Soumalya Sarkar

Target Detection and Classification Using Seismic and PIR Sensors

Unattended ground sensors (UGS) are widely used to monitor human activities, such as pedestrian motion and detection of intruders in a secure region. Efficacy of UGS systems is often limited by high false alarm rates, possibly due to inadequacies of the underlying algorithms and limitations of onboard computation. In this regard, this paper presents a wavelet-based method for target detection and classification. The proposed method has been validated on data sets of seismic and passive infrared sensors for target detection and classification, as well as for payload and movement type identification of the targets. The proposed method has the advantages of fast execution time and low memory requirements and is potentially well-suited for real-time implementation with onboard UGS systems.

Performance comparison of feature extraction algorithms for target detection and classification

This paper addresses the problem of target detection and classification, where the performance is often limited due to high rates of false alarm and classification error, possibly because of inadequacies in the underlying algorithms of feature extraction from sensory data and subsequent pattern classification. In this paper, a recently reported feature extraction algorithm, symbolic dynamic filtering (SDF), is investigated for target detection and classification by using unmanned ground sensors (UGS). In SDF, sensor time series data are first symbolized to construct probabilistic finite state automata (PFSA) that, in turn, generate low-dimensional feature vectors. In this paper, the performance of SDF is compared with that of two commonly used feature extractors, namely Cepstrum and principal component analysis (PCA), for target detection and classification. Three different pattern classifiers have been employed to compare the performance of the three feature extractors for target detection and human/animal classification by UGS systems based on two sets of field data that consist of passive infrared (PIR) and seismic sensors. The results show consistently superior performance of SDF-based feature extraction over Cepstrum-based and PCA-based feature extraction in terms of successful detection, false alarm, and misclassification rates.

Symbolic Dynamic Analysis of Transient Time Series for Fault Detection in Gas Turbine Engines

This brief paper presents a symbolic dynamics-based method for detection of incipient faults in gas turbine engines. The underlying algorithms for fault detection and classification are built upon the recently reported work on symbolic dynamic filtering. In particular, Markov model-based analysis of quasi-stationary steady-state time series is extended to analysis of transient time series during takeoff. The algorithms have been validated by simulation on the NASA Commercial Modular Aero Propulsion System Simulation (C-MAPSS) transient test-case generator.

Sensor fusion for fault detection and classification in distributed physical processes

This paper proposes a feature extraction and fusion methodology to perform fault detection & classification in distributed physical processes generating heterogeneous data. The underlying concept is built upon a semantic framework for multi-sensor data interpretation using graphical models of Probabilistic Finite State Automata (PFSA).While the computational complexity is reduced by pruning the fused graphical model using an information-theoretic approach, the algorithms are developed to achieve high reliability via retaining the essential spatiotemporal characteristics of the physical processes. The concept has been validated on a simulation test bed of distributed shipboard auxiliary systems.

Dynamic data driven sensor array fusion for target detection and classification

Target detection and classification using unattended ground sensors (UGS) has been addressed in literature. Various techniques have been proposed for target detection, but target classification is a challenging task to accomplish using the limited processing power on each sensor module. The major hindrance in using these sensors reliably is, that, the sensor observations are significantly affected by external conditions, which are referred to as context. When the context is slowly time-varying (e.g., day-night cycling and seasonal variations) the usage of the same classifier may not be a good way to perform target classification. In this paper, a new framework is proposed as a Dynamic Data Driven Application System (DDDAS) to dynamically extract and use the knowledge of context as feedback in order to adaptively choose the appropriate classifiers and thereby enhance the target classification performance. The features are extracted by symbolic dynamic filtering (SDF) from the time series of sensors in an array and spatio-temporal aggregation of these features represents the context. Then, a context evolution model is constructed as a deterministic finite state automata (DFSA) and, for every context state in this DFSA, an event classifier is trained to classify the targets. The proposed technique of detection and classification has been compared with a traditional method of training classifiers without using any contextual information.

Multi-sensor information fusion for fault detection in aircraft gas turbine engines

The article addresses data-driven fault detection in commercial aircraft gas turbine engines in the framework of multi-sensor information fusion and symbolic dynamic filtering. The hierarchical decision and control structure, adopted in this article, involves construction of composite patterns, namely, atomic patterns extracted from single sensors, and relational patterns representing cross-dependence between a pair of sensors. While the underlying theories are presented along with necessary assumptions, the proposed method is validated on the NASA C-MAPSS simulation test bed of aircraft gas turbine engines; both single-fault and multiple-fault scenarios have been investigated. Since aircraft engines undergo natural degradation during the course of their normal operation, the issue of distinguishing between a fault and natural degradation is also addressed.

