Siamese Network Data Preprocessing for Semiconductor Manufacturing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data comparing methods in semiconductor manufacturing facilities face challenges in determining data sameness due to varying data sizes, which can cause Siamese networks to lose learned content when input sizes change.
Innovation Solution
A pre-processing method that divides data according to each process step and converts it to a uniform size using ID convolution, ensuring that the data can be effectively processed and compared using a Siamese network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data sizes are not adjusted before inputting to Siamese network, then data processing can be performed on original data, but the Siamese network loses learned content when input size changes
Solution Approach 1:
The patent applies preliminary action by adjusting data sizes to be uniform before inputting to the Siamese network. The processor performs size adjustment on first data and second data before they are input to the neural network, ensuring that the network maintains its learned content across different input batches. This preprocessing step prevents the loss of learned content that would occur if variable-sized data were input directly.
2Reliability
If data sizes are adjusted to be uniform, then Siamese network can maintain learned content, but data processing time increases due to size adjustment
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the size parameter of input data based on the actual data characteristics. The processor determines the actual size of first data and second data, then adjusts them to a target size that is at least as large as the actual sizes. This parameter adjustment ensures model consistency while minimizing the time penalty by only adjusting to the necessary size rather than a fixed large size.
3Ease of operation
If maximum data size is used for all inputs, then all data can be processed uniformly, but processing efficiency decreases due to handling larger than necessary data
Solution Approach 1:
The patent applies dynamics by making the target data size dynamic rather than fixed. The processor determines the target size based on the actual sizes of the data being processed, setting it to be at least as large as the maximum actual size. This dynamic approach maintains processing uniformity (all data are adjusted to the same target size) while improving efficiency by only allocating the necessary size for each processing batch rather than always using a conservative maximum size.
Data Source
AI summary
The inventive concept provides a method for processing data generated at a substrate treating. The method includes dividing the data according to each process of the substrate treating; and converting the divided data to a same size. In an embodiment, converting the divided data t a same size comprises converting the divided data to the same size using an ID convolution.