Dinkar Mylaraswamy

Model-based diagnostics for an aircraft auxiliary power unit

This paper describes a case study of model-based diagnostics system development for an aircraft auxiliary power unit (APU) turbine system. The off-line diagnostics algorithms described in the paper work with historical data of a flight cycle. The diagnostics algorithms use detailed engine systems models and fault model knowledge available to Honeywell as the engine manufacturer. The developed algorithms provide fault condition estimates that allow for consistent detection of incipient performance faults and abnormal conditions.

A comparison of model-based and neural network-based diagnostic methods

Abstract Considerable attention has been devoted to the study of computer-aided process fault detection and diagnosis in recent years. A number of automated approaches have been proposed for this problem. Unfortunately, there exists very little body of work that compare and evaluate the different approaches to highlight their relative merits and demerits. This paper reports one such comparative study. This work compares the performance of a model-based diagnostic method (diagnostic model processor) and a neural network-based approach (ellipsoidal neural networks) on an industrial case study. The relative advantages and drawbacks of these two approaches are contrasted suggesting when one might use one of these alternatives.

Tiered architecture for on-line detection, isolation and repair of faults in wireless sensor networks

Wireless sensor networks fuse data from a multiplicity of sensors of different modalities and spatiotemporal scales to provide information for reconnaissance, surveillance, and situational awareness in many defense applications. For decisions to be based on information returned by sensor networks it is crucial that such information be of sustained high quality. While the Quality of Information (QoI) depends on many factors, perhaps the most crucial is the integrity of the sensor data sources themselves. Even ignoring malicious subversion, sensor data quality may be compromised by non-malicious causes such as noise, drifts, calibration, and faults. On-line detection and isolation of such misbehaviors is crucial not only for assuring QoI delivered to the end-user, but also for efficient operation and management by avoiding wasted energy and bandwidth in carrying poor quality data and enabling timely repair of sensors. We describe a two-tiered system for on-line detection of sensor faults. A local tier running at resource-constrained nodes uses an embedded model of the physical world together with a hypothesis-testing detector to identify potential faults in sensor measurements and notifies a global tier. In turn, the global tier uses these notifications on the one hand during fusion for more robust estimation of physical world events of interest to the user, and on the other hand for consistency checking among notifications from various sensors and generating feedback to update the embedded physical world model at the local nodes. Our system eliminates the undesirable attributes of purely centralized and purely distributed approaches that respectively suffer from high resource consumption from sending all data to a sink, and high false alarms due to lack of global knowledge. We demonstrate the performance of our system on diverse real-life sensor faults by using a modeling framework that permits injection of sensor faults to study their impact on the application QoI.

A Hybrid Prognostic Model Formulation and Health Estimation of Auxiliary Power Units

Prognostic health monitoring is an important element of condition-based maintenance and logistics support. The accuracy of prediction and the associated confidence in prediction greatly influence overall performance and subsequent actions either for maintenance or logistics support. Accuracy of prognosis is directly dependent on how closely one can capture the system and component interactions. Traditionally, such models assume a constant and univariate prognostic formulation-that is, components degrade at a constant rate and are independent of each other. Our objective in this paper is to model the degrading system as a collection of prognostic states (health vectors) that evolve continuously over time. The proposed model includes an age dependent deterioration distribution, component interactions, as well as effects of discrete events arising from line maintenance actions and/or abrupt faults. Mathematically, the proposed model can be summarized as a continuously evolving dynamic model, driven by non-Gaussian input and switches according to the discrete events in the system. We develop this model for aircraft auxiliary power units, but it can be generalized to other progressive deteriorating systems. The system identification and recursive state estimation scheme for the developed non-Gaussian model under a partially specified distribution framework has been deduced. The diagnostic/prognostic capabilities of our model and algorithms have been demonstrated using simulated and field data.

A hybrid prognostic model formulation system identification and health estimation of auxiliary power units

Prognostic health monitoring (PHM) is an important element of condition-based maintenance and logistics support. The accuracy of prediction and the associated confidence in prediction, greatly influences overall performance and subsequent actions either for maintenance or logistics support. Accuracy of prognosis is directly dependent on how closely one can capture the system and component interactions. Traditionally, such models assume constant and univariate prognostic formulation - that is, components degrade at a constant rate and are independent of each other. Our objective in this paper is to model the degrading system as a collection of prognostic states (health vectors) that evolve continuously over time. The proposed model includes an age dependent deterioration distribution, component interactions, as well as effects of discrete events arising from line maintenance actions and/or abrupt faults. Mathematically, the proposed model can be summarized as a continuously evolving dynamic model, driven by non-Gaussian input and switches according to the discrete events in the system. We develop this model for aircraft auxiliary power units (APU), but it can be generalized to other progressive deteriorating systems. We derive the system identification and recursive state estimation scheme for the developed non-Gaussian model under a partially specified distribution framework. The diagnostic/prognostic capabilities of our model and algorithms have been demonstrated using simulated and field data

