Search

The width of a neural network matters since increasing the width will necessarily increase the model capacity. However, the performance of a network does not improve linearly with the width and soon gets saturated. In this case, we argue that increasing the number of networks (ensemble) can achieve better accuracy-efficiency trade-offs than purely increasing the width. To prove it, one large network is divided into several small ones regarding its parameters and regularization components. Each of these small networks has a fraction of the original one's parameters. We then train these small networks together and make them see various views of the same data to increase their diversity. During this co-training process, networks can also learn from each other. As a result, small networks can achieve better ensemble performance than the large one with few or no extra parameters or FLOPs, \ie, achieving better accuracy-efficiency trade-offs. Small networks can also achieve faster inference speed than the large one by concurrent running. All of the above shows that the number of networks is a new dimension of model scaling. We validate our argument with 8 different neural architectures on common benchmarks through extensive experiments. The code is available at \url{https://github.com/FreeformRobotics/Divide-and-Co-training}., Comment: accepted by T-IP 2022, code is at https://github.com/FreeformRobotics/Divide-and-Co-training

This paper introduces a cosmic expansion model with constant speed of cosmic spatial expansion via derivation and simulations, where the speed of cosmic spatial expansion equals the speed of light c. Simulation results show that the earliest observable universe time is t = 5.084 Gyrs where the current universe time T = 13.82 Gyrs, and the furthest observable distance at the earliest observable universe time t is S = 0.632R, where R is the cosmic radius at current universe time T. The above constant cosmic expansion model does not consider the inflation period in the early universe according to the Big Bang model, nor does it considered the cosmic acceleration in recent universe time. However, this simplified cosmic expansion model could be a benchmark that will be helpful to understand the cosmic expansion and the observable universe. Based on the derivation and simulation of the constant cosmic expansion model, the threshold of observable universe for the accelerated cosmic expansion model can also be calculated similarly, as far as the speed of cosmic spatial expansion at any universe time t can be provided.

Humans have a remarkable ability to learn continuously from th e environment and inner experience. One of the grand goals of robots is to build an artificial "lifelong learning" agent that can shape a cultivated understanding of the world from the current scene and previous knowledge via an autonomous lifelong development. It is challenging for the robot learning process to retain earlier knowledge when robots encounter new tasks or information. Recent advances in computer vision and deep -learning methods have been impressive due to large-scale data sets, such as ImageNet and COCO. However, robotic vision poses unique new challenges for applying visual algorithms developed from these computer vision data sets because they implicitly assume a fixed set of categories and time -invariant task distributions.

Scene recognition is a fundamental task in robotic perception. For human beings, scene recognition is reasonable because they have abundant object knowledge of the real world. The idea of transferring prior object knowledge from humans to scene recognition is significant but still less exploited. In this paper, we propose to utilize meaningful object representations for indoor scene representation. First, we utilize an improved object model (IOM) as a baseline that enriches the object knowledge by introducing a scene parsing algorithm pretrained on the ADE20K dataset with rich object categories related to the indoor scene. To analyze the object co-occurrences and pairwise object relations, we formulate the IOM from a Bayesian perspective as the Bayesian object relation model (BORM). Meanwhile, we incorporate the proposed BORM with the PlacesCNN model as the combined Bayesian object relation model (CBORM) for scene recognition and significantly outperforms the state-of-the-art methods on the reduced Places365 dataset, and SUN RGB-D dataset without retraining, showing the excellent generalization ability of the proposed method. Code can be found at https://github.com/hszhoushen/borm., Comment: 8 pages, 5 figures, conference, Accepted by IROS2021

Accurate perception of the surrounding scene is helpful for robots to make reasonable judgments and behaviours. Therefore, developing effective scene representation and recognition methods are of significant importance in robotics. Currently, a large body of research focuses on developing novel auxiliary features and networks to improve indoor scene recognition ability. However, few of them focus on directly constructing object features and relations for indoor scene recognition. In this paper, we analyze the weaknesses of current methods and propose an Object-to-Scene (OTS) method, which extracts object features and learns object relations to recognize indoor scenes. The proposed OTS first extracts object features based on the segmentation network and the proposed object feature aggregation module (OFAM). Afterwards, the object relations are calculated and the scene representation is constructed based on the proposed object attention module (OAM) and global relation aggregation module (GRAM). The final results in this work show that OTS successfully extracts object features and learns object relations from the segmentation network. Moreover, OTS outperforms the state-of-the-art methods by more than 2\% on indoor scene recognition without using any additional streams. Code is publicly available at: https://github.com/FreeformRobotics/OTS., Comment: Accepted by IROS2021

