44 Neural networks Machine learning techniques such as multi-layer perceptron (MLP) neural networks can be used to analyse vibration data and detect anomalies in real time, even in the presence of noise and dynamic operating conditions. Unlike convolutional neural networks (CNNs), which are optimised for image and spatial data, or recurrent neural networks suitable for sequential data processing, MLPs are highly effective at analysing non-sequential data such as vibration signals. Their inherent simplicity and adaptability make MLPs a suitable choice for capturing the complex, nonlinear relationships present in UAV rotor imbalance scenarios in performance monitoring. The lower computational requirements make them suitable for real-time onboard implementation in resource-constrained UAV systems. The MLP system uses micromachined micro-electromechanical systems (MEMS) sensors such as accelerometers for vibration data acquisition, preprocessing techniques for noise reduction and feature extraction, and an optimised MLP architecture tailored to high-dimensional vibration data. Rotor imbalances typically generate vibrations in the range of 10-100 Hz, which fall comfortably within the accelerometer’s range of up to 260 Hz. This ensures that the sensor can detect the primary vibration signals, the fundamental frequencies and their harmonics. The sensor’s range also supports capture of transient vibrations during rotor startup, shutdown or rapid speed changes, as well as steady-state vibrations during sustained flight. The upper limit of 260 Hz aids the detection of higher-order vibrations and harmonics that might arise from structural resonances or aerodynamics. MLPs also excel in scenarios with small- to medium-sized datasets, a common challenge in UAV rotor imbalance detection where extensive real-world data collection can be difficult. Their relatively straightforward and customisable structure also makes them more interpretable than more complex counterparts such as long short-term memory models or hybrid approaches, thereby enabling easier debugging. On a dataset of 800 samples representing both balanced and imbalanced rotor conditions, an optimised MLP model with five layers achieved a root mean squared error of 0.1414 Hz and a correlation coefficient of 0.9224 on the test dataset, demonstrating high accuracy and reliability. For a model customised for UAV applications, the framework addresses challenges such as non-stationary signals, dynamic flight conditions and real-time performance requirements. This bridges the gap between traditional analytical methods and machine learning-based diagnostics for improved UAV safety and maintenance efficiency, and provides a scalable and effective solution for rotor imbalance detection in dynamic environments. While hybrid models or CNN-based architectures might achieve slightly higher accuracy, they often come with increased complexity and computational overheads. Thus, MLPs strike a balance between performance and simplicity, making them a practical choice for rapid deployment and efficient operation. Motor monitoring The high risk of failure of bearings means that the identification and characterisation of bearing faults in motors via non-destructive evaluation methods have been studied extensively, amongst which vibration analysis has been found to be a promising technique for early diagnosis of anomalies. For smaller UAVs that cannot accommodate MEMS sensors, a systematic approach is needed for inflight non-destructive evaluation and health monitoring to enable efficient and safe operations in low-altitude airspace, as well as to mitigate associated risks to people and property on the ground. Data might be vehicle-specific such as battery state of charge (SoC; a component health status), from thirdparty sources such as weather, obstacle and terrain information providers, or from UAV Traffic Management such as realtime traffic within an airspace. Bearings Current practice for ensuring the safety of a motor includes a prescribed preflight and post-flight inspection of the motors followed by replacement of the bearing parts if a motor is deemed to be warmer or noisier than its counterparts. Bearings, like any component undergoing fatigue, are not supposed to fail instantaneously during a short flight of one to two hours, but lack of in-flight monitoring and the highly subjective nature of the prescribed checks August/September 2025 | Uncrewed Systems Technology Quadcopter arm and MEMS sensor placement (Image courtesy of Putra University)
RkJQdWJsaXNoZXIy MjI2Mzk4