Dustin Harvey

Automated Feature Design for Numeric Sequence Classification by Genetic Programming

Pattern recognition methods rely on maximum-information, minimum-dimension feature sets to reliably perform classification and regression tasks. Many methods exist to reduce feature set dimensionality and construct improved features from an initial set; however, there are few general approaches for the design of features from numeric sequences. Any information lost in preprocessing or feature measurement cannot be recreated during pattern recognition. General approaches are needed to extend pattern recognition to include feature design and selection for numeric sequences, such as time series, within the learning process itself. This paper proposes a novel genetic programming (GP) approach to automated feature design called Autofead. In this method, a GP variant evolves a population of candidate features built from a library of sequence-handling functions. Numerical optimization methods, included through a hybrid approach, ensure that the fitness of candidate algorithms is measured using optimal parameter values. Autofead represents the first automated feature design system for numeric sequences to leverage the power and efficiency of both numerical optimization and standard pattern recognition algorithms. Potential applications include the monitoring of electrocardiogram signals for indications of heart failure, network traffic analysis for intrusion detection systems, vibration measurement for bearing condition determination in rotating machinery, and credit card activity for fraud detection.

Automated extraction of damage features through genetic programming

Robust damage detection algorithms are a fundamental requirement for development of practical structural health monitoring systems. Typically, structural health-related decisions are made based on measurements of structural response. Data analysis involves a two-stage process of feature extraction and classification. While classification methods are well understood, feature design is difficult, time-consuming, and requires application experts and domain-specific knowledge. Genetic programming, a method of evolutionary computing closely related to genetic algorithms, has previously shown promise when adapted to problems involving structured data such as signals and images. Genetic programming evolves a population of candidate solutions represented as computer programs to perform a well-defined task. Importantly, genetic programming conducts an efficient search without specification of the size of the desired solution. In this study, a novel formulation of genetic programming is introduced as an automated feature extractor for supervised learning problems related to structural health monitoring applications. Performance of the system is evaluated on signal processing problems with known optimal solutions.

Characterization and Prognosis of Multirotor Failures

Multirotor (MR) unmanned aviation systems are becoming more prevalent in the commercial, philanthropic, and military communities. Because of these public environment applications, hardware malfunctions pose serious safety concerns. Propeller, motor, and structural damage can cause substantial failure of the MR vehicle and endanger surrounding people and structures; thus, early identification and prognosis of these failure modes is necessary to mitigate harm. An embedded structural health monitoring (SHM) system is optimal for identification and diagnosis of these failure modes in time to alter or abort the mission. To achieve autonomous SHM, statistical data must be accrued from a series of sensor measurements. This information is utilized in the development of appropriate damage metrics for failure modes of interest, which determine the real-time state of hardware elements. A comprehensive sensor network was successfully designed and implemented on an MR vehicle to determine which instruments provide valuable information. Utilizing this relevant data, a compatible set of tools was developed for signal processing, and the resulting SHM system is capable of classifying propeller, motor, and structural hardware failures.

Structural health monitoring feature design by genetic programming

Structural health monitoring (SHM) systems provide real-time damage and performance information for civil, aerospace, and other high-capital or life-safety critical structures. Conventional data processing involves pre-processing and extraction of low-dimensional features from in situ time series measurements. The features are then input to a statistical pattern recognition algorithm to perform the relevant classification or regression task necessary to facilitate decisions by the SHM system. Traditional design of signal processing and feature extraction algorithms can be an expensive and time-consuming process requiring extensive system knowledge and domain expertise. Genetic programming, a heuristic program search method from evolutionary computation, was recently adapted by the authors to perform automated, data-driven design of signal processing and feature extraction algorithms for statistical pattern recognition applications. The proposed method, called Autofead, is particularly suitable to handle the challenges inherent in algorithm design for SHM problems where the manifestation of damage in structural response measurements is often unclear or unknown. Autofead mines a training database of response measurements to discover information-rich features specific to the problem at hand. This study provides experimental validation on three SHM applications including ultrasonic damage detection, bearing damage classification for rotating machinery, and vibration-based structural health monitoring. Performance comparisons with common feature choices for each problem area are provided demonstrating the versatility of Autofead to produce significant algorithm improvements on a wide range of problems.

