Search

The results are presented of a comparative study evaluating the performance of neural network (NN) and fuzzy logic reconstructors (FLRs) for the development of a virtual flight data recorder (VFDK). Typical flight data recorders (FDRS) on commercial airliners do not record the aircraft control surface deflections. These dynamic parameters are critical in the investigation of an accident or an uncommanded maneuver. The results are shown relative to a VFDR based on a neural network simulator (NNS) along with a neural network reconstructor (NNR) or a FLR The NNS is trained off-line, using available flight data for the particular aircraft, for the purpose of simulating any desired dynamic output recorded in current FDRs. The NNS is then interfaced with the NNR or with the FLR. The output of the two reconstructors are the control surface deflections which minimize a performance index based on the differences between the available data from the FDR and the output from the NNS. The study tested with night data from a B737-300 shows that both schemes, the one with the NNR and the one with the FLR, provide accurate reconstructions of the control surface deflections time histories.

This paper presents preliminary results of the development of a virtual flight data recorder (VFDR) for commercial airliners. Federal Aviation Administration (FAA) regulations, currently being revised, mandate the recording of 11 dynamic parameters, not including the control surface deflections. The absence of these data can be critical for crash investigation purposes. This paper proposes the introduction of a VFDR based on a neural network simulator (NSS) and a neural network reconstructor (NNR). The NNS is trained, using flight data for the particular aircraft, to simulate any desired control surface deflections (or any other parameters of interest not recorded by the FDR), minimizing a cost function based on the differences between the available data from the FDR and the output from the NNS. The VFDR scheme has been introduced, tested, and validated with flight data from a Boeing 737-300 with an FDR with extended recording capabilities showing accurate reconstruction of the control surface deflections’ time histories. The VFDR can be considered a tool for crash investigations where control surface failures are believed to be a factor.

Detection, identification, and accommodation of sensor failures can be a challenging task for complex dynamic systems. This paper presents the comparison of two different approaches for the task of sensor failure detection, identification, and accommodation in a flight control system assumed to be without physical redundancy in the sensory capabilities. The first approach is based on the use of a set of online learning neural networks; the second approach is based on the use of a bank of Kalman filters. The objective is to evaluate the robustness of both schemes; the comparison is performed through testing of the schemes for several types of failures presenting different level of complexity in terms of detectability. The required computational effort for both schemes is also evaluated. For each of these failure types this comparison is performed at nominal conditions, that is with the system model and its noise perfectly modeled (with the Kalman filter scheme performing at nominal conditions), and at conditions, where discrepancies occur for the modeling of the system as well as the system and measurement noises. While the Kalman-filter-based scheme takes advantage of its robustness capabilities, the neural-network-based scheme, starting from a random numerical architecture, relies on the learning accumulated either online or from off-line simulations. The study reveals that online learning neural architectures have potential for online estimation purposes in a sensor validation scheme, particularly in the case of poorly modeled dynamics.

An the investigation of an aircraft crash the flight recorder data, the voice recorder tracks, and the inspection of the aircraft wreckage are key elements. At times these elements are not sufficient to allow the investigators to reach a conclusion about a cause or a set of causes for the failure(s) leading to the crash. Federal rules do not mandate for a "black box" to record deflections of the aircraft control surfaces. Nevertheless, it is clear that these dynamic time histories can provide critical information leading to the discovery of the cause(s) of the crash. In this paper a neural network approach is proposed for the reconstruction of the time histories of the control surfaces deflections. The neural inputs are the discretized "black box" data; the neural outputs are the estimates of the control surface deflections which minimize weighted quadratic differences between the time histories data from a parallel simulation code and the actual "black box" data. The approach is iterative in nature; starting from an initial time instant, the signal reconstruction scheme will proceed to the next time instant only when the estimates of the control surfaces deflections at the current instant minimize the weighted quadratic differences relative to the next time instant. The results relative to the reconstruction of the control surfaces deflections for a single type of failure under different conditions are presented and discussed.

The objective of this document Is to show the capabilities of parallel hardware-based on-line learning neural networks (NNs). This specific application is related to an on-line estimation problem for sensor validation purposes. Neural-network-based microprocessors are starting to be commercially available. However, most of them feature a learning performed with the classic back-propagation algorithm (BPA). To overcome this lack of flexibility a customized motherboard with transputers was implemented for this investigation, The extended BPA (EBPA), a modified and more effective BPA, was used for the on-line learning, These parallel hardware-based neural architectures were used to implement a sensor failure detection, identification, and accommodation scheme in the model of a night control system assumed to be without physical redundancy in the sensory capabilities. The results of this study demonstrate the potential for these neural schemes for implementation in actual flight control systems of modern high performance aircraft, taking advantage of the characteristics of the extended back-propagation along with the parallel computation capabilities of NN customized hardware.