Development of thin surface virtual sensors for predictive maintenance

Abstract

In the manufacturing field, stamping and plastic injection are among the essential processes, and companies in this sector must optimize them to gain a competitive advantage. In this context, this work is part of the On-Surf Project, which aims to develop advanced solutions for the manufacturing industry through surface engineering techniques. This study proposes the development of virtual sensors capable of estimating, in real time, the temperature at multiple points of a plastic mold injection chamber, regardless of geometry, using data from one or multiple physical sensor inputs. The proposed method is divided into three stages. The first stage employs the Computer-Aided Engineering (CAE) software ANSYS to model the temperature distribution of the injection system using the Transient Thermal module. The second stage uses the mathematical software MATLAB to obtain the virtual sensor equations from the temperature time series generated by ANSYS. The third stage involves applying these virtual sensor equations in a performance test through an algorithm developed in Python, which calculates the virtual sensor values based on the input of one or more physical temperature sensors. The study begins with case analyses of simple metallic geometries to understand their thermal behavior and boundary conditions, and later advances to the analysis of the mold. The method was experimentally validated using simple metallic geometries, plates and tubes made of aluminum and copper, subjected to heat conduction tests in a testing jig. The testing jig is equipped with a datalogger capable of recording the readings of four temperature sensors onto a MicroSD card. For validation purposes, the positions of the four physical sensors coincide with those of the virtual sensors, enabling quantitative and qualitative comparisons. Sensor S1 is positioned on the heat source and serves as the input for the virtual model, while sensors S2, S3, and S4 are used for comparison. In the virtual modeling process, logarithmic and polynomial expressions of orders 4, 6, and 8 were tested, with fourth-order polynomials providing the best balance between computational cost and accuracy. In the tests performed with a topology consisting of one real input and seven virtual output sensors, the expression resolution time was approximately 230 milliseconds (ms). For the topology with two real inputs and seven virtual outputs, the resolution time was around 473 ms. During validation, for all geometries, the inferred value of sensor S2 remained within the range of the measured real values for most of the time interval. The results from sensor S3 showed better performance for the tubular geometries of aluminum and copper compared to the plate geometries, whose inferred values were slightly lower than those measured experimentally. The inferred values of sensor S4 exhibited variations across all geometries, with values slightly below the real measurements. These results are directly related to the conditions of the real measurements, as well as to the configuration of the boundary conditions in ANSYS. The proposed technique shows potential for improvement, providing relevant measurements from an equipment monitoring perspective and generating useful information on usage trends that can contribute to Predictive Maintenance (PdM) strategies.