In this paper, a novel soft inflatable arm is proposed for telepresence robots. The new proposed structure of the arm is achieved by a very common and low-cost inflatable material and it is very light, weighing only about 50 g. However, it can realize agile movement by driving six tiny cables installed in the shoulder and elbow joints. The soft inflatable arm can work by pumping air at a very low pressure ( $7.32 \pm 3.45$ kPa) and allows direct and soft human contact without any external force sensors. This paper proposes joint compressed models, joint kinematic models, and kinematic models for the whole inflatable robot arm. These models can be easily applied to multijoint arms. In addition, a new redundancy resolution method is also developed for the inflatable arm, which makes it easier to control resolution and is less complex than other traditional approaches. Numerous experiments have been conducted, including performances of accuracy, repeatability, motion trajectories, human-safe interaction, and remote interaction. Results are satisfactory and validate the expected performance of the proposed robotic arm.

Redundant space manipulators, including Space Station Remote Manipulator System (SSRMS), Special Purpose Dexterous Manipulator (SPDM) and European Robotic Arm (ERA), have been playing important roles in the construction and maintenance of International Space Station (ISS). They all have 7 revolute joints arranged in similar configurations, and are referred to as SSRMS-type manipulators. When a joint is locked in an arbitrary position due to some failures, a 7R manipulator degrades to a 6R manipulator. Without a spherical wrist or three consecutive parallel joints, the inverse kinematics of the 6R manipulator is very complex. In this paper, we propose effective methods to resolve the inverse kinematics for different cases of any joint locked in an arbitrary position. Firstly, configuration characteristics of the SSRMS-type redundant manipulators are analyzed. Then, an existing of closed-form inverse kinematics is discussed for locking different joints. Secondly, D–H frames and corresponding D–H parameters of the new 6-DOF manipulator formed by locking a joint in an arbitrary position are re-constructed. A unified table is then created to describe the kinematics for all possible cases of single joint locking failure. Thirdly, completely analytical and semi-analytical methods are presented to solve the inverse kinematics equations, and the former is used for locking joint 1, 2, 6 or 7 while the latter for locking joint 3, 4 or 5. Finally, typical cases for single joint locking are studied. The results verify the proposed methods.

This paper presents an online dynamic hand gesture recognition system with an RGB-D camera, which can automatically recognize hand gestures against complicated background. For background subtraction, we use a model-based method to perform human detection and segmentation in the depth map. Since a robust hand tracking approach is crucial for the performance of hand gesture recognition, our system uses both color information and depth information in the process of hand tracking. To extract spatio-temporal hand gesture sequences in the trajectory, a reliable gesture spotting scheme with detection on change of static postures is proposed. Then discrete HMMs with Left-Right Banded (LRB) topology are utilized to model and classify gestures based on multi-feature representation and quantization of the hand gesture sequences. Experimental evaluations on two self-built databases of dynamic hand gestures show the effectiveness of the proposed system. Furthermore, we develop a human-robot interactive system, and the performance of this system is demonstrated through interactive experiments in the dynamic environment.

Improvements in hybrid electric vehicle (HEV) fuel economy and emissions heavily depend on an efficient energy management strategy (EMS). However, the uncertainty of future driving conditions generally cannot be easily tackled in EMS design. Most existing EMSs act upon fixed parameters and cannot adapt to varying driving conditions. Therefore, they usually fail to fully explore the potential of these advanced vehicles. In this paper, a novel EMS design procedure based on neural dynamic programming (NDP) is proposed. The NDP is a generic online learning algorithm, which combines stochastic dynamic programming (SDP) and the temporal difference (TD) method. Instead of computing the utility function and optimal control actions through Bellman equations, the NDP algorithm uses two neural networks to approximate them. The weights of these neural networks are updated online by the TD method. It avoids the high computational cost that SDP suffers from and is suitable for real-time implementation. The main advantages of NDP EMS is that it does not rely on prior information related to future driving conditions, and can self-tune with a wide variance in operating conditions. The NDP EMS has been applied to “Qianghua-I”, a prototype of a parallel HEV, using a revolving drum test bench for verification. Experiment results illustrate the potential of the proposed EMS in terms of fuel economy and in keeping state of charge (SOC) deviations at a low level. The proposed research ensures the optimality of NDP EMS, as well as real-time applicability.

