Detecting abnormal situations using the Kullback–Leibler divergence

This article develops statistics based on the Kullback–Leibler (KL) divergence to monitor large-scale technical systems. These statistics detect anomalous system behavior by comparing estimated density functions for the current process behavior with reference density functions. For Gaussian distributed process variables, the paper proves that the difference in density functions, measured by the KL divergence, is a more sensitive measure than existing work involving multivariate statistics. To cater for a wide range of potential application areas, the paper develops monitoring concepts for linear static systems, that can produce Gaussian as well as non-Gaussian distributed process variables. Using recorded data from a glass melter, the article demonstrates the increased sensitivity of the KL-based statistics by comparing them to competitive ones.     

An edge detection algorithm using multi-state adalines

An edge detection algorithm using multi-state adaptive linear neurons (ADALINES) is presented. Although the tri-state ADALINE is only considered in this work, general multi-state input vectors with extreme values are shown to be linearly separable from the rest of the vectors with the same dimension. The input state of each ADALINE is defined using the local mean in a predefined mask. In addition to the binary input states ± 1, the 0 input state is introduced for controlling the noise effect. If the input pattern matches one of the predefined edge patterns, the corresponding pixel is detected as an edge pixel. Experimental results are shown where the proposed detector is compared with both the Canny and LOG edge detectors.     

Behavior monitoring under uncertainty using Bayesian surprise and optimal action selection

The increasing trend towards delegating tasks to autonomous artificial agents in safety–critical socio-technical systems makes monitoring an action selection policy of paramount importance. Agent behavior monitoring may profit from a stochastic specification of an optimal policy under uncertainty. A probabilistic monitoring approach is proposed to assess if an agent behavior (or policy) respects its specification. The desired policy is modeled by a prior distribution for state transitions in an optimally-controlled stochastic process. Bayesian surprise is defined as the Kullback–Leibler divergence between the state transition distribution for the observed behavior and the distribution for optimal action selection. To provide a sensitive on-line estimation of Bayesian surprise with small samples twin Gaussian processes are used. Timely detection of a deviant behavior or anomaly in an artificial pancreas highlights the sensitivity of Bayesian surprise to a meaningful discrepancy regarding the stochastic optimal policy when there exist excessive glycemic variability, sensor errors, controller ill-tuning and infusion pump malfunctioning. To reject outliers and leave out redundant information, on-line sparsification of data streams is proposed.     

Abnormality detection based on the Kullback–Leibler divergence for generalized Gaussian data

This paper is on abnormality detection, where the observed data under the normal condition is assumed to be independent and identically distributed (i.i.d.) and follow the generalized Gaussian distribution (GGD) with shape parameter greater than 1. The Kullback–Leibler divergence (KLD) between the estimated GGD of the observed data and the normal one is used as the test statistic. An analytical expression of the KLD is derived under the normal condition when the number of samples is large; then, two algorithms with constant and adaptive thresholds are proposed. Extensive simulated and industrial case studies are conducted to verify the analytical results and to show the effectiveness of the proposed algorithms.     

Fault detection and isolation based on fuzzy automata

Fuzzy automata are proposed for fault diagnosis. The output of the monitored system is partitioned into linear segments which in turn are assigned to pattern classes (templates) with the use of membership functions. A sequence of templates is generated and becomes input to fuzzy automata which have transitions that correspond to the templates of the properly functioning system. If the automata reach their final states, i.e. the input sequence is accepted by the automata with a membership degree that exceeds a certain threshold, then normal operation is deduced, otherwise, a failure is diagnosed. Fault diagnosis of a DC motor and detection of abnormalities in the ECG signal are used as case studies.     

An automatic algorithm for the detection of Trypanosoma cruzi parasites in blood sample images

Chagas disease is a tropical parasitic disease caused by the flagellate protozoan Trypanosoma cruzi (T. cruzi) and currently affecting large portions of the Americas. One of the standard laboratory methods to determine the presence of the parasite is by direct visualization in blood smears stained with some colorant. This method is time-consuming, requires trained microscopists and is prone to human mistakes. In this article we propose a novel algorithm for the automatic detection of T. cruzi parasites, in microscope digital images obtained from peripheral blood smears treated with Wright's stain. Our algorithm achieved a sensitivity of 0.98 and specificity of 0.85 when evaluated against a dataset of 120 test images. Experimental results show the versatility of the method for parasitemia determination.     

