Geri Dön

Elektrikli araç motorlarında mekanik arızaların derin öğrenme ile tespiti için yeni bir yöntemin geliştirilmesi

Development of a new method for the detection of mechanical faults in electric vehicle motors with deep learning

PDF İndir

  1. Tez No: 923225
  2. Yazar: EYUP SÖNMEZ
  3. Danışmanlar: PROF. DR. SEZGİN KAÇAR
  4. Tez Türü: Doktora
  5. Konular: Mekatronik Mühendisliği, Mechatronics Engineering
  6. Anahtar Kelimeler: Belirtilmemiş.
  7. Yıl: 2024
  8. Dil: Türkçe
  9. Üniversite: Sakarya Uygulamalı Bilimler Üniversitesi
  10. Enstitü: Lisansüstü Eğitim Enstitüsü
  11. Ana Bilim Dalı: Mekatronik Mühendisliği Ana Bilim Dalı
  12. Bilim Dalı: Belirtilmemiş.
  13. Sayfa Sayısı: 150

Özet

Bu tezde, elektrikli araç motorlarında, özellikle de endüstriyel ve otomotiv uygulamalarında yaygın olarak kullanılan BLDC motorlarında arıza teşhisi için yapay zekâ tabanlı yeni bir yöntem geliştirilmiştir. Elektrikli araçların kullanımının artmasıyla birlikte, elektrik motorlarının güvenilirliğini ve güvenliğini sağlamak amacıyla doğru ve erken hata teşhisi büyük önem taşımaktadır. Bu çalışma, WDD-CNN adlı, bir boyutlu ve iki boyutlu evrişimli sinir ağı (1D-CNN ve 2D-CNN) yapısını birleştiren hibrit bir model önererek bu ihtiyaca yanıt vermektedir. Model, titreşim sinyallerini analiz ederek farklı hata türlerini çeşitli çalışma koşullarında erken ve doğru bir şekilde tespit etmeyi amaçlamaktadır. WDD-CNN modeli, yüksek gürültü seviyeleri ve değişken çalışma koşulları ele alacak şekilde geliştirilmiştir. Modelin çift yol mimarisi, titreşim sinyallerindeki hem yerel hem de küresel özellikleri yakalamak üzere tasarlanmıştır; 1D-CNN yolu ham sinyallerden zamansal özellikler çıkarırken, 2D-CNN yolu, sinyalleri matrislere dönüştürerek mekansal desenleri tanımlar. Bu yapı, modelin genelleme kabiliyetini artırarak yeni arıza durumlarına uyum sağlamasını kolaylaştırır. Model, yüksek boyutlu verileri daha düşük hesaplama yüküyle işleyerek çeşitli veri türlerinden anlamlı bilgiler çıkarır. Ayrıca WDD-CNN modelinin kapsamlı arızayla ilgili kalıpları öğrenmesini sağlayarak çeşitli operasyonel koşullara uyum yeteneğini artırmak amacıyla kayan pencere yöntemi uygulanarak, sınıf dengesizliğine yol açmadan her arıza durumu için ek eğitim örnekleri üretilmiştir. Veri artırma sürecinde, yalnızca kullanılabilir eğitim verilerinin miktarı artırılmış, ancak daha gerçekçi sonuçlar elde etmek için test verileri orijinal biçiminde bırakılmıştır. Çalışmada model, yaygın CWRU veri seti ile birlikte çalışmaya özgü üç farklı deneysel veri seti ile değerlendirilmiştir. CWRU veri setinde model %96,45 doğruluk oranına ulaşarak diğer yöntemlerden üstün performans sergilemiştir. Özgün veri setlerinde yapılan testlerde ise WDD-CNN modeli %100 doğruluğa ulaşmıştır. Orijinal veri setiyle yapılan deneylerde, WDD-CNN modeli test edilen tüm koşullar altında dikkate değer bir tanı performansı elde etmiş ve kesinlik, duyarlılık ve F1 puanı gibi temel değerlendirme ölçütlerinin her biri 1.0'lık maksimum değere ulaşmıştır. Sonuç olarak, WDD-CNN modeli, farklı hız ve çalışma koşullarında istikrarlı ve yüksek doğrulukla arıza teşhisi yaparak endüstriyel uygulamalar için güvenilir bir bakım sistemi olarak büyük potansiyel taşımaktadır. Gelecek çalışmalar için modelin farklı arıza türleri ve çoklu sensör verileri ile genişletilmesi ile daha fazla genelleme kabiliyeti kazandırılması önerilmektedir.