In this paper, we propose an autonomous tree climbing strategy for a novel tree climbing robot that is named Treebot. The proposed algorithm aims to guide Treebot in climbing along an optimal path by the use of minimal sensing resources. Inspired by inchworms, the algorithm reconstructs the shape of a tree simply by the use of tactile sensors. It reveals how the realization of an environment can be achieved with limited tactile information. An efficient nonholonomic motion planning strategy is also proposed to make Treebot climb on an optimal path. This is accomplished by the prediction of the future shape of the tree. The study that is presented in this paper also includes the formulation of Treebot kinematics and an analysis of the workspace of Treebot on different shapes of a tree. Numerous experiments have been conducted to evaluate the proposed autonomous climbing algorithm and to unveil the ability of Treebot.

Space robotic systems are expected to play an increasingly important role in future space activities. Nevertheless, dynamics modeling and motion planning of a space robot are much more complex than those of a fixed-base robot, due to the dynamic coupling between the manipulator and its base. On the other hand, in order to assure the success of on-orbital missions, many experiments are required to verify the key algorithms on the ground before the space robot is launched. In this paper, the main research achievements on dynamics modeling, path planning, and ground verification are reviewed, and future studies are recommended. Firstly, we summarize the essential modeling concepts, and deduce the kinematics and dynamics equations of a space robot. Secondly, the main motion planning approaches are discussed. Then, different ground verification systems, including the air-bearing table, neutral buoyancy, airplane flying, free-falling motion, suspension system, and hybrid system, are introduced. Finally, the future research trends are forecasted.

Space robotic systems are expected to play an increasingly important role in the future. However, the control methods based on the inverse kinematics are affected by singularities. In this paper, practical approaches are proposed to solve the problems of a wrist-partitioned space manipulator. For spacecraft-referenced end-point motion control, we presented the singularity separation plus damped reciprocal (SSPDR) method, which separates the singularity parameters from the inverse of the Jacobian, and replaces their reciprocals using the damped reciprocals. For another control strategy, i.e. inertially referenced end-point motion control, including spacecraft attitude-controlled mode and free-floating mode, the linear momentum equation is used to eliminate three independent variables. With modifying some expressions, the SSPDR method is utilized to solve the singularities of spacecraft attitude-controlled space robot. When the space robot is free-floating, the singularities, i.e. the so-called dynamic singularities, cannot be predicted according to its kinematic structure. Combining with the measured angular velocity of the base, the dynamic singularity handling problem is transformed into real-time kinematic singularity avoiding problem, which can be solved by the SSPDR method. Since the SVD decomposition, the estimation of the minimum singularity value, and the calculation of the generalized Jacobian matrix are not required, the algorithm has lower computation load. Another advantage is that, only the accuracy of part velocity components is reduced by adding the damped coefficients. Simulation results verify the proposed approaches.

Early automatic detection of cardiovascular diseases is of great importance to provide timely treatment and reduce fatality rate. Although many efforts have been devoted to detecting various arrhythmias, classification of other common cardiovascular diseases still lacks comprehensive and intensive studies. This work aims at developing an automatic diagnosis system for myocardial infarction, valvular heart disease, cardiomyopathy, hypertrophy, and bundle branch block, based on the clinic recordings provided by PTB Database. The proposed diagnosis system consists of the components as baseline wander reduction, beat segmentation, feature extraction, feature reduction and classification. The selected features are the location, amplitude and width of each wave, exactly the parameters of ECG dynamical model. We also propose a mean shift algorithm based method to extract these features. To demonstrate the availability and efficacy of the proposed system, we use a total of 13,564 beats to conduct a large scale experiment, where only 25% beats are utilized to train the eigenvectors of generalized discriminant analysis in the feature reduction phase and 25% beats are applied to train the support vector machine in the classification phase. The average sensitivity, specificity and positive predicitivity for the test set, containing 75% beats, are respectively 96.06%, 99.32% and 97.29%.

