Cluster Analysis for Ultrasound Part Sorting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional pattern recognition systems, such as the VIPR system, are less effective in processing parts with large process variations, leading to the grouping of dissimilar good parts into one large group, which can result in the inclusion of bad parts within the good group.
Innovation Solution
A cluster analysis method and apparatus that utilizes a training set to form smaller subsets or clusters of parts, using broadband signatures acquired through ultrasound scanning, to improve the statistical analysis/pattern recognition tool's ability to distinguish between non-defective and defective parts by creating tighter clusters and enhancing the sensitivity of the sorting process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single large group is used to identify good parts, then all non-defective parts can be accepted, but bad parts may be inadvertently included due to large process variation
Solution Approach 1:
The patent divides the training set into multiple subsets using cluster analysis based on broadband signatures. Instead of treating all good parts as a single homogeneous group, the system segments them into distinct clusters that reflect natural variations in the manufacturing process. This segmentation allows the pattern recognition tool to establish multiple reference profiles, improving the ability to distinguish between good and bad parts while maintaining high acceptance rates for non-defective parts.
2Device complexity
If conventional VIPR system groups all good parts into one large group, then the sorting process is simple, but the variation in good parts obscures the differences between good and bad parts
Solution Approach 1:
The patent performs cluster analysis as a preliminary step before applying the pattern recognition tool. By pre-processing the training data to identify and segment natural clusters, the system prepares optimized reference groups that enhance the subsequent classification accuracy. This preliminary action ensures that when the VIPR system compares unknown parts against the training set, it is comparing against multiple refined profiles rather than a single broad category, thereby detecting subtle differences between good and bad parts more effectively.
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
The cluster analysis method improves the accuracy of the statistical analysis/pattern recognition tool by creating smaller subsets that enhance the ability to accept all non-defective parts and reject defective parts, reducing the impact of process variation and improving the overall sorting performance.
Implementation Method 1
The data collected for each part includes the frequency of a resonance or peak, and the values for that peak, such as amplitude, zero-crossing width, etc.
Implementation Method 2
broadband signatures acquired through ultrasound scanning
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for classifying an unknown part includes acquiring a broadband frequency response for a plurality of parts in a training set of parts, the training set of parts including a plurality of non-flawed parts and a plurality of flawed parts, performing a statistical analysis on the broadband frequency responses to form a plurality of part subsets, the plurality of part subsets including at least one subset of non-flawed parts and at least one subset of flawed parts, and utilizing the plurality of part subsets to form a blended subset of parts, the blended subset of parts being used to classify an unknown part as either a defective part or a non-defective part. A tool for implementing the method is also described.