Algorithms for leak detection, estimation, isolation and localization in open water channels

In this paper, a methodology for leak detection and isolation in open water channels is proposed. It consists of on-line leak detection and off-line leak isolation, estimation and localization. The on-line leak detection is carried out pool by pool using local information only, and a comparison between three leak detection methods in terms of detection performance and ease of implementation is presented. Information from neighboring pools is used to distinguish the effect of a leak from sensor faults. The proposed methods were tested on real experimental data from the Coleambally Channel no. 6 in Australia and showed good performance.     

SCaNME: Location tracking system in large-scale campus Wi-Fi environment using unlabeled mobility map

In this paper, we propose a novel location tracking system called SCaNME (Shotgun Clustering-aided Navigation in Mobile Environment) which iteratively sequences the clusters of sporadically recorded received signal strength (RSS) measurements and adaptively construct a mobility map of the environment for location tracking. In the SCaNME system, the location tracking problem is solved by first matching the people’s locations to the location points (LPs) with small Kullback–Leibler (KL) divergence. Then, Allen’s logics are applied to reveal the person’s activities, assist the on-line location tracking and finally obtain a refined path estimate. The experimental results conducted on the large-scale HKUST campus demonstrate that the SCaNME tracking system provides better precision and reliability than the conventional location tracking systems. Furthermore, the experiments of SCaNME tracking system show its capability of providing people’s real-time locations without fingerprint calibration in large-scale Wi-Fi environment.     

Fault detection for a diesel engine actuator - A benchmark for FDI

An electro-mechanical position servo is introduced as a benchmark for mode-based Fault Detection and Identification (FDI). The purpose is to provide a simple, industrial system as a platform for comparison of model-based FDI methods, and for the gathering of design experience. Despite a simple system structure, FDI design is intricate when realistic obstacles are included: measurement noise. FDI performance specifications are provided. They include requirements for detection probability and time to detect. Two mathematical models are given: a simple model for use during design, and a complex, nonlinear one for simulation and verification.     

Plant-wide detection and diagnosis using correspondence analysis

This paper presents an approach based on the correspondence analysis (CA) for the task of fault detection and diagnosis. Unlike other data-based monitoring tools, such as principal components analysis/dynamic PCA (PCA/DPCA), the CA algorithm has been shown to use a different metric to represent the information content in the data matrix X. Decomposition of the information represented in the metric is shown here to yield superior performance from the viewpoints of data compression, discrimination and classification, as well as early detection and diagnosis of faults. Metrics similar to the contribution plots and threshold statistics that have been developed and used for PCA are also proposed in this paper for detection and diagnosis using the CA algorithm. Further, using the benchmark Tennessee Eastman problem as a case study, significant performance improvements are demonstrated in monitoring and diagnosis (in terms of shorter detection delays, smaller false alarm rates, reduced missed detection rates and clearer diagnosis) using the CA algorithm over those achievable using the PCA and DPCA algorithms.     

Distributed fault detection for interconnected second-order systems

In this paper, the existence of unknown input observers for networks of interconnected second-order linear time invariant systems is studied. Two classes of distributed control systems of large practical relevance are considered. It is proved that for these systems, one can construct a bank of unknown input observers, and use them to detect and isolate faults in the network. The result presents a distributed implementation. In particular, by exploiting the system structure, this work provides further insight into the design of UIO for networked systems. Moreover, the importance of certain network measurements is shown. Infeasibility results with respect to available measurements and faults are also provided, as well as methods to remove faulty agents from the network. Applications to power networks and robotic formations are presented. It is shown how the developed methodology apply to a power network described by the swing equation with a faulty bus. For a multi-robot system, it is illustrated how a faulty robot can be detected and removed.     

AIRBUS state of the art and practices on FDI and FTC in flight control system

This paper deals with industrial practices and strategies for Fault Tolerant Control (FTC) and Fault Detection and Isolation (FDI) in civil aircraft by focusing mainly on a typical Airbus Electrical Flight Control System (EFCS). This system is designed to meet very stringent requirements in terms of safety, availability and reliability that characterized the system dependability. Fault tolerance is designed into the system by the use of stringent processes and rules, which are summarized in the paper. The strategy for monitoring (fault detection) of the system components, as a part of the design for fault tolerance, is also described in this paper. Real application examples and implementation methodology are outlined. Finally, future trends and challenges are presented.This paper is a full version of the invited plenary talk presented by the author on the 1st July 2009 at the 7th IFAC Symposium Safeprocess '09, Barcelona.     

