Deep learning-based intelligent fault diagnosis and prognosis of rotating machinery in an automated food production line**

加拿大公司温哥华冻干有限公司生产各种类型的冻干食品，包括水果，蔬菜，乳制品和营养成分。这些产品在大多数超市和零售商店都能买到。在满足加拿大的需求后，它们也被出口。为了应对激烈的全球竞争和经济挑战，公司计划在保持高标准的产品质量和操作安全的同时提高生产率。自动化生产线中旋转部件的意外故障降低了生产效率，引起了安全问题。拟议的研究将开发和实施将解决这一问题的机电一体化技术。具体而言，该项目将开发一种改进的机器状态监测系统，采用合适的传感、信号处理、故障分类和预测技术。潜在的挑战包括各种机械故障和操作条件；机器退化过程分析建模的复杂性和不可行性；具有感知误差和噪声的数据预处理；基于小数据集的深度神经网络模型构建与训练并在在线监测中进行实时检测和预测。有计划的研究活动将克服这些挑战。特别是，深度神经网络的智能端到端学习将解决建模问题；智能传感器融合将解决精度和可靠性的感官信息，提高鲁棒性；卷积神经网络、循环神经网络结构和具有共享模型参数的迁移学习将解决小数据集的训练问题；基于图形处理单元（GPU）的处理将克服实时应用的计算速度问题。项目成果将包括一个包括多个传感器和数据采集设备的传感系统、改进的故障诊断和预测模型、实验验证和技术转让。由此产生的加拿大工业的经济优势将是显著的。

The Canadian company Vancouver Freeze Dry Ltd. manufactures various types of freeze dry food items including fruits, vegetables, dairy products, and nutraceutical ingredients. The products are available in most supermarkets and retail stores. They are also being exported after meeting the Canadian demand. To meet the stiff global competition and economic challenges, the company plans to increase the productivity while maintaining high standards of product quality and operation safety. The unexpected failures of the rotating components in the automated production line have decreased the productivity and caused safety concerns. The proposed research will develop and implement mechatronic technologies that will solve this problem. Specifically, the project will develop an improved machine condition monitoring system with suitable sensing, signal processing, fault classification and prediction technologies. The underlying challenges include the various machinery faults and operation conditions; the complexity and even infeasibility of analytical modeling of the machine degradation processes; data preprocessing with sensing errors and noise; deep neural network model construction and training with small data sets; and real-time detection and prediction in online monitoring. Planned research activities will overcome these challenges. In particular, intelligent end-to-end learning of deep neural networks will address modeling problems; intelligent sensor fusion will address precision and reliability of sensory information, with improved robustness; convolutional neural networks, recurrent neural network structures, and transfer learning with shared model parameters will solve the problems of training with small data sets; and graphical processing unit (GPU)-based processing will overcome the computing speed problem for real-time application. The project outcomes will include a sensing system including multiple sensors and data acquisition devices, improved fault diagnosis and prognosis models, experimental verification, and technology transfer. The resulting economic advantage for the Canadian industry will be significant.