Fault Detection of Sensor Data in Semiconductor Processing Using Neural Network with Dynamic Time Wrapping Loss

The monitoring of machine sensor data was of great significance in modern semiconductor wafer manufacturing. These process data were used for early detection of anomalies, facilitating later process adjustments and avoiding further economic loss. In this paper, a deep variational auto encoder neural network model was proposed to improve the accuracy of anomaly detection, reduce manual intervention and adapt to the differences between processes and machines, and at last eliminate the cause of the faults and thus reduce abnormal yield loss. Dynamic time wrapping (DTW) loss was utilized to better reflect sensor data shape change and avoid the affection of inconstancy sampling rates. Experiments carried out on various sensor datasets reveal good performance of DTW compared to models trained with the standard Mean Squared Error (MSE) loss function. Since the anomalies of semiconductor processing were usually reflected as comprehensive abnormality of multi-sensor in same time ranges, the reconstruction loss of all sensors in different stages could effectively identify the positioning of the processing abnormal stage.