Posted by Admin: System Admin
More than half of all commercial aircraft operation accidents could have been prevented by executing a go around. Making timely decision to execute a go-around manoeuvre can potentially reduce overall aviation industry accident rate. In this paper, we describe a cockpit-deployable machine learning system to support flight crew go-around decision-making based on the prediction of a hard landing event. This work presents a hybrid approach for hard landing prediction that uses features modelling temporal dependencies of aircraft variables as inputs to a neural network. Based on a large dataset of 58177 commercial flights, the results show that our approach has 85% of average sensitivity with 74% of average specificity at the go-around point. It follows that our approach is a cockpit-deployable recommendation system that outperforms existing approaches. Machine learning is an important component of the growing field of data science. Through the use of statistical methods, different type of algorithms is trained to make classifications or predictions, and to uncover key insights in this project. These insights subsequently drive decision making within applications and businesses, ideally impacting key growth metrics. Machine learning algorithms build a model based on this project data, known as training data, in order to make predictions or decisions without being explicitly programmed to do so. Machine learning algorithms are used in a wide variety of datasets, where it is difficult or unfeasible to develop conventional algorithms to perform the needed tasks.
A Hard Landing (HL) is a phenomenon in which the airplane has an excessive impact on the ground at the moment of landing. This impact is directly related to the vertical (or normal) acceleration, therefore, HL can be defined as flights where the vertical acceleration exceeds the limited value of the aircraft type during the landing phase. A threshold on such normal acceleration (Airbus uses vertical acceleration > 2G at Touch Down, TD) triggers maintenance requirement, so that can be considered as a criterion for HL detection. Under the former definition of HL, existing approaches for HL prediction can be split into two groups: those based on a classifier to discriminate flights with normal acceleration at TD above a given threshold from other flights and those based on a regressor that predicts the normal acceleration with the aim of using this predicted value as the HL detector. Classifiers can be categorized into machine learning and deep learning approaches. Machine learning methods [17]– [19] apply a classifier to UAV flight data recorded using the Quick Access Recorder (QAR) sampled at a discrete set of heights that define the feature space. Most methods [17], [19] use the values of variables describing aircraft dynamics sampled between 9 and 2 meters before TD. Others, like [18], use statistical descriptors (panel data) of such variables also sampled at the very last meters before TD. On one hand, it is not clear what is the capability of these approaches to capture time-sequence dependencies that variables might have across the approach phase. On the other hand, the temporal window (9-2 meters before landing) used for predictions in UAV flights might not be appropriate for HL predictions in commercial flights. The approximate limit altitude (known as Decision Height -DH-) in commercial flights to decide a go around is 100 feet (38 meters). Thus, regardless of their accuracy in predicting HL, these ML methods are not applicable for commercial flights due to the altitude range required. Deep learning approaches are mainly based on Long Short-Term Memory Recurrent Neural Network (LSTM) architectures. Proposed by [20], these networks are a variant of Recurrent Neural Networks (RNN) [21] able to model long term dependencies within temporal data. In particular, the very recent work in [22] used the signals of 3 kinds of landing related features (aircraft dynamics, atmospheric environment, and pilot operations) as inputs to a LSTM network predicting HL. Their comparison to classic machine learning approaches in terms of precision and recall of HL events of A320 flights indicates a potentially higher performance in terms of HL recall with 70% of HL detection while keeping with a percentage (76%) of precision similar to the one obtained by classic machine learning approaches. Despite the promising results, we consider that the experimental design of [22] lacks in some aspects for properly assessing the potential for deployment in the cockpit. First, the test set used is balanced with almost the same number of HL and non HL cases. However, in a real situation, HL cases are rare events that represent only 3-4% of flights [23]. By balancing the test set, precision might be too optimistic and, even unrealistic. Disadvantages ? An existing system not implemented Sources of errors and capability for go-around recommendation. ? An existing system not implemented hybrid approach for hard landing prediction that uses features modeling temporal dependencies of aircraft variables as inputs to a neural network.
This paper presents an analysis of approaches for early prediction of hard-landing events in commercial flights. Unlike previous works, experiments are designed to analyze to what extend methods can be deployable in the cockpit as goaround recommendation systems. With this final goal, we contribute to the following aspects: 1) Hybrid model with optimized net architecture. We propose a hybrid approach that uses features modeling temporal dependencies of aircraft variables as input to a neural network with an optimized architecture. In order to avoid any bias caused by a lack of convergence of complex models (like LSTM), we use a standard network and model potential temporal dependencies associated with unstable approaches as the variability of different types of aircraft variables at a selected set of altitudes. The concatenation of such variability for variables categorized into 4 main types (physical, actuator, pilot operations and all of them) are the input features of different architectures in order to determine the optimal subset. 2) Exhaustive comparison to SoA in a large database of commercial flights. A main contribution compared to existing works is that our models have been tested and compared to SoA methods on a large database of Flight Management System (FMS) recorded data of an airline no longer in operation that includes 3 different aircraft models (A319, A320, A321). Results show that the optimal classification network when all variable types are considered achieves an average recall of HL events of 85% with a specificity of 75% in average, which outperforms current LSTM methods found in the literature. Regarding regression networks, our hybrid model performs similarly to LSMT methods with an average MSE of the order of 10????3 in accelerations estimated at TD. 3) Analysis of the performance of classifiers and regressors. With the final goal of developing a cockpit deployable recommendation system we have conducted a study of the performance of classification and regression models in terms of the flight height and different aircraft variables including the impact of automation and pilot manoeuvres. Results on our large dataset of commercial flights, show that although our regression networks performs similarly to SoA methods (with MSE of 10????3 in estimations at TD), the accuracy for detecting HL is very poor (46% of sensitivity). This indicates that regression models might not be the most appropriate for the detection of HL events in a cockpit deployable support system. 4) Sources of errors and capability for go-around recommendation. Unlike previous approaches, we analyze the capability of networks for the detection of HL before the decision height, as well as, the influence of the operational context. We have also performed an analysis of the sources of errors, including selection of the best variable type, optimal altitude range used for predictions, biases due to aircraft type and capability of regressors for HL prediction. Advantages ? The machine learning approach can also be improved in several aspects. Although results appear superior to existing methods, our models would benefit from a more complex analysis of temporal dependencies using a convolutional neural network to extract deep dependencies. ? In the proposed system, for a cockpit-deployable machine learning system to support flight crew go-around decision, some results regarding the hardware and software requirements, especially for the speed of networks should be investigated.