Automated Selection of Damage Detection Features by Genetic Programming

Robust damage detection algorithms are the first requirement for development of practical structural health monitoring systems. Typically, a damage decision is made based on time series measurements of structural responses. Data analysis involves a two-stage process, namely feature extraction and classification. While classification methods are well understood, no general framework exists for extracting optimal, or even good, features from time series measurements. Currently, successful feature design requires application experts and domain-specific knowledge. Genetic programming, a method of evolutionary computing closely related to genetic algorithms, has previously shown promise as an automatic feature selector in speech recognition and image analysis applications. Genetic programming evolves a population of candidate solutions represented as computer programs to perform a well-defined task such as classification of time series measurements. Importantly, genetic programming conducts an efficient search without specification of the size of the desired solution. This preliminary study explores the use of genetic programming as an automated feature extractor for two-class supervised learning problems related to structural health monitoring applications.

Cointegration as a data normalization tool for structural health monitoring applications

The structural health monitoring literature has shown an abundance of features sensitive to various types of damage in laboratory tests. However, robust feature extraction in the presence of varying operational and environmental conditions has proven to be one of the largest obstacles in the development of practical structural health monitoring systems. Cointegration, a technique adapted from the field of econometrics, has recently been introduced to the SHM field as one solution to the data normalization problem. Response measurements and feature histories often show long-run nonstationarity due to fluctuating temperature, load conditions, or other factors that leads to the occurrence of false positives. Cointegration theory allows nonstationary trends common to two or more time series to be modeled and subsequently removed. Thus, the residual retains sensitivity to damage with dependence on operational and environmental variability removed. This study further explores the use of cointegration as a data normalization tool for structural health monitoring applications.

Echo: Data and Analysis Management

This document describes, Echo, a comprehensive suite of tools for data wrangling, management, and analysis, and its application.

Multi-Source Sensing and Analysis for Machine-Array Condition Monitoring

Early detection of damage in machines can eliminate the expenses and safety hazards associated with failure. Current methods use distributed systems to monitor individual machines, but the associated costs of instrumentation, data acquisition hardware, and facility retrofits are high. A centralized, remote, multi-source monitoring system would increase both the cost efficiency and ease of user operation. This study investigates how data from multiple sensing streams can best be utilized for decision making processes and to what extent a remote, centralized instrumentation package can effectively monitor multiple machines. An array of duct fans was used in this study to represent an arbitrary set of systems from which data can be collected and fused. Data was acquired using multiple measurement types (vibrations, acoustics, current, voltage, RF) using a centralized instrumentation system. The voltage and current were measured from the single supply used to power the fan array. The remaining sensors, including an accelerometer, microphone, and antenna, were placed in a single, central location among the array of fans. Data analysis focused on determining whether separate, nominally identical machines could be uniquely identified and characterized from measurements. Spectral analysis and signature development were used to characterize the state of each machine. These methods can be implemented in other applications involving the fusion of data from several sources to obtain information about the identification, location, and characterization of one or more dynamic systems.

Developing Conservative Mechanical Shock Specifications

Mechanical shock testing and analysis are integral parts of developing new high-value items and ensuring their capability to withstand the environments to which they will be exposed. Most conventional methods for specifying mechanical shock environments provide no mathematically defensible correlation between their parameters and the damage-causing potential of the environment, thus warranting the study of new methods to specify shock environments. In this paper, a variety of parameters that correlate to the damage-causing potential of a shock environment are identified. The parameters investigated are restricted to those that can be constrained during a real-time laboratory test. An analytical study using a sensitivity analysis determined the effect of each parameter on the damage-causing potential of a shock environment. These parameters are used to create shock specifications and investigate the simulated response of a structure. The parameters generated in this study improve mechanical shock testing by providing a strong correlation between the shock environment and the damage-causing potential of a mechanical shock.

Robust evaluation of time series classification algorithms for structural health monitoring

Structural health monitoring (SHM) systems provide real-time damage and performance information for civil, aerospace, and mechanical infrastructure through analysis of structural response measurements. The supervised learning methodology for data-driven SHM involves computation of low-dimensional, damage-sensitive features from raw measurement data that are then used in conjunction with machine learning algorithms to detect, classify, and quantify damage states. However, these systems often suffer from performance degradation in real-world applications due to varying operational and environmental conditions. Probabilistic approaches to robust SHM system design suffer from incomplete knowledge of all conditions a system will experience over its lifetime. Info-gap decision theory enables nonprobabilistic evaluation of the robustness of competing models and systems in a variety of decision making applications. Previous work employed info-gap models to handle feature uncertainty when selecting various components of a supervised learning system, namely features from a pre-selected family and classifiers. In this work, the info-gap framework is extended to robust feature design and classifier selection for general time series classification through an efficient, interval arithmetic implementation of an info-gap data model. Experimental results are presented for a damage type classification problem on a ball bearing in a rotating machine. The info-gap framework in conjunction with an evolutionary feature design system allows for fully automated design of a time series classifier to meet performance requirements under maximum allowable uncertainty.