Full length articleReal-time anomaly detection framework using a support vector regression for the safety monitoring of commercial aircraft
Graphical abstract
Introduction
According to the International Air Transport Association (IATA), the number of passengers traveling by air is anticipated to double by 2035. This increases the likelihood of safety-related issues for the aviation infrastructure. The current air infrastructure strongly depends on manual systems such as human air traffic controllers, manual inspection, and communication between airliners and airport officers [1]. Therefore, there is a significant need to automate air traffic-related diagnostics for air traffic to progress smoothly. This is especially true for aircraft system safety; automated systems will ideally detect or predict system malfunctions and faults or at least detect performance anomalies that may be further analyzed by domain experts. Such an automated system can provide rapid and accurate diagnostic capabilities and prevent hazardous flight conditions and associated domino effects on air traffic systems [2], [3]. For instance, the Next Generation Air Transportation System (NextGen), currently in development by the Federal Aviation Administration (FAA), is expected to enable automatic and proactive in-air and on-ground aviation safety management through interconnected information processing [4].
Current aircraft monitoring methods depend on exceedance detection with predefined thresholds within subsystems to identify anomalous behavior [5]. However, exceedance criteria may have limited ability to capture unknown safety issues that are not explicitly defined by thresholds [6]. Furthermore, excessively strict thresholds may provoke unnecessary aircraft maintenance. For effective risk management, a robust health monitoring methodology should be capable of the early detection of operational anomalies and providing on-time targeted maintenance of the associated faulty system processes, which can prevent unnecessary cost and downtime of an aircraft [7], [8]. However, the development of such a comprehensive health monitoring framework for aircraft is challenging because of the complex and dynamic interactions of aircraft subsystems, which require simultaneous health monitoring and analysis at the subsystem level for safety evaluation of the global system [9].
Many approaches have been developed for monitoring system status and detecting anomalies to assure system reliability. In general, these approaches are placed into two categories: model-based and data-driven. Model-based methods use an explicit mathematical model to estimate system behavior through system identification techniques such as observer-based estimation and Kalman filters [10], [11], [12], [13], [14], [15], [16]. Wang and Lum [11] proposed an adaptive observer-based fault detection and isolation (FDI) approach and used an multiple adaptive observer that was capable of capturing actuator faults in an F-16 aircraft. Kobayashi and Simon [14] utilized multiple Kalman filters to capture faults in sensors and actuators through health parameter estimation based on fault indicators. However, modeling a complete aircraft system along with all related subsystems is limited by the assumptions made and uncertainties in the numerical model [17]. Furthermore, these models are strictly system-dependent and cannot be generalized [18]. Data-driven techniques, on the other hand, do not rely on a priori knowledge and utilize historical data to model the system response and provide detection capabilities [19], [20]. However, a primary issue in complex data-driven approaches with significant predictive abilities is the associated computational expense during the training process, which is attributed to the high dimensionality of the training data. Feature extraction and information fusion techniques such as principal component analysis (PCA), neural networks (NNs), and support vector machine (SVM) models have been utilized to address this issue [20], [21], [22], [23], [24], [25], [26], [27], [28], [29], [30], [31], [32], [33], [34]. Fujimaki et al. [28] investigated spacecraft performance by using kernel principal component analysis (KPCA) to detect malfunctions of the engine systems. Vanini et al. [31] adopted multiple dynamic neural networks (DNNs) for aircraft turbine FDI, where classical NNs cannot represent the system behavior because of the complexity of jet engine systems. Kromanis and Kripakaran [34] utilized the support vector regression (SVR) model with a moving fast Fourier transform (MFFT) to predict the behavior of large structures and capture a damage index for abnormal scenarios. However, these methods were tested on well posed faulty conditions introduced by simulation, and they assessed system integrity through a limited number of features. Therefore, the scope of such monitoring techniques may be limited in the context of detecting realistic faults that occur in complex scenarios and include inherent random noise.
In addition, outlier detection algorithms have also been adopted for anomaly detection, where abnormal conditions are identified as outliers [6], [35], [36], [37], [38], [39]. Das et al. [36] developed a multiple kernel anomaly detection (MKAD) algorithm that uses a one-class SVM model to detect potential anomalies in discrete and continuous flight parameters. Li et al. [6] proposed a clustering-based outlier detection algorithm that transforms time series flight information into a hyperspace and reduces the feature dimensions through PCA to capture operational anomalies as outliers. Tested flight features are flagged if they deviate from predefined standard operational bounds during the takeoff and approach phases. However, the developed methodologies were aiming at capturing unknown operational anomalies by analyzing historical flight dataset rather than detecting real-time anomalies. Therefore, these methodologies may be insufficient for real-time assessment of system health.