Learning Bayesian Network Structures to Augment Aircraft Diagnostic Reference Models

Fault detection and isolation schemes are designed to detect the onset of adverse events during operations of complex systems, such as aircraft and industrial processes. The state-of-the-art fault diagnosis systems on aircraft combine an expert-created reference model of the associations between faults and symptoms, and a Naïve Bayes reasoner. For complex systems with many dependencies between components, the expert-generated reference models are often incomplete, which hinders timely and accurate fault diagnosis. Mining aircraft flight data is a promising approach to finding these missing relations between symptoms and data. However, mining algorithms generate a multitude of relations, and only a small subset of these relations may be useful for improving diagnoser performance. In this paper, we adopt a knowledge engineering approach that combines data mining methods with human expert input to update an existing reference model and improve the overall diagnostic performance. We discuss three case studies to demonstrate the effectiveness of this method.

On handling dependent evidence and multiple faults in knowledge fusion for engine health management

Diagnostic architectures that fuse outputs from multiple algorithms are described as knowledge fusion or evidence aggregation. Knowledge fusion using a statistical framework such as Dempster-Shafer (D-S) has been used in the context of engine health management. Fundamental assumptions made by this approach include the notion of independent evidence and single fault. In most real world systems, these assumptions are rarely satisfied. Relaxing the single fault assumption in D-S based knowledge fusion involves working with a hyper-power set of the frame of discernment. Computational complexity limits the practical use of such extension. In this paper, we introduce the notion of mutually exclusive diagnostic subsets. In our approach, elements of the frame of discernment are subsets of faults that cannot be mistaken for each other, rather than failure modes. These subsets are derived using a systematic analysis of connectivity and causal relationship between various components within the system. Specifically, we employ a special form of reachability analysis to derive such subsets. The theory of D-S can be extended to handle dependent evidence for simple and separable belief functions. However, in the real world the conclusions of diagnostic algorithms might not take the form of simple or separable belief functions. In this paper, we present a formal definition of algorithm dependency based on three metrics: the underlying technique an algorithm is using, the sensors it is using, and the feature of the sensor that the algorithm is using. With this formal definition, we partition evidence into highly dependent, weakly dependent and independent evidence. We present examples from a Honeywell auxiliary power unit to illustrate our modified D-S method of evidence aggregation

Information Quality aware sensor network services

Sensor networks with a number of wirelessly inter-connected devices have proven useful for many applications in diverse domains. The challenges of scale and resource constraints posed by this system have led to development of novel network protocols and services, but their focus has been on traditional metrics of quality of service of network data transport. Arguing that a holistic view of these systems requires an end-to-end view that combines networking quality of service concerns with the data quality and integrity of sensor sources and performance of sensor fusion algorithms, we propose to use the ldquoquality of informationrdquo (QoI) as a performance metric associated with the end result produced by a sensor network. In this paper we describe the factors that affect QoI, and describe how network protocols and services can be designed to be QoI aware.

Model-based diagnostics for small-scale turbomachines

This paper describes a case study of model-based diagnostics system development for an aircraft auxiliary power unit (APU) turbine system. The off-line diagnostics algorithms described in the paper work with historical data of a flight cycle. The diagnostics algorithms use detailed turbine engine systems models and fault model knowledge available to an engine manufacturer. The developed algorithms provide fault condition estimates and allow for consistent detection of incipient performance faults and abnormal conditions.

Cooperative sensor anomaly detection using global information

Sensor networks are deployed in many application areas nowadays ranging from environment monitoring, industrial monitoring, and agriculture monitoring to military battlefield sensing. The accuracy of sensor readings is without a doubt one of the most important measures to evaluate the quality of a sensor and its network. Therefore, this work is motivated to propose approaches that can detect and repair erroneous (i.e., dirty) data caused by inevitable system problems involving various hardware and software components of sensor networks. As information about a single event of interest in a sensor network is usually reflected in multiple measurement points, the inconsistency among multiple sensor measurements serves as an indicator for data quality problem. The focus of this paper is thus to study methods that can effectively detect and identify erroneous data among inconsistent observations based on the inherent structure of various sensor measurement series from a group of sensors. Particularly, we present three models to characterize the inherent data structures among sensor measurement traces and then apply these models individually to guide the error detection of a sensor network. First, we propose a multivariate Gaussian model which explores the correlated data changes of a group of sensors. Second, we present a Principal Component Analysis (PCA) model which captures the sparse geometric relationship among sensors in a network. The PCA model is motivated by the fact that not all sensor networks have clustered sensor deployment and clear data correlation structure. Further, if the sensor data show non-linear characteristic, a traditional PCA model can not capture the data attributes properly. Therefore, we propose a third model which utilizes kernel functions to map the original data into a high dimensional feature space and then apply PCA model on the mapped linearized data. All these three models serve the purpose of capturing the underlying phenomenon of a sensor network from its global view, and then guide the error detection to discover any anomaly observations. We conducted simulations for each of the proposed models, and evaluated the performance by deriving the Receiver Operating Characteristic (ROC) curves.