Research On Induction Heating Temperature Control Without A Temperature Sensor

In the industrial induction heating process, a common thermocouple or a thermal resistance with a metal casing is easily heated in the electromagnetic field, which seriously affects the measurement accuracy. To get a high measure precision, a bare thermocouple should be spot on the work piece. Spot welding is time-consuming and laborious, which cannot be applied in the industrial production. A radiation pyrometer can realize non-contact measurement. But the emissivity change is caused by the oxide skin generating on the metal surface at a high temperature, and the radiation pyrometer measurement produces error.In this paper, a NARX neural network model is introduced to predict the temperature of the induction heating work piece, and is applied in PID temperature control without a temperature sensor.The induction heating system is composed of a medium frequency power supply, an electrothermal capacitor, a transformer, a coil, and a20#steel pipe. The controller is an Advantech IPC510and a PCL812PG card. The impact factor of the induction heating temperature was analyzed, and the input and output parameters of the neural network was determined. The time series tool of Matlab2010b was used to train and simulate the NARX neural network. The training mode is series-parallel. The network input is delayed actual DC current and delayed actual temperature, and the output is the next temperature. The simulation mode is parallel. The input is delayed actual DC current and delayed network output, and the output is the next temperature. The neural network model was trained by samples of a step signal and a sinusoidal signal, and was respectively tested by samples of a constant signal, a sinusoidal signal, a step signal and a PID control signal. It was approved that the network trained by the sinusoidal signal generalizes better than the one trained by the step signal, so the sinusoidal network was used in PID control.Real-Time Workshop (RTW) was employed to establish a real-time hardware-in-the-loop simulation control system. The temperature was set from32℃to1000℃, and steady at1000℃. DC current of the medium frequency power supply was collected by IPC, and input to the NARX neural network model. Based on the error of the prediction temperature and the set temperature, PID controller adjusted output power of the medium frequency power supply, and therefore the work piece temperature was controlled without a temperature sensor. The control result is as follows:the error of the prediction temperature and actual temperature is large in the initial rising stage with the maximum168℃; the error in later rising stage is small with the maximum38℃; in the steady stage the error changes from large to small, and then from small to large, with the maximum±25℃. From the trend of the actual temperature, the error is likely to continue to increase after180s. Three reasons cause the error. One is that the initial state of the training model is different from the one of the actual control, which results in the large error in the initial rising stage. The other is that training samples are fewer, and the error will be larger in longer control time. The third is that neural network structure parameters need to be optimized.From the study of the subject, it is proved that the NARX neural network prediction model can be used in PID temperature control in induction heating, and achieve accuracy of±2.5%in the steady stage. This method can simplify device complexity, reduce cost of equipment, and propose a way of temperature control without a temperature sensor.