Ihab Samy

Neural-network-based flush air data sensing system demonstrated on a mini air vehicle

Flush air data sensing systems have been widely applied to large (manned) aircraft, where pressure orifices are typically located at the nosetip. This paper investigates the feasibility of a flush air data sensing system designed to estimate the air data states of a small unmanned air vehicle flown at speeds as low as Mach 0.07. Furthermore, due to the presence of a nose propeller, the pressure orifices are located at the wing leading edge. The motivation behind this project is the fact that traditional air data booms are physically impractical for small unmanned air vehicles. Overall, an 80 and 97% reduction in instrumentation weight and cost, respectively, were achieved. Both parametric and multilayer perceptron neural network models have been previously applied in the literature to model the aerodynamic relationship between aircraft surface pressure and the air data states. In this paper, an extended minimum resource allocating network radial basis function neural network is used as the flush air data sensing system model, due to its good generalization capabilities and compact structure. Computational fluid dynamic simulations are implemented to identify the ideal pressure port locations, and wind-tunnel tests are carried out to train and test the extended minimum resource allocating network radial basis function neural network.

Sensor fault detection and accommodation using neural networks with application to a non-linear unmanned air vehicle model

Abstract Model-based fault detection and isolation (FDI) systems have become a crucial step towards autonomy in aerospace engineering. Traditional methods such as observer-based methods have already been developed and widely applied. Novel approaches make use of online learning neural networks (NN) which have seen an increase in FDI applications over the years. However, few publications consider FDI applications to unmanned air vehicles (UAV) where high levels of autonomy are required. This article demonstrates such an application, where an extended minimum resource allocation network radial basis function (RBF) NN is used for modelling purposes. A novel residual generation approach is also presented and found to outperform a conventional approach by reducing the number of false alarms and missed faults. All tests are carried out in simulation where single sensor faults are assumed to occur in the pitch gyro of a non-linear UAV model.

A comparative study of NN- and EKF-based SFDA schemes with application to a nonlinear UAV model

In this article, we propose two schemes for sensor fault detection and accommodation (SFDA): one based on a neural network (NN) and the other on an extended Kalman filter (EKF). The objective of this article is to compare both approaches in terms of execution time, robustness to poorly modelled dynamics and sensitivity to different fault types. The schemes are tested on an unmanned air vehicle (UAV) application where traditional sensor redundancy methods can be too heavy and/or costly. In an attempt to reduce the false alarm rates and the number of undetected faults, a modified residual generator, originally proposed in Samy, Postlethwaite, and Gu in 2008 (Samy, I., Postlethwaite, I., and Gu, D.-W. (2008a). Neural Network Sensor Validation Scheme Demonstrated on a UAV Model, in IEEE Proceedings of CDC, Cancun, Mexico, pp. 1237–1242) is implemented. Simulation work is presented for use on a UAV demonstrator under construction with support from BAE systems and EPSRC. Results have shown that the NN-SFDA scheme outperforms the EKF-SFDA scheme with only one missed fault, zero false alarms and an average estimation error of 0.31°/s for 112 different test conditions.

Unmanned air vehicle air data estimation using a matrix of pressure sensors: a comparison of neural networks and look-up tables

Flush airdata sensing (FADS) systems are cost- and weight- effective alternatives to current air data booms for measuring important air data parameters such as airspeed, angle of attack, sideslip, etc. Most applications consider large manned/unmanned air vehicles where the Pitot-static tube is located at the nose tip. However, traditional air data booms can be physically impractical for micro- (unmanned) air vehicles (MAVs) and, in this article, a FADS system mounted on the wing leading edge of a MAV flown at low speeds of Mach 0.07 (wind tunnel experiments under corresponding conditions) is designed. Moreover, two approaches for converting the FADS system pressure to meaningful air data are compared: a neural network (NN) approach and a look-up table (LUT). Results have shown that instrumentation weight and cost were reduced by 80 per cent and 97 per cent, respectively, in comparison to a traditional air data boom. Overall, the NN estimation accuracies were 0.51°, 0.44 lb/ft2, and 0.62 m/s and the LUT estimation accuracies 1.32°, 0.11 lb/ft2, and 0.88 m/s for the angle of attack, static pressure, and airspeed, respectively. It was also found that the LUT has faster execution times while the NN was in most cases more robust to sensor faults. However, while the LUT requires high memory usage, especially for higher dimensions, the NN can be executed in a few lines of code.

Neural network based sensor validation scheme demonstrated on an unmanned air vehicle (UAV) model

Nowadays model-based fault detection and isolation (FDI) systems have become a crucial step towards autonomy in aerospace engineering. However few publications consider FDI applications to unmanned air vehicles (UAV) where full-autonomy is obligatory. In this paper we demonstrate a sensor fault detection and accommodation (SFDA) system, which makes use of analytical redundancy between flight parameters, on a UAV model. A Radial-Basis Function (RBF) neural network (NN) trained online with Extended Minimum Resource Allocating Network (EMRAN) algorithms is chosen for modelling purposes due to its ability to adapt well to nonlinear environments while maintaining high computational speeds. Furthermore, in an attempt to reduce false alarms (FA) and missed faults (MF) in current SFDA systems, we introduce a novel residual generator. After 47 minutes (CPU running time) of NN offline training, the SFDA scheme is able to detect additive and constant bias sensor faults with zero FA and MF. It also shows good global approximation capabilities, essential for fault accommodation, with an average pitch gyro estimation error of 0.0075 rad/s.