Denoising of electrocardiogram (ECG) is the fundamental technique for manual or automatic ECG diagnosis. Model-based denoising has attracted initial studies since the ECG dynamical model was established in 2003 and been demonstrated to outperform most model-less denoising methods. The focus of this paper is robust denoising of abnormal ECG signals, which do not satisfy the assumption in previous model-based studies that morphological or physiological variations are small from one beat to another. A mean shift based initializer is proposed to provide a much more robust estimation of initial model parameters for each heart beat. Together with physiological knowledge based wave sub-segmentation and enhanced strategies, the novel initializer has been demonstrated to achieve satisfactory performance for both normal and abnormal heart beats under both white and pink noises. Utilizing records from Massachusetts Institute of Technology (MIT)-Beth Israel Hospital (BIH) database, this paper also applies various filters to denoise noisy signals and the denoising performances verify the availability and efficacy of the proposed denoising method.

For the aging population, surveillance in household environments has become more and more important. In this paper, we present a household robot that can detect abnormal events by utilizing video and audio information. In our approach, moving targets can be detected by the robot using a passive acoustic location device. The robot then tracks the targets by employing a particle filter algorithm. To adapt to different lighting conditions, the target model is updated regularly based on an update mechanism. To ensure robust tracking, the robot detects abnormal human behavior by tracking the upper body of a person. For audio surveillance, Mel frequency cepstral coefficients (MFCC) is used to extract features from audio information. Those features are input to a support vector machine classifier for analysis. Experimental results show that the robot can detect abnormal behavior such as "falling down" and "running". Also, a 88.17% accuracy rate is achieved in the detection of abnormal audio information like "crying", "groan", and "gun shooting". To lower the false alarms by abnormal sound detection system, the passive acoustic location device directs the robot to the scene where abnormal events occur and the robot can employ its camera to further confirm the occurrence of the events. At last, the robot will send the image captured by the robot to the mobile phone of master.

Space robots are playing an increasingly important role in on-orbital servicing, including repairing, refueling, or de-orbiting the satellite. The target must be captured and berthed before the servicing task starts. However, the attitude of the base may lean much and needs re-orientating after capturing. In this paper, a method is proposed to berth the target, and re-orientate the base at the same time, using manipulator motion only. Firstly, the system state is formed of the attitude quaternion and joint variables, and the joint paths are parameterized by sinusoidal functions. Then, the trajectory planning is transformed to an optimization problem. The cost function, defined according to the accuracy requirements of system variables, is the function of the parameters to be determined. Finally, we solve the parameters using the particle swarm optimization algorithm. Two typical cases of the spacecraft with a 6-DOF manipulator are dynamically simulated, one is that the variation of base attitude is limited; the other is that both the base attitude and the joint rates are constrained. The simulation results verify the presented method.

In this paper, we discuss the problem of how human control strategy can be represented as a parametric model using a support vector machine (SVM) and how an SVM-based controller can be used in controlling dynamically stable systems. The SVM approach has been implemented in the balance control of Gyrover, which is a dynamically stable, statically unstable, single-wheel mobile robot. The experimental results that compare SVM with general artificial neural network approaches clearly demonstrate the superiority of the SVM approach with regard to human control strategy learning.

This paper introduces a face detection, classification and counting system that is robust and works in real-time. It tracks multiple people, which is useful for face counting. The classification problem is defined as differentiating and then classifying the front of a face into Asian or non-Asian categories. The first step, is principal component analysis (PCA) for feature generation and independent component analysis (ICA) for feature extraction. Then, we employ support vector machine (SVM) for the training process and combine different SVM classifiers to create new classifiers, which improves the classification rate. Based on this, we can count the number of Asians and non-Asians. Experiments show that our system achieves a classification rate of 82.5% based on a database containing 750 face images from FERET.

