Ultrasonic Defect Classification Using Siamese Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current ultrasonic non-destructive testing methods rely heavily on human operators for defect detection and classification in semiconductor wafers and chip packages, which is labor-intensive and requires significant training, while existing automated algorithms often fail to accurately identify defects.
Innovation Solution
Employing deep neural networks with 3D image tensors and Siamese Neural Networks to analyze ultrasonic scan data, utilizing cross-correlation layers and incorporating reference images and parameter graphs to enhance defect detection and classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If human operators are used to detect and classify defects, then detection accuracy can be improved through experience and training, but labor intensity and training requirements increase significantly
Solution Approach 1:
The patent replaces the mechanical human visual inspection system with an automated machine learning system. The neural network model processes ultrasonic scan data to detect and classify defects, substituting human operators with an automated computational system that maintains high detection accuracy while eliminating labor-intensive manual inspection
Solution Approach 2:
The system enables self-service defect detection through automated machine learning models that process ultrasonic data without human intervention. The neural network automatically learns defect patterns from training data and independently performs detection and classification, reducing dependency on trained human operators
2Extent of automation
If existing automated algorithms are used for defect detection, then automation level increases, but detection accuracy and reliability deteriorate
Solution Approach 1:
The patent changes the parameters of the automated detection system by implementing a neural network model with specific architectural parameters (convolutional layers, pooling layers, fully connected layers). This transforms the automated algorithm into a more reliable system capable of learning complex defect patterns from training data, improving both automation level and reliability simultaneously
Solution Approach 2:
The system performs preliminary training action by pre-training the neural network model with labeled defect data before deployment. This preliminary learning phase enables the automated system to acquire defect recognition capabilities in advance, ensuring reliable defect identification when the system operates in production mode
3Productivity
If traditional inspection methods are used, then system complexity remains low, but productivity and detection capability are limited
Solution Approach 1:
The patent segments the defect detection task into distinct computational stages: data preprocessing, feature extraction through convolutional layers, classification through fully connected layers, and result output. This segmentation enables the complex neural network to process ultrasonic data efficiently, improving productivity while managing system complexity through modular architecture
Solution Approach 2:
The system transitions from traditional 2D image analysis to processing ultrasonic scan data in additional dimensions (depth, time, multiple scan angles). This dimensional expansion enables more comprehensive defect detection and classification capability, increasing productivity despite the increased complexity of handling multi-dimensional data
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Automates the detection and classification of defects with high precision, reducing the need for human intervention and improving the reliability of defect identification in semiconductor wafers and chip packages.
Implementation Method 1
A mechanical positioning system can synchronously scan an ultrasonic transceiver transducer, and a receiver transducer on the opposite side of a device under test (DUT). The reflected signal back to the transceiver from the DUT is a TDR, time domain reflection, waveform
Data Source
AI summary
In an automated defect detection and classification system, one or more computing devices access scan data acquired in an ultrasonic scan of an object. A first input feature map, including a two-dimensional (2D) scan image, is built from the scan data and input to a first deep neural network to generate a first output feature map. A second input feature map, including an image of a defect-free object, is input to a second deep neural network, having the same structure and weight values as first deep neural network, to produce a second output feature map. The scanned object is determined to contain a defect when a distance between first and second output feature maps is large. In an alternative approach, the 2D scan image and one or more images of the defect-free object are input to different channels of neural network trained using color images.