Automatic Borescope Image Analysis | 01
Borescope inspections (BSI) is a critical routine process for monitoring the aircraft engine heath status. BSI technology allows inspectors to visually inspect inside aircraft engines without opening the engine component, by guiding a small camera into the engine. Inspector will look for any sign of damage and overall condition of the engine. At the end of the inspection, depends on the detected defects and severity of the damage, inspector may recommend to send the engine for repair or approve it to continue its operation with condition to next inspection follow up.
In this process the inspector experience is vital to avoid unnecessary disruption (False Positive) which is a very costly mistake for airlines, or missing detection of damage (False Negative) which has safety implications.
The objective here is to develop image analytics to assist the inspector during inspection.
What I did:
I have used various machine learning and deep learning technologies in this project, including classification models (CNN), Semantic segmentation (FCN , U-net), Object detection (Yolo) and image restoration (GAN).
The outcome of the project was improving the efficiency of inspection by 30% with more than 98% accuracy ( 92% IOU) for TBC loss, 97% accuracy ( 72% IOU) for crack detection.
Customer Identification | 02
Majority of airline passengers are not part of loyalty program, so Airlines knows a as little as they should about their behaviour, preference and requirements.
The objective of the project is to identify all previous reservation records of the same passenger in the DB. What makes is super difficult are three reasons:
Inherently changing values. Some identifiable data is changing over the course of the person life, for example passport detail.
Multi value data: for example multiple email addresses (personal, work email), multiple nationals, etc.
Non Exclusive data: for example multiple people lives in the same address.
Missing data: There are few data fields that are mandatory, others are optional.
Typo: due to the data entry error, we experience a massive amount of typos. Some of them are genuine spelling mistakes such as Mohammad vs Mohamed.
Non unified format and free form text entry: for example travel documents is a free form text which can holds (multiple) passport detail, expiry date, visa numbers, visa expiry date, etc.
By identifying passengers historical records, first we can identify important passengers, second get 360 view of the passengers, such as most frequent travel origin/destinations, seasonal pattern, purpose of travel, etc.
What I did:
In this project I used de-duplication methodology, fuzzy string matching and NLP technologies to train a machine learning model to predict if two reservation records belongs to the same person or not. The accuracy of the model achieved above 87% with 95% precision.