Enhancing UAV Safety with Machine Learning

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.May 23 2024

In a paper published in the journal Heliyon, researchers explored the versatile applications of unmanned aerial vehicles (UAVs) across diverse sectors, including environmental monitoring, recreation, healthcare, military operations, and transportation. As UAV usage grows with advancements in battery technology, brushless direct current (BLDC) motors have become integral for propulsion. However, BLDC motor failures can pose serious risks, potentially leading to accidents and loss of control. 

To mitigate this, the authors introduced a novel machine learning-based method to anticipate defects in BLDC motors within UAVs. Their study rigorously compared the effectiveness of three key machine learning models—k-nearest neighbor (KNN), support vector machine (SVM), and Bayesian network (BN)—to optimize performance and enhance safety in UAV operations.

UAV Fault Detection

Previous studies have explored diverse approaches to detect faults in unmanned aerial vehicle (UAV) components. For example, sound signal analysis combined with machine learning achieved accuracies ranging from 90.53% to 100%. Acceleration sensor data enabled the detection of propeller failures with over 95% accuracy. Artificial neural networks (ANNs) identified BLDC motor faults with a 99% accuracy rate, while the adaptive neuro-fuzzy inference system (ANFIS) classified bearing defects with 97% accuracy.

Thermal imaging and convolutional neural networks (CNN) yielded 100% detection rates for motor failures, as did probabilistic neural networks (PNN) for winding faults. These studies collectively demonstrate the effectiveness of machine learning and signal analysis in enhancing UAV operational safety.

BLDC Motor Defect Identification

The methodology for identifying defects in BLDC motors encompasses three main approaches: model-based, data-driven, and hybrid. Model-based methods require a deep understanding of the physical and chemical processes underlying degradation. Data-driven methods leverage large datasets, often from Internet of Things (IoT) sensors, using machine learning to estimate remaining useful life (RUL) and diagnose defects. This study employed a data-driven method, utilizing historical data from sensor monitoring to classify defects in BLDC motors, including chipped propellers, eccentric shafts, and electronic speed controller (ESC) faults.

The process begins with preprocessing historical data, extracting defective features, and labeling them to capture relevant information about the analyzed defects. Both historical and real-time data undergo preprocessing steps such as noise extraction, outlier elimination, missing value completion, and normalization. Feature extraction, crucial for improving classification performance, occurs in the time domain, utilizing statistical indicators like mean, standard deviation, kurtosis, skewness, and root mean square.

The training dataset, derived from an experimental setup, contains information from sensors monitoring motor behavior under various defect classes. Machine learning algorithms, specifically KNN, SVM, and BN, are then employed to classify defects in BLDC motors. Performance metrics such as accuracy, precision, recall, and F1 score are used to evaluate the effectiveness of each classifier in accurately identifying fault types.

The performance analysis of classifiers involves measuring key indicators such as accuracy, precision, recall, and F1 score for each defect class. Additionally, macro-average metrics provide an overall assessment of classifier performance across all defect classes, facilitating a comprehensive evaluation of the methodology's effectiveness in fault classification.

Experimental Findings

The experimental setup section provides a detailed overview of the data collected during evaluating BLDC motors. It showcases variations in temperature, electrical current intensity, and acceleration magnitude across different classes and motor speeds. The graphs illustrate significant amplitude variations for the ESC fault class, indicating high energy consumption and noisy operation. These findings underscore the distinct characteristics of each fault class, which are essential for accurate defect classification.

Moreover, the section highlights the influence of rotor shaft speed on parameter variations, emphasizing the unique fingerprint created by each analyzed class. The methodology's implementation used Python 3.6 with Tensorflow and Pandas packages. Experimental results, facilitated by a personal computer (PC) with specified hardware configurations, reveal the performance of three BLDC motors, three propellers, and two ESCs in generating the dataset. Each class exhibits distinct parameter variations, contributing to creating unique data fingerprints.

The evaluation metrics, including accuracy, precision, recall, and F1 score, demonstrate the SVM classifier's superior performance in defect classification compared to KNN and BN classifiers. Further insights suggest potential enhancements in classifier performance by including additional datasets featuring varied chip shapes and motor configurations. In conclusion, the study underscores the significance of robust classification algorithms in accurately identifying BLDC motor defects.

The SVM classifier emerges as the most effective model, offering promising results in fault classification. However, opportunities for improvement exist, such as augmenting training datasets with diverse sensor values and monitoring parameters. Additionally, considerations for algorithm efficiency, highlighted by the faster processing speed of SVM compared to KNN and BN, further underscore avenues for optimizing defect classification methodologies.

Conclusion

To sum up, the study presented a novel approach for fault classification of BLDC motors in UAVs using ML. By comparing SVM, KNN, and BN classifiers, SVM emerged as the top performer. The training dataset encompassed various fault classes, each exhibiting unique features. Experimental results highlighted distinct characteristics of motor operation under different fault classes, particularly the ESC fault class. The proposed model demonstrated superior predictive capabilities, making it suitable for UAV predictive maintenance systems.