Space robotic systems are expected to play an increasingly important role in the future. The robotic on-orbital service, whose key is the capturing technology, becomes research hot in recent years. In this paper, the authors propose an autonomous path planning method for target capturing. The task is described in Cartesian space and it can drive the manipulator to approach the target along the closest path. Firstly, the target feature is extracted based on the measured information via the hand-eye camera, and the target pose (position and orientation) and velocities (linear velocity and angular velocity) are estimated using Kalman filtering technology. Then, a numerically feasible approach is presented to plan the manipulator motion and avoid the dynamic singularities, which are transformed into real-time kinematic singularities avoiding problem. Thirdly, the potential disturbance on the base due to the manipulator's motion is estimated, and the joint rates are autonomously adjusted to reduce the disturbance if it is beyond the allowed bound. At last, a ground experiment system is set up based on the concept of dynamic emulation and kinematic equivalence. With the experiment system, the autonomous target capturing experiments are conducted. The experiment results validate the proposed algorithm.

Detecting human abnormal behaviors in a crowded situation is a challenging problem for public security departments. Human abnormal behaviors are seen either in an individual or in a crowd. This paper is focused on the latter, which is more crucial at some important spots and has been less studied. To achieve this goal, we define two categories of video energy based on intensity variation and motion features and adopt two surveillance methods for the two energy accordingly. Using wavelet analysis of the energy curves, we have obtained a result which shows that both methods can be used to deal with crowd modeling and real-time surveillance satisfactorily. A comparison of the two methods is then made in the actual environment in a metro surveillance system.

This paper presents a method for modeling abnormal human gait using hidden Markov model under the framework of a shoe-integrated system. The intelligent system focuses on modeling the following patterns: normal gait, toe in abnormality, and toe out abnormality. In the developed prototype, an inertial measurement unit (IMU) consisting of three-dimensional gyroscopes and accelerometers is employed to measure the angular velocities and accelerations of the foot. Four force sensing resistors (FSRs) and one bend sensor are arranged on the insole of each foot for force and flexion information acquisition. The proposed method is mainly based on supervised Principal Component Analysis (SPCA) for feature generation and hidden Markov model (HMM) for multi-pattern modeling. The "similarity distance measure" criterion is introduced to do model-to-model evaluation. Experimental results demonstrate the proposed approach is robust and efficient in detecting abnormal gait patterns. Our goal is to provide a cost-effective system for detecting gait abnormalities in order to assist persons with abnormal gaits in developing a normal walking pattern in their daily life.

Coded structured light is an active approach that obtains the shape of object. In our previous work, we proposed a new binary projection mechanism for inspecting semiconductor products. The mechanism consists of only a single light source in combination with a binary grating for projecting binary pattern. By shifting the binary grating in space and in every shifting taking a separate image of the illuminated surface, each position on the illuminated surface is attached with a unique binary string for correspondence. In the mechanism the inspection speed is governed by the number of needed images which also equals to the number of shiftings of the grating. To reduce image number, we propose that two neighbor bits are combined together to produce unique codeword, which is referred to as bit-pairing mechanism. This paper addresses how the spatial shifting of bit-pairing mechanism can be designed optimally for minimizing this number for faster inspection speed. An optimal solution to shifting strategy optimization is proposed that is applicable to any given binary patterns. Theoretical analysis and real image experiments are presented to illustrate the workability of the solutions.

Human gait is a dynamic biometrical feature which is complex and difficult to imitate. It is unique and more secure than static features such as passwords, fingerprints and facial features. In this paper, we present intelligent shoes for human identification based on human gait modeling and similarity evaluation with hidden Markov models (HMMs). Firstly we describe the intelligent shoe system for collecting human dynamic gait performance. Using the proposed machine learning method hidden Markov models, an individual wearer's gait model is derived and we then demonstrate the procedure for recognizing different wearers by analyzing the corresponding models. Next, we define a hidden-Markov-model-based similarity measure which allows us to evaluate resultant learning models. With the most likely performance criterion, it will help us to derive the similarity of individual behavior and its corresponding model. By utilizing human gait modeling and similarity evaluation based on hidden Markov models, the proposed method has produced satisfactory results for human identification during testing.