Although many techniques have been studied to develop aircraft health assessment methodologies, the drawbacks of current methods show that there is still an urgent need to develop a robust real-time health monitoring framework with the following features: (i) integrated system monitoring that enables health assessment of the aircraft subsystems with the overall aircraft operation and performance, and (ii) accurate detection capability with high computational efficiency for real-time assessment and early stage anomaly detection.
Motivated by these challenges, this paper proposes a robust, real-time, and data-driven anomaly detection framework that can potentially be applicable for systems such as the NextGen approach. In particular, the proposed framework utilizes on-board sensor data from commercial flights to predict possible flight performance anomalies. It should be noted that the proposed framework is mainly focused on achieving accurate detection capabilities with high computational efficiency through use of efficient application of training and detection techniques. Decimation and Savitzky–Golay filtering are used for preprocessing the signal data before training and test. Feature subset selection is performed with a correlation-based method to extract highly correlated flight features for training the model. This process improves the computational efficiency and prediction accuracy by reducing the dimensionality of the feature space and mitigating the overfitting issue for the trained model. The selected features are then used to train the SVR model and predict the flight performance as a function of the on-board flight sensor data. Anomalies are detected when the predicted flight performance violates a safety interval, which is defined by the mean and standard deviation of the flight dynamic features derived from the flight dataset. The proposed monitoring framework was validated through a simulation using historical flight datasets to demonstrate real-time anomaly detection. Other anomalous flight situations not included in historical flight datasets were also investigated to evaluate the detection robustness and degree of generalization.
Section snippets
Historical flight information and assumption
A sanitized commercial flight data recorder (FDR) dataset was utilized for training and testing the proposed anomaly detection methodology. The FDR (also called the black box in contemporary aviation literature) records all on-board sensor data in commercial aircraft during each flight and is generally used by airlines for flight analytics and occasionally digital forensics after flight malfunctions and/or crashes [40]. Although such data are difficult to acquire in their raw form outside of
Overview
Fig. 2 shows a schematic of the real-time anomaly detection framework, which involves a training phase and monitoring phase. In the training phase, the original FDR dataset is initially preprocessed with decimation and the Savitzky–Golay filter to synchronize the sampling frequency and reduce random noise in the sensor signal, as described in Section 3.2. Subsequently, the preprocessed flight features are reduced through feature subset selection, where features that are highly correlated with
Model training and evaluation
This section discusses the training process of the –SVR model and the prediction accuracy of the trained model. Initially, the selected feature and monitoring features to be predicted were introduced into the SVR model as normalized input (scaled between 0 and 1) to reduce divergence issues in the loss function during the training process. The predicted output features were then scaled back to their original form. A sequential minimal optimization algorithm was utilized in the training
Conclusion
In this study, a real-time aircraft health monitoring framework was developed to detect flight performance anomalies to potentially improve aviation safety while decreasing aircraft downtime. A historical flight dataset recorded from commercial aircraft was utilized to analyze the flight performance behavior and anomalies. Decimation and the Savitzky–Golay filter were employed to synchronize the feature dimensionality and remove noise from the raw signals. Feature selection was performed based
Acknowledgement
The research reported in this paper was supported by funds from NASA University Leadership Initiative program (Contract No. NNX17AJ86A, Project Technical Officer: Dr. Anupa Bajwa, Project Manager: Dr. Koushik Datta). The support is gratefully acknowledged.
References (62)
Model-based fault-detection and diagnosis - Status and applications
Annu. Rev. Control.
(2005)- et al.
Robust Adaptive Kalman Filter for estimation of UAV dynamics in the presence of sensor/actuator faults
Aerosp. Sci. Technol.
(2013) - et al.
Decentralized damage detection of seismically-excited buildings using multiple banks of Kalman estimators
Adv. Eng. Informatics
(2018) - et al.
From model-based control to data-driven control: Survey, classification and perspective
Inf. Sci. (Ny)
(2013) - et al.
Practical options for selecting data-driven or physics-based prognostics algorithms with reviews
Reliab. Eng. Syst. Saf.
(2015) Survey on data-driven industrial process monitoring and diagnosis
Annu. Rev. Control.
(2012)- et al.
Model-free data interpretation for continuous monitoring of complex structures
Adv. Eng. Informatics
(2008) - et al.
Exploration and evaluation of AR, MPCA and KL anomaly detection techniques to embankment dam piezometer data
Adv. Eng. Informatics
(2015) - et al.
Multi-sensor data fusion using support vector machine for motor fault detection
Inf. Sci. (Ny)
(2012) - et al.
Support vector regression for on-line health monitoring of large-scale structures
Struct. Saf.
(2006)