Akademik Veri Yönetim Sistemi
Tezin Türü: Yüksek Lisans
Tezin Yürütüldüğü Kurum: Galatasaray Üniversitesi, Fen Bilimleri Enstitüsü, Türkiye
Tezin Onay Tarihi: 2023
Tezin Dili: İngilizce
Öğrenci: Candan ERGELDİ
Danışman: Orhan Feyzioğlu
Özet:
Today, the majority of food created is wasted rather than consumed, which has a negative impact on worldwide hunger and the economy. Improvements to aggregate supply chains are at the forefront of the actions needed to meet the nutritional requirements of an expanding population. One of such improvements noted in this research was aggregate food storage. The ESP8266-Microcontroller, along with the DHT11 temperature and humidity sensor and the MQ3 alcohol sensor, is put in the storage area to measure the storage conditions of fruit products on a regular basis. The data gathered is sent to the IoT Application in AWS cloud computing service via the microcontroller and MQTT communication protocol and is stored in both the S3 Bucket and Firehose Kinesis databases using the rules defined in this console. As result, the sensor data stored in the database is examined using AWS-IoT-Analysis. When the temperature rises, the crops begin to decompose. Accordingly, a rule in the AWS IoT Application is defined to fire with out-of-range measurements, and the AWS Simple Notification Service is triggered to send ambient temperature, humidity, and methanol values to user via SMS and e-mail. A Convolutional Neural Network model was developed to classify fruits based on their variety and whether they are fresh or rotten using binary and multiclass classification. Models was first taught using images of 1693 fresh apples, 1581 fresh bananas, 1466 fresh oranges, 2342 rotten apples, 2224 rotten bananas, and 1595 rotten oranges over 50 epochs. Then, images of 395 fresh apples, 381 fresh bananas, 381 fresh oranges, and 388 rotten apples, 601 rotten bananas, and 530 rotten oranges were evaluated. Models had a training accuracy of 99.15% and an assessment accuracy of 97.59% for the binary and training accuracy of 99.59% and training loss of 1.27% and a testing accuracy of 98.99% and testing loss of 2.32% over 50 epochs both for the multiclass models.