status. In the end, the outcomes of this research will significantly reduce filament waste and improve FDM printer productivity. The partner organization will benefit from this self-diagnosed system through improved sales, profits, revenue, and the service quality provided to its clients. The results will also help 3D printing technology expand into a wider range of industrial and commercial applications, generating a positive influence on the Canadian economy and business environment.