Dynamic data-driven prediction of lean blowout in a swirl-stabilized combustor

This paper addresses dynamic data-driven prediction of lean blowout (LBO) phenomena in confined combustion processes, which are prevalent in many physical applications (e.g., land-based and aircraft gas-turbine engines). The underlying concept is built upon pattern classification and is validated for LBO prediction with time series of chemiluminescence sensor data from a laboratory-scale swirl-stabilized dump combustor. The proposed method of LBO prediction makes use of the theory of symbolic dynamics, where (finite-length) time series data are partitioned to produce symbol strings that, in turn, generate a special class of probabilistic finite state automata (PFSA). These PFSA, called D-Markov machines, have a deterministic algebraic structure and their states are represented by symbol blocks of length D or less, where D is a positive integer. The D-Markov machines are constructed in two steps: (i) state splitting, i.e., the states are split based on their information contents, and (ii) state merging, i.e., two or more states (of possibly different lengths) are merged together to form a new state without any significant loss of the embedded information. The modeling complexity (e.g., number of states) of a D-Markov machine model is observed to be drastically reduced as the combustor approaches LBO. An anomaly measure, based on Kullback-Leibler divergence, is constructed to predict the proximity of LBO. The problem of LBO prediction is posed in a pattern classification setting and the underlying algorithms have been tested on experimental data at different extents of fuel-air premixing and fuel/air ratio. It is shown that, over a wide range of fuel-air premixing, D-Markov machines with D > 1 perform better as predictors of LBO than those with D = 1.

Context-aware Dynamic Data-driven Pattern Classification ∗

This work aims to mathematically formalize the notion of context, with the purpose of allowing contextual decision-making in order to improve performance in dynamic data driven classification systems. We present definitions for both intrinsic context, i.e. factors which directly affect sensor measurements for a given event, as well as extrinsic context, i.e. factors which do not affect the sensor measurements directly, but do affect the interpretation of collected data. Supervised and unsupervised modeling techniques to derive context and context labels from sensor data are formulated. Here, supervised modeling incorporates the a priori known factors affecting the sensing modalities, while unsupervised modeling autonomously discovers the structure of those factors in sensor data. Context-aware event classification algorithms are developed by adapting the classification boundaries, dependent on the current operational context. Improvements in context-aware classification have been quantified and validated in an unattended sensor-fence application for US Border Monitoring. Field data, collected with seismic sensors on different ground types, are analyzed in order to classify two types of walking across the border, namely, normal and stealthy. The classification is shown to be strongly dependent on the context (specifically, soil type: gravel or moist soil).

Anomaly detection in flight recorder data: A dynamic data-driven approach

This paper presents a method of feature extraction in the context of aviation data analysis. The underlying algorithm utilizes a feature extraction algorithm called symbolic dynamic filtering (SDF) that was recently published. In SDF, time-series data are partitioned for generating symbol sequences that, in turn, construct probabilistic finite state automata (PFSA) to serve as features for pattern classification. The SDF-based algorithm of feature extraction, which enjoys both flexibility of implementation and computational efficiency, is directly applicable to detection, classification, and prediction of anomalies and faults. The results of analysis with real-world flight recorder data show that the SDF-based features can be derived at a desired level of abstraction from the information embedded in the time-series data. The performance of the proposed SDF-based feature extraction is compared with that of standard temporal feature extraction for anomaly detection. Our study on flight recorder data shows that SDF-based features can enable discovering unique anomalous flights and improve the performance of the detection algorithm. We also theoretically show that under certain conditions it may be possible to achive a better or comparable time complexity with SDF based features.

Dynamic data-driven prediction of instability in a swirl-stabilized combustor:

Combustion instability poses a negative impact on the performance and structural durability of both land-based and aircraft gas turbine engines, and early detection of combustion instabilities is of paramount importance not only for performance monitoring and fault diagnosis, but also for initiating efficient decision and control of such engines. Combustion instability is, in general, characterized by self-sustained growth of large-amplitude pressure tones that are caused by a positive feedback arising from complex coupling of localized hydrodynamic perturbations, heat energy release, and acoustics of the combustor. This paper proposes a fast dynamic data-driven method for detecting early onsets of thermo-acoustic instabilities, where the underlying algorithms are built upon the concepts of symbolic time series analysis (STSA) via generalization of D-Markov machine construction. The proposed method captures the spatiotemporal co-dependence among time series from heterogeneous sensors (e.g. pressure and chemiluminescence) to generate an information-theoretic precursor, which is uniformly applicable across multiple operating regimes of the combustion process. The proposed method is experimentally validated on the time-series data, generated from a laboratory-scale swirl-stabilized combustor, while inducing thermo-acoustic instabilities for various protocols (e.g. increasing Reynolds number (Re) at a constant fuel flow rate and reducing equivalence ratio at a constant air flow rate) at varying air-fuel premixing levels. The underlying algorithms are developed based on D-Markov entropy rates, and the resulting instability precursor measure is rigorously compared with the state-of-the-art techniques in terms of its performance of instability prediction, computational complexity, and robustness to sensor noise.