Technology for the capturing of space target is very important for on-orbit servicing. In order to assure the task is accomplished successfully, ground experimentations are required for the verification of the planning and control algorithms of space robotic system before it is launched. In this paper, an experiment concept is used, which is a hybrid approach, i.e. it combines the mathematical model with the physical model. The key issues of the concept are dynamic emulation and kinematic equivalence, in which the behaviors of the space robotic system are calculated by its dynamic equations. The motion of its end-effector and the space target is realized by two industrial robots. According to different observation spots, two modes of capturing process are emulated: one is observed from the inertial frame, the other is from the space base. Based on the concept proposed above, a ground experiment system is set up, which is composed of two industrial robots, a set of global visual system and five industrial computers. Using the system, algorithms of space robot of any geometry and mass properties can be tested. As an example, the autonomous trajectory planning algorithm is verified by the experiment of capturing a moving target. Moreover, a real-time 3D simulation system is developed to emulate the capturing process in 3D space. Numeric simulation and experiment results show that the ground system is effective in evaluating the planning and control algorithms of space robot.

Recently, audio analysis has been proposed by some researchers for surveillance application. Compared with video surveillance, the effectiveness of audio surveillance is not influenced by the occlusions. In this paper, we propose a learning based approach to intelligent audio surveillance in household environment. This approach can be applied to static or mobile audio surveillance systems which work solely or work as the supplements of video surveillance systems. After extracting Mel frequency cepstral coefficients (MFCC) from a 1.0 s waveform, a classifier trained from a labeled dataset using modest AdaBoost is employed to determine whether this waveform is normal or abnormal. The result and simulation show the effectiveness of this approach.

In practice, the environments in which mobile robots operate are usually modeled in highly complex forms and, as a result, autonomous navigation and localization can be difficult. The difficulties are exacerbated for practical robots with limited on-board computational resources and complex planning algorithms, since this paradigm of environmental modeling requires enormous computational power. A novel navigation/localization learning methodology is presented to abstract and transfer the human sequential navigational skill to a robotic wheelchair by showing the platform how to respond in different local environments along a demonstrated, static cluttered route using a lookup table representation. This method utilizes limited on-board range sensing information to concisely model local unstructured environments, with respect to the robot, for navigation or localization along the learned route in order to achieve good performance with low on-line computational demand and low-cost hardware requirements. Experimental study demonstrates the feasibility of this method and some interesting characteristics of navigation, localization, and environmental modeling problems. Analysis is also conducted to investigate performance evaluation, advantages of the approach, choices of lookup table inputs and outputs, and potential generalization of this paper

This paper introduces a real-time video surveillance system which can track people and detect human abnormal behaviors. In the blob detection part, an optical flow algorithm for crowd environment is studied experimentally and a comparison study with respect to traditional subtraction approach is carried out. The different approaches in segmentation and tracking enable the system to track persons when they change movement unpredictably in occlusion. We developed two methods for the human abnormal behavior analysis. The first one employs Principal Component Analysis for feature selection and Support Vector Machine for classification of human behaviors. The proposed feature selection method is based on the border information of four consecutive blobs. The second approach computes optical flow to obtain the velocity of each pixel for determining whether a human behavior is normal or not. Both algorithms are successfully developed in crowded environments to detect the following human abnormal behaviors: (1) Running people in a crowded environment; (2) falling down movement while most are walking or standing; (3) a person carrying an abnormal bar in a square; (4) a person waving hand in the crowd. Experimental results demonstrate these two methods are robust in detecting human abnormal behaviors.

Global manufacturing chains have become common in the 21st century. To perform remote monitoring and diagnosis in such chains, real-time data compression has become a core factor in the efficient and effective exchange of information via computer networks. This paper presents a new technique for compressing data using a kernel-based method. Overcoming the drawbacks of support vector techniques – that is, fast decompression but slow compression – the new method exhibits high speed in both phases. In addition, the new method can also be applied for pattern classification. Based on strain signal, example tests derived from sheet metal stamping operations, the new method is shown to be very effective. The proposed technology has enormous potential in the application of advanced manufacturing processes monitoring and control through Internet or Intranet.

Surveillance of public places has become a worldwide concern in recent years. The ability to identify abnormal human behaviors in real-time is fundamental to the success of intelligent surveillance systems. The recognition of abnormal and suspicious human walking patterns is an important step towards the achievement of this goal. In this research, we develop an intelligent visual surveillance system that can classify normal and abnormal human walking trajectories in outdoor environments by learning from demonstration. The system takes into account both the local and global characteristics of the observed trajectories and is able to identify their normality in real-time. By utilizing support vector learning and a similarity measure based on hidden Markov models, the developed system has produced satisfactory results on real-life data during testing