부산대학교 도서관

There is a growing need for industrial robots to undertake high tolerance operations in line with Industry 4.0 (I4.0) demands and requirements. This requires robotic accuracy to be understood and quantified within appropriate bounds, relative to their application area. One such area is complex aerospace assembly where robotic precision is of particular importance if the full potential and benefits of I4.0 are to be realised in the sector. Model-free robot error compensation, including machine learning (ML) methods, has demonstrated capability in reducing robot error. Despite this, the variation in ML algorithms tested is limited. Artificial Neural Networks (ANNs) have been the primary ML algorithm investigated for this purpose. However, the application of other potentially suitable ML algorithms for robot error compensation, remains relatively under-researched. All robot error compensation methods have yet to consistently reduce robotic error such that they attain the level of accuracy required in aerospace assembly. They have also not been found to validate their methods when a robot program is repeated continually, despite the fact that this is how the robot would operate on a production line. Further limitations of these approaches surround the lack of a standard approach in identifying the robot base frame, even though this must be established accurately to calculate the robot's actual attained positions accurately. This research, therefore, aims to investigate, and optimise, ML approaches such that they can support an I4.0, autonomous production system, by operating accurately throughout its entire operation, day-to-day. First, the basis for a common approach to the determination of a robot base frame was established. The method integrated a Universal Robot (UR), metrology hardware, including a laser tracking system, to capture robot positions, with a Design of Experiments (DOE) approach to select an appropriate measurement routine. The proposed novel approach was found to increase the repeatability in establishing the RBF origin point by 93.4%, compared to a previous method that used arbitrarily chosen factors. The UR was then tested to verify the need to compensate for both robotic error and positional drift over time. The results from this analysis verified the presence of positional error and drift of the robot. The x-axis returned the lowest mean error value of 0.241mm, and the drift of one point in the x-axis was 0.081mm. The scale of error exceeds that typically required by the aerospace industry (this thesis used a tolerance goal of ±0.2mm), thus confirming the need for error and drift compensation in this robot. This novel investigation found three main robot controller outputs that were correlated with the error; Joint Temperature, Tool Centre Point Forces, and Current. An ANN and non-ANN ML algorithm were investigated as robot error compensation models. This research found that the proposed non-ANN compensation method was capable of reducing robotic error on a similar scale to the ANN model. The success of both algorithms supported their novel adaption to compensate for robot error, with the addition of predicted robot controller outputs as model inputs, to determine if they could improve the models' error prediction, and manage drift. The ANN models consistently stabilised drift. The ANN models returned the greatest reduction in the range of error and mean error. However, no single model variation reduced the error the greatest across each axis. The ANN model with Current as an additional model input reduced the mean error the most in the x and y-axis, by 104.1% in both axes. The ANN model with only positional data and time as the inputs reduced the mean error in the z-axis the greatest, by 99.6%. A goal tolerance of ±0.2mm was set, and it was a non-ANN model (without robot controller outputs used as an additional model input) that attained the greatest number of point-axis combinations that satisfied this tolerance. The success of both the ANN and non-ANN algorithms enforces the need for further research into other ML algorithms as robot error compensation models, and there is also still potential to further optimise the algorithms used. While the addition of predicted robot controller outputs did prove beneficial for the ANN algorithms, their use increased the overall computational cost of the analysis and further work is required to justify their addition. In summary, this thesis addresses key limitations of robot error compensation models and the results may be used to aid in the development of accurate, autonomous, robotic systems in an I4.0 environment.