Fault detection and flight data measurement: Demonstrated on unmanned air vehicles using neural networks

Introduction.- Fault detection and isolation (FDI).- Introduction to FADS systems.- Neural Networks.- SFDA-Single sensor faults.- SFDIA-Multiple sensor faults.- FADS system applied to a MAV.- Conclusions and Future Work.

Detection and accommodation of sensor faults in UAVs- a comparison of NN and EKF based approaches

In this paper we propose two schemes for sensor fault detection and accommodation (SFDA); one based on a neural network (NN) and the other an extended Kalman filter (EKF). The objective is to compare both approaches in terms of execution time, robustness to poorly modelled dynamics and sensitivity to different fault types. The schemes are tested on an unmanned air vehicle (UAV) application where traditional sensor redundancy methods can be too heavy and/or costly. In an attempt to reduce the false alarm rates and the number of undetected faults, a modified residual generator, originally proposed in [11], is implemented. Simulation work is presented for use on a UAV demonstrator under construction with support from BAE Systems and EPSRC. Results have shown that the NN-SFDA scheme outperforms the EKF-SFDA scheme with only 1 missed fault, zero false alarms and an average estimation error of 0.31deg/s for 112 different test conditions.

SFDIA of consecutive sensor faults using neural networks – demonstrated on a UAV

Neural network based sensor fault detection, isolation and accommodation (NN-SFDIA) is becoming a popular alternative to traditional linear time-invariant model-based sensor fault detection, isolation and accommodation (SFDIA) schemes, such as observer-based methods. Their online training capabilities and ability to model complex nonlinear systems have attracted much research interest in the applications area of neural networks. In this article, we design an NN-SFDIA scheme to detect multiple sensor faults in an unmanned air vehicle (UAV). Model-based SFDIA is a direction of development in particular with UAVs where sensor redundancy may not be an option due to weight, cost and space implications. In this article, a maximum of three consecutive faults are assumed in the pitch gyro, normal accelerometer and angle of attack sensor of a nonlinear UAV model. Furthermore, a novel residual generator which is designed to minimise the false alarm rates and missed faults, is implemented. After 33 separate SFDIA tests implemented on a 1.6 GHz Pentium processor, the NN-SFDIA scheme detected all but three faults with a fast execution time of 0.55 ms per flight data sample.

A comparison of neural networks for FDI of rolling element bearings – demonstrated on experimental rig data

In this article, a fault detection and isolation (FDI) approach for bearing faults in rotating machinery using a combination of vibration analysis and expert systems via neural networks (NNs) is proposed. The NN chosen is the extended minimum resource allocating network (EMRAN) radial basis function (RBF) due to its good performance characteristics. While the EMRAN RBF NN structure is itself not novel, the application to bearing FDI has, to the authors knowledge, been less frequently explored. The EMRAN RBF NN is used for pattern classification of four types of bearing health conditions: healthy, inner race, outer race, and ball bearing faults. A machine fault simulator is used to simulate the bearing faults and the input nodes of the NN include five features extracted from the time-domain vibration data: peak, root mean square, standard deviation, kurtosis, and normal negative log-likelihood value. Using real experimental data from a machine fault simulator, it was found that a 3-7-4 EMRAN RBF NN structure outperforms a 5-20-4 multilayered perceptron NN with zero false alarms, fewer undetected faults, higher pattern correlation factors, and faster execution times.

Fault diagnosis of rolling element bearings using an EMRAN RBF neural network- demonstrated using real experimental data

rolling element bearings are critical components of rotating machinery. Failure diagnosis of bearing faults is necessary and can often avoid more catastrophic failure consequences. Nowadays vibration condition monitoring is the most frequently used failure diagnostic method for rotating machinery. Several designs have been proposed in the literature and in this paper we propose a different approach using a radial basis function (RBF) neural network (NN) trained with extended minimum resource allocating network (EMRAN) algorithms, for pattern classification of 4 types of bearing health conditions: healthy, inner race, outer race and ball bearing faults. The input nodes of the NN consist of five features extracted from the time domain vibration data: peak, root mean square, standard deviation, kurtosis and normal negative log-likelihood value. Furthermore the NN is analyzed in terms of sensitivity to the different input features in order to remove significant and/or redundant inputs. The accuracy of the pattern classification technique is compared for both longitudinal and vertical accelerations. Using real experimental data from a machine fault simulator it was found that the EMRAN RBF NN requires only a few features and classifies the 4 types of bearing faults with good accuracy. The effectiveness of the approach proposed in this paper has illustrated its feasibility for real time condition monitoring of rotating machinery.