Robotic Cleaning of Fish Processing Plants: Kinematics, vision and Optimization using Artificial Intelligence

Abstract

A robotic cleaning system for fish processing plants has been designed and implemented. The system is based on a custom-made manipulator specifically designed for robotic cleaning of fish processing lines. This manipulator is mounted on a horizontal rail, and is designed to have the required reach and mobility, and to withstand the harsh environment characterized by high humidity and aggressive chemicals. The development includes a vision system for robotic inspection of the cleaning quality. Methods from artificial intelligence, such as genetic algorithms and convolutional neural networks were successfully implemented for the robotic system. The robotic cleaning system is evaluated in extensive experiments in an industrial setting. The experiments include a comparison with human operators where it was demonstrated that the robotic system performed as well as a human operator with 15 years of experience. This thesis presents the mechanical design and control of the custom-made robotic manipulator. The inverse kinematics is also presented. A new genetic algorithm approach to optimize the mechanical design of robotic manipulators is presented in this thesis. The proposed algorithm is tested in experiments using the 6DOF custom manipulator. Two variations of the genetic algorithm are implemented, and optimization with respect to payload, reach, weight and stiffness is performed. The results show that the proposed algorithm converges quickly towards a near-optimal solution. The proposed algorithm enables a faster and more optimal design process of robotic manipulators. The proposed algorithm also offers the possibility to optimize the design with regards to parts of the workspace of the robotic manipulator. A vision system for quality assessment of the cleaning is proposed in this thesis. The vision system uses convolutional neural networks to detect residual fish blood on cleaned surfaces. The performance of different convolutional neural network architectures and parameters are evaluated. Data sets that simulate various conditions in fish processing plants are generated using data augmentation techniques. Tests using further augmented training data to increase the performance of the neural network are performed, which results in a substantial increase in performance both compared to a baseline color thresholding technique, and also compared to the same neural network architecture without augmented training data. A proposed approach for predicting remaining useful life based on deep learning and genetic algorithm is also presented in this thesis. The proposed approach achieves state-ofthe-art result on the Commercial Modular Aero-Propulsion System Simulation dataset.