Özet (Çeviri)

In this thesis, a new artificial intelligence-based method is developed to improve fault diagnosis in electric vehicle motors, especially focusing on BLDC motors, which are widely used in industrial and automotive applications. Considering the increasing use of electric vehicles, it is very important to ensure the reliability and safety of electric motors through accurate and early fault detection. This study addresses this need by proposing a robust fault diagnosis model based on a hybrid structure of one-dimensional and two-dimensional convolutional neural networks (1D-CNN and 2D-CNN), called the WDD-CNN model. The model aims to detect a range of motor faults by analyzing vibration signals under varying operating conditions and facilitate early and accurate determination of potential faults. The WDD-CNN model is developed to address the challenges associated with high noise levels, variable operating conditions, and fault identification across classes. The dual-path architecture of the model is designed to capture both local and global features of vibration signals: the 1D-CNN path processes raw vibration signals to extract critical temporal features, while the 2D-CNN path uses transformed signal matrices to identify spatial patterns. This dual approach enhances the generalization ability of the model and enables it to effectively adapt to new fault conditions. Moreover, the 1D and 2D CNN paths are combined using an additional layer, which allows the extracted features of each path to improve the overall learning ability without increasing the model complexity. This structure is particularly advantageous for high-dimensional data as it minimizes the computational burden while preserving essential information from different dimensions of the input. The research includes four core datasets to comprehensively evaluate the effectiveness of the model: the Case Western Reserve University (CWRU) bearing dataset and three different original experimental datasets obtained from experimental sets specifically created for this study. The CWRU dataset is widely recognized and contains fault data for engine bearings, including different types and sizes of defects on the inner ring, outer ring, and ball. This dataset provides a controlled environment to evaluate the performance of fault diagnosis models. Experiments conducted on this dataset revealed that the WDD-CNN model outperformed the traditional methods and other compared methods by achieving an impressive average accuracy of 96.45%. This high level of accuracy demonstrates the strong generalization ability of the model, as it can identify various fault conditions at multiple speeds and loads, providing a reliable criterion for fault detection in industrial environments. In addition, three unique datasets have been created to further evaluate the model's adaptability and performance in more realistic environments. Each custom dataset is designed to reflect different operational scenarios and failure conditions, using specific hardware configurations to collect vibration data under different conditions: Original dataset 1: This dataset was collected using a BLDC drone motor (A2212/10T 1400KV) equipped with an ADXL-345 MEMS accelerometer. Data was recorded for both healthy and soft foot fault conditions at motor speeds of 2000, 4000 and 6000 rpm using a sampling rate of 25 kHz. The accelerometer was mounted close to the motor to capture x, y and z axis vibrations for 500 seconds per speed, resulting in 250,000 data points per condition. Original dataset 2: Collected using a 2 kW outer rotor BLDC motor, this data set includes a piezoelectric model 607A11 vibration sensor with a frequency response of up to 10 kHz. The motor's performance was recorded at 250, 400 and 600 rpm under both normal and imbalance fault conditions, with each test lasting 10 seconds and generating 250000 data points. Data acquisition was managed using an NI USB-6211 DAQ card, and the data was stored and processed on a PC for further analysis. Original dataset 3: This dataset utilized a 2 kW inrunner BLDC motor coupled to a shaft system, simulating conditions similar to those found in light commercial vehicle applications. Data were gathered with the Model MVS.101.51 piezoelectric accelerometer, capturing vibrations under normal and mechanical friction fault conditions at 200 and 400 rpm. Data acquisition involved the NI MyDAQ DAQ card with a sampling rate of 25 kHz, collecting 250000 data points per test condition. In this study, a sliding window technique was applied to augment the training data to ensure robust model training. This approach involved splitting continuous vibration signals into overlapping sequences, each representing a training sample, thus increasing the volume of training data and capturing a wider range of features. Approximately 5568 samples per failure class were generated for the CWRU training dataset. For the original datasets, approximately 1950 training samples were generated per failure class. A window length of 3600 samples with 128 steps was used in this data augmentation process, which significantly enriched the dataset and addressed class imbalance issues. This augmentation process enabled the WDD-CNN model to learn extensive failure-related patterns, increasing its adaptability to various operational conditions. In the data augmentation process, only the amount of available training data was increased, but the test data was left in its original form to obtain more realistic results. The architecture of the WDD-CNN model is designed to capture both temporal and spatial features that are critical for effective fault detection in vibration data. The 1D-CNN path processes raw vibration signals by focusing on temporal features indicating motor faults, while the 2D-CNN path transforms these signals into 60x60 matrices to capture spatial patterns. This dual-path structure enables the model to learn more complex fault-related patterns and enhances its robustness under different fault conditions. In the experiments with the original dataset, the WDD-CNN model achieved remarkable diagnostic performance under all tested conditions, and each of the key evaluation metrics, such as precision, recall, and F1-score, reached a maximum value of 1.0. Moreover, the accuracy metric is 100% for each original data set. This performance demonstrates the high reliability and effectiveness of the model in identifying faults even under harsh operational variations. Additionally, the adaptability of the model was evaluated by training it at one speed (e.g. 200 rpm) and testing it at another speed (e.g. 400 rpm) and vice versa.This approach simulated real-world conditions where motors may operate under varying loads and speeds. The results indicate that the WDD-CNN model retains its diagnostic accuracy across different speeds, underscoring its robustness in practical applications. This cross-speed testing showed that the model can generalize well to new operational settings without requiring retraining, which is a significant advantage for deployment in dynamic industrial environments. Overall, the results obtained in this study affirm that the proposed WDD-CNN model can effectively diagnose faults in electric motors with high accuracy and stability across varying speeds and conditions. The capability of the model to operate reliably under different working conditions highlights its potential for real-time applications in predictive maintenance systems. Furthermore, the findings suggest that integrating this model into industrial and automotive systems could lead to improved operational efficiency, reduced downtime, and enhanced motor lifespan by enabling timely fault detection and prevention. There are several avenues for future research that could improve the functionality and extend the applicability of the proposed model. First, expanding the range of fault types to include mechanical as well as electrical failures could provide a more comprehensive diagnostic capability. Furthermore, the inclusion of data from multiple sensors, such as temperature and acoustic sensors, could improve diagnostic ability by providing a richer dataset. Another promising direction is to apply transfer learning techniques to enable the model to adapt to new conditions or engine types with minimal additional training, which will further increase its practicality in various industrial settings.

