To achieve high productivity and product quality, we are deriving ideal filling curves for controlling the ladle tilting speed, plunger injection speed, and filling pressure based on physical laws or numerical analysis. However, since multi-step velocity input type casting machines are widely used in many foundries, it is necessary to convert the derived filling curves into discrete values (panel input data) consisting of time and velocity (or pressure) data to be input as target values to the pressurization control panel of low pressure casting machines. Another problem that can be encountered is that the output could differ from the pre-calculated behavior due to delay and error. Conventionally, research has been conducted to construct input-output models that take into account the delay and error of the output relative to the control input. However, with multi-step velocity input type casting machines, the number of data points that can be input to the panel is limited, and partial errors such as, behavior that differs from the pre-calculation, still result. In recent years, digital transformation (DX) approaches are increasing, and research studies are being conducted to construct input-output models based on deep learning. In this research, we constructed a model that predicts the output from the panel input data using deep learning for a multi-step velocity input type casting machine. In addition, for errors that can occur during the conversion of ideal filling curves to panel input data, we propose a method of determining the panel input data for achieving operations based on the ideal filling curves on the actual machine with a limited number of panel input data. Finally, the effectiveness was verified by conducting air pressure control experiments on a water experimental unit for low pressure castings.
