Atypical Manufacturing Text Processing for Explainable Failure Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models struggle with atypical text data from production sites, leading to reduced understandability and accuracy in failure identification due to mismatched formats and irrelevant keywords, making it difficult to accept outputs at the site.
Innovation Solution
A system that utilizes region knowledge to select specific parts of speech, such as nouns and adjectives, from atypical text data, excluding insignificant phrases, and trains an explainable AI model to improve readability and accuracy of failure identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If machine learning models are trained on atypical text data from production sites, then the model can process real-world data, but the understandability and accuracy of failure identification deteriorates due to mismatched formats and irrelevant keywords
Solution Approach 1:
The patent segments the text data processing into distinct stages: morphological analysis to identify parts of speech, selection of relevant phrases based on part-of-speech filtering, and training with cleaned feature amounts. This segmentation allows the model to handle atypical text data while improving accuracy by processing it through structured steps that eliminate irrelevant information.
Solution Approach 2:
The patent extracts and removes irrelevant keywords and phrases from the atypical text data before training. By using morphological analysis to identify and filter out insignificant parts of speech, the system extracts only the relevant features needed for accurate failure identification, thereby improving measurement precision while maintaining adaptability to real-world data.
2Quantity of substance
If all phrases from atypical text data are used as feature amounts, then the model captures more information, but the readability and relevance of explanations deteriorates due to inclusion of insignificant phrases
Solution Approach 1:
The patent extracts and removes insignificant phrases from the feature amounts by using morphological analysis to identify and filter out parts of speech that do not contribute meaningfully to failure identification. This extraction process maintains the quantity of relevant information while improving the readability and relevance of explanations.
Solution Approach 2:
The patent applies different quality standards to different parts of the text data. By using part-of-speech filtering, the system treats different words differently based on their linguistic properties, keeping only those with local quality characteristics that are relevant to failure identification, thereby improving explanation readability.
3Ease of manufacture
If natural language processing is applied to decompose memo content, then the extraction of words improves, but the accuracy of failure identification deteriorates when keywords have no direct relevance
Solution Approach 1:
The patent applies local quality filtering by using morphological analysis to identify and select only those words with specific part-of-speech characteristics that are relevant to failure identification. This selective approach improves measurement precision by ensuring that only locally relevant keywords are used, while the decomposition process remains automated.
Solution Approach 2:
The patent incorporates feedback mechanisms where the system iteratively refines the selection of keywords based on their relevance to failure patterns. The training process uses labeled data to feedback and adjust the feature selection, ensuring that only keywords with direct relevance to failure identification are retained, thereby improving accuracy.
Data Source
AI summary
A non-transitory computer-readable recording medium having stored therein an information processing program for causing a computer to execute a process including selecting a phrase corresponding to a specific part of speech from among phrases extracted from atypical text data in a manufacturing process of a product, and performing training of a machine learning model that outputs a determination result corresponding to an input feature amount and a feature amount contributing to the determination, using training data that associates the input feature amount including the selected phrase and configuration information of the product with label information indicating the determination result regarding the product.