Life Prediction Model for High-Voltage Power Supply of Exposure Machine

As the core power unit in semiconductor lithography processes, the service life of the high-voltage power supply (HVPS) for exposure machines directly affects equipment utilization rate and production yield. Traditional life assessment mostly relies on empirical values, which have problems such as delayed early warning and blind maintenance, easily leading to sudden shutdowns. According to industry data, shutdowns of exposure machines caused by power supply failures account for more than 35% of total shutdowns, with a single shutdown loss exceeding 100,000 yuan.
Building a life prediction model requires overcoming three core difficulties: first, the coupling of multiple failure factors. The service life of the power supply is affected by capacitor aging (high temperature accelerates electrolyte volatilization), thermal stress of power devices (accumulation of switching losses), and voltage fluctuation impact (deterioration of insulation layer under high-frequency pulses). It is necessary to quantify the weight of each factor; second, the integrity of data collection. It is necessary to real-time monitor 12 key parameters such as output voltage ripple, module temperature, load current fluctuation, and insulation resistance to avoid model deviation caused by missing data; third, dynamic adaptability. The operating conditions of the power supply vary greatly under different lithography processes (e.g., 7nm and 28nm processes), so the model needs to be compatible with multiple scenarios.
The solution adopts a hybrid architecture of "data-driven + mechanism modeling": first, a real-time monitoring network is built through IoT sensors to collect the full-life-cycle operation data of the power supply, and wavelet transform is used to extract health characteristic factors such as capacitor capacitance attenuation and IGBT junction temperature change; second, a basic life function is established based on the failure physics model (e.g., Arrhenius equation describing the impact of temperature on life), and the LSTM neural network is used to correct the prediction deviation under dynamic working conditions, realizing dual calibration of "static mechanism + dynamic data"; finally, through cross-validation (dividing the 3-year operation data of 30 power supplies into training set and test set), the model prediction error is controlled within 8%, which is 60% more accurate than traditional methods.
The application of this model can realize "predictive maintenance": when the system determines that the remaining service life of the power supply is less than 300 hours, it automatically triggers a maintenance reminder, reducing the probability of sudden failures by 70%, and at the same time reducing resource waste caused by over-maintenance. After application in a semiconductor factory, the annual maintenance cost decreased by 22%, and the utilization rate of exposure machines increased from 92% to 96.5%.