Two fault detection and isolation schemes for robot manipulators using soft computing techniques

With growing technology, fault detection and isolation (FDI) have become one of the interesting and important research areas in modern control and signal processing. Accomplishment of specific missions like waste treatment in nuclear reactors or data collection in space and underwater missions make reliability more important for robotics and this demand forces researchers to adapt available FDI studies on nonlinear systems to robot manipulators, mobile robots and mobile manipulators.In this study, two model-based FDI schemes for robot manipulators using soft computing techniques, as an integrator of Neural Network (NN) and Fuzzy Logic (FL), are proposed. Both schemes use M-ANFIS for robot modelling. The first scheme isolates faults by passing residual signals through a neural network. The second scheme isolates faults by modelling faulty robot models for defined faults and combining these models as a generalized observers scheme (GOS) structure. Performances of these schemes are tested on a simulated two-link planar manipulator and simulation results and a comparison according to some important FDI specifications are presented.     

Fault detection and isolation using viability theory and interval observers

This paper proposes the use of interval observers and viability theory in fault detection and isolation (FDI). Viability theory develops mathematical and algorithmic methods for investigating the viability constraints characterisation of dynamic evolutions of complex systems under uncertainty. These methods can be used for checking the consistency between observed and predicted behaviour by using simple sets that approximate the exact set of possible behaviour (in the parameter or state space). In this paper, FDI is based on checking for an inconsistency between the measured and predicted behaviours using viability theory concepts and sets. Finally, an example is provided in order to show the usefulness of the proposed approach.     

Structural analysis for the sensor location problem in fault detection and isolation

In this paper we tackle the sensor location problem for fault detection and isolation based on structural analysis for linear systems with faults. We deal with this problem when the system under consideration is structured, that is, the entries of the system matrices are either fixed zeros or free parameters. With such structured systems one can associate a graph. A dedicated residual set is designed using a bank of observers for solving the problem. A major tool in this analysis is the notion of input separator in the associated graph, these separators form a lattice structure. The main contribution of this paper is the formulation of necessary and sufficient solvability conditions for the problem in terms of number of additional sensors measuring variables between faults and input separators in the associated graph.     

Actuator fault diagnosis for a class of nonlinear systems and its application to a laboratory 3D crane

In this paper, an actuator fault diagnosis scheme is proposed for a class of affine nonlinear systems with both known and unknown inputs. The scheme is based on a novel input/output relation derived from the considered nonlinear systems and the use of the recently developed high-order sliding-mode robust differentiators. The main advantages of the proposed approach are that it does not require a design of nonlinear observer and applies to systems not necessarily detectable. Conditions are provided to characterize the feasibility of fault detection and isolation using the proposed scheme and the maximum number of isolatable actuator faults. The efficacy of the proposed actuator fault diagnosis approach is tested through experiments on a laboratory 3D Crane, and the experimental results show that the proposed actuator fault diagnosis approach is promising and can achieve fault detection and isolation satisfactorily.     

On the disturbed fault detection and isolation problem

In this paper we study the fault detection and isolation problem in presence of disturbances. In the case of observer-based residual generation, the problem amounts to finding two gain matrices such that two problems are simultaneously solved. These problems are insensitivity of the residuals to disturbances and the existence of some special structure for the transfer from faults to residuals. We prove in this paper that this joint problem can be solved if and only if the usual (undisturbed) fault detection and isolation problem can be solved for a system with a reduced number of outputs.     

Active fault detection and isolation of discrete-time linear time-varying systems: a set-membership approach

Active fault detection and isolation (AFDI) is used for detection and isolation of faults that are hidden in the normal operation because of a low excitation signal or due to the regulatory actions of the controller. In this paper, a new AFDI method based on set-membership approaches is proposed. In set-membership approaches, instead of a point-wise estimation of the states, a set-valued estimation of them is computed. If this set becomes empty the given model of the system is not consistent with the measurements. Therefore, the model is falsified. When more than one model of the system remains un-falsified, the AFDI method is used to generate an auxiliary signal that is injected into the system for detection and isolation of faults that remain otherwise hidden or non-isolated using passive FDI (PFDI) methods. Having the set-valued estimation of the states for each model, the proposed AFDI method finds an optimal input signal that guarantees FDI in a finite time horizon. The input signal is updated at each iteration in a decreasing receding horizon manner based on the set-valued estimation of the current states and un-falsified models at the current sample time. The problem is solved by a number of linear and quadratic programming problems, which result in a computationally efficient algorithm. The method is tested on a numerical example as well as on the pitch actuator of a benchmark wind turbine.