Benzer Tezler

  1. Tez No

    884109

    Neuro classifiers for condition and bearing health assessment of an electric motor

    Elektrik makinasında durum ve rulman sağlığı değerlendirmesi için nöro sınıflandırıcılar

    MINA GHORBAN ZADEH BADELI

    Yüksek Lisans

    İngilizce

    İngilizce

    2022

    Elektrik ve Elektronik Mühendisliğiİstanbul Teknik Üniversitesi

    Elektrik Mühendisliği Ana Bilim Dalı

    DR. ÖĞR. ÜYESİ DUYGU BAYRAM KARA

  2. Tez No

    931620

    The lifespan comparison of an NMC811 pouch cell based high voltage battery with silicone and polyurethane foam usage

    NMC811 poşet hücre tabanli bir yüksek voltajli bataryada silikon ve poliüretan sünger kullaniminin ömür karşilaştirmasi

    SERHAT SOYER

    Yüksek Lisans

    İngilizce

    İngilizce

    2025

    Enerjiİstanbul Teknik Üniversitesi

    Makine Mühendisliği Ana Bilim Dalı

    DOÇ. DR. TURGUT GÜLMEZ

  3. Tez No

    926003

    Comparative analysis of multi-environment sensor data for reliable detection of electrical and mechanical faults in electrical motors

    Elektrik motorlarında elektriksel ve mekanik arızaların güvenilir tespiti için çoklu ortam sensör verilerinin karşılaştırmalı analizi

    GÜRCÜ NUR AYAS

    Yüksek Lisans

    İngilizce

    İngilizce

    2025

    Elektrik ve Elektronik MühendisliğiDokuz Eylül Üniversitesi

    Elektrik ve Elektronik Mühendisliği Ana Bilim Dalı

    DOÇ. DR. TANER GÖKTAŞ

  4. Tez No

    853732

    Investigation of solder joint durability in electric drive unit terminals under the effects of thermal aging and vibration

    Termal yaşlanma ve titreşim etkileri altında elektrik tahrik ünitesi terminallerinde lehim eklemi dayanıklılığının araştırılması

    BURAK YILMAZ

    Yüksek Lisans

    İngilizce

    İngilizce

    2024

    Makine MühendisliğiYeditepe Üniversitesi

    Makine Mühendisliği Ana Bilim Dalı

    DOÇ. DR. NAMIK ÇIPLAK

  5. Tez No

    381827

    Adaptive control of a novel tilt-roll rotor quadrotor UAV

    Adaptif dört rotorlu bir insansız hava aracının modellenmesi ve kontrolü

    ABDULKERİM FATİH ŞENKUL

    Yüksek Lisans

    İngilizce

    İngilizce

    2015

    Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrolİstanbul Teknik Üniversitesi

    Makine Mühendisliği Ana Bilim Dalı

    DOÇ. DR. ERDİNÇ ALTUĞ

Geri Dön