Finn Tseng

An Industrial Strength Novelty Detection Framework for Autonomous Equipment Monitoring and Diagnostics

This paper presents a practical framework for autonomous monitoring of industrial equipment based on novelty detection. It overcomes limitations of current equipment monitoring technology by developing a “generic” structure that is relatively independent of the type of physical equipment under consideration. The kernel of the proposed approach is an “evolving” model based on unsupervised learning methods (reducing the need for human intervention). The framework employs procedures designed to temporally evolve the critical model parameters with experience for enhanced monitoring accuracy (a critical ability for mass deployment of the technology on a variety of equipment/hardware without needing extensive initial tune-up). Proposed approach makes explicit provision to characterize the distinct operating modes of the equipment, when necessary, and provides the ability to predict both abrupt as well as gradually developing (incipient) changes. The framework is realized as an autonomous software agent that continuously updates its decision model implementing an unsupervised recursive learning algorithm. Results of validation of the proposed methodology by accelerated testing experiments are also discussed.

Novelty Detection Based Machine Health Prognostics

In this paper we present a new novelty detection algorithm for continuous real time monitoring of machine health and prediction of potential machine faults. The kernel of the system is a generic evolving model that is not dependent on the specific measured parameters determining the health of a particular machine. Two alternative strategies are introduced in order to predict abrupt and gradually developing (incipient) changes. This algorithm is realized as an autonomous software agent that continuously updates its decision model implementing an unsupervisory recursive learning algorithm. Results of validation of the proposed algorithm by accelerated testing experiments are also discussed

Driving pattern identification for EV range estimation

This paper presents a driving pattern recognition method based on trip segment clustering. Driving patterns categorize various driving behaviors that contain certain energy demand property in common. It can be applied to various applications including intelligent transportation, emission estimation, passive/active safety controls and energy management controls. In this paper, pattern features are first identified from high impact factors from static and quasi-static environmental and traffic information. A feature based trip/route partitioning algorithm is then developed based on data clustering methods. The driving patterns are finally recognized by synthesizing all partitioned feature zones along the trip/route where each partitioned road section is distinguished by an attribute of feature combination that will result in a distinctive drive energy demand property. The driving pattern recognition is a critical technology especially in solving problems like range estimation and energy consumption preplanning for the plug-in capable electrified vehicles.

Real-time driver characterization during car following using stochastic evolving models

This paper studies characterizing the driving behavior during steady-state and transient car-following. An approach utilizing the online learning of an evolving Takagi-Sugeno fuzzy model that is combined with a probabilistic model is applied to capture the multi-model and evolving nature of the driving behavior. The approach is validated by testing on a vehicle during different driving conditions.

Real Time Novelty Detection Modeling for Machine Health Prognostics

The paper deals with a real time algorithm for modeling and prediction of machine health status. It utilizes the concepts of fuzzy k-nearest neighbor clustering and the Gaussian mixture model to model the machine feature space as a loose collection of clusters representing the dynamics of the main operating modes

Hidden-Markov model based sequential clustering for autonomous diagnostics

Despite considerable advances over the last few decades in sensing instrumentation and information technology infrastructure, monitoring and diagnostics technology has not yet found its place in health management of mainstream machinery and equipment. The fundamental reason for this being the mismatch between the growing diversity and complexity of machinery and equipment employed in industry and the historical reliance on ldquopoint-solutionrdquo diagnostic systems that necessitate extensive characterization of the failure modes and mechanisms (something very expensive and tedious). While these point solutions have a role to play, in particular for monitoring highly-critical assets, generic yet adaptive solutions, meaning solutions that are flexible and able to learn on-line, could facilitate large-scale deployment of diagnostic and prognostic technology. We present a novel approach for autonomous diagnostics that employs model-based sequential clustering with hidden-Markov models as a means for measuring similarity of time-series sensor signals. The proposed method has been tested on a CNC machining test-bed outfitted with thrust-force and torque sensors for monitoring drill-bits. Preliminary results revealed the competitive performance of the method.

A two-stage-training support vector machine approach to predicting unintentional vehicle lane departure

ABSTRACT Advanced driver assistance systems, such as unintentional lane departure warning systems, have recently drawn much attention and efforts. In this study, we explored utilizing the nonlinear binary support vector machine (SVM) technique to predict unintentional lane departure, which is innovative, as the SVM methodology has not previously been attempted for this purpose in the literature. Furthermore, we developed a two-stage training scheme to improve SVMs prediction performance in terms of minimization of the number of false positive prediction errors. Experiment data generated by VIRTTEX, a hydraulically powered, 6-degrees-of-freedom moving base driving simulator at Ford Motor Company, were used. All the vehicle variables were sampled at 50 Hz and there were 16 drowsy drivers (about 3 hours of driving per subject) and six control drivers (approximately 20 minutes f driving each). In total, 3,508 unintentional lane departures occurred for the drowsy drivers and 23 for the control drivers. Our study involving these 22 drivers with a total of more than 7.5 million prediction decisions demonstrates that (a) excellent SVM prediction performance, measured by numbers of false positives (i.e., falsely predicted lane departures) and false negatives (i.e., lane departures failed to be predicted), was achieved when the prediction horizon was 0.6 seconds or less, (b) lateral position and lateral velocity worked the best as SVM input variables among the nine variable sets that we explored, and (c) the radial basis function performed the best as the SVM kernel function.

A support vector machine approach to unintentional vehicle lane departure prediction

Advanced driver assistance systems, such as unintentional lane departure warning systems, have recently drawn much attention and R & D efforts. Such a system may assist the driver by monitoring the driver or vehicle behaviors to predict/detect driving situations (e.g, lane departure) and alert the driver to take corrective action. In this paper, we show how the support vector machine (SVM) methodology can potentially provide enhanced unintentional lane departure prediction, which is a new method relative to literature. Our binary SVM employed the Radial Basis Function kernel to classify time series of select vehicle variables. The SVM was trained and tested using the driver experiment data generated by VIRTTEX, a hydraulically powered 6-degrees-of-freedom moving base driving simulator at Ford Motor Company. The data that we used represented 16 drowsy subjects (three-hour driving time per subject) and six control subjects (20 minutes driving per subject), all of which drove a simulated 2000 Volvo S80. The vehicle variables were all sampled at 50 Hz. There were a total of 3,508 unintentional lane departure occurrences for the drowsy drivers and only 23 for four of the six control drivers (two had none). The SVM was trained by over 60,000 time series examples (the actual number depended on the prediction horizon) created from 50% of the lane departures. The training data were removed from the testing data. During the testing, the SVM made a lane departure prediction at every sampling time for every one of the 22 drivers (over 6.8 million predictions in total). The overall sensitivity and specificity of the SVM with a 0.2-second prediction horizon for the 22 drivers were 99.77465% and 99.99997%, respectively. The SVM predicted, on average 0.200181 seconds in advance, lane departure correctly for all the control drivers, but missed 4 of the 1,758 and gave false positives for another 2 for the drowsy drivers. For the prediction horizon of 0.4s, there was 1 false positive case for the control subjects, and the false negative and false positive cases rose substantially to 10 and 137 for the drowsy drivers, respectively.

Force-based weld quality monitoring algorithm

Traditionally, to check weld quality, destructive (the dominant method at industry until now) and non-destructive tests are used on randomly sampled work pieces at the production site. These processes tend to be predominantly off-line or end-of-line processes. The objective of the force-based weld monitoring algorithm is to classify the weld status, i.e. cold, normal and expulsion, by using non-intrusive force sensor. It is composed mainly from two sub-algorithms: expulsion and normal welds detection. This algorithm has been tested on medium frequency direct current (MFDC) constant current (CC) control. The results were as follows: the percentage of the false alarm (I) type-1 error is the lowest for cold welds case at 3.6%, 11.0% for normal welds and 21.7% for expulsion welds. As for the failed alarm (β) type-2 error, the lowest percentage is for expulsion welds 1.9%, 3.4% for cold welds and 16.5% for normal welds.

Building Efficient Probability Transition Matrix Using Machine Learning from Big Data for Personalized Route Prediction

Abstract Personalized route prediction is an important technology in many applications related to intelligent vehicles and transportation systems. Current route prediction technologies used in many general navigation systems are, by and large, based on either the shortest or the fastest route selection. Personal traveling route prediction is a very challenging big data problem, as trips getting longer and variations in routes growing. It is particularly challenging for real-time in-vehicle applications, since many embedded processors have limited memory and computational power. In this paper we present a machine learning algorithm for modeling route prediction based on a Markov chain model, and a route prediction algorithm based on a probability transition matrix. We also present two data reduction algorithms, one is developed to map large GPS based trips to a compact link-based standard route representation, and another a machine learning algorithm to significantly reduce the size of a probability transition matrix. The proposed algorithms are evaluated on real-world driving trip data collected in four months, where the data collected in the first three months are used as training and the data in the fourth month are used as testing. Our experiment results show that the proposed personal route prediction system generated more than 91% prediction accuracy in average among the test trips. The data reduction algorithm gave about 8:1 reduction in link-based standard route representation and 23:1 in reducing the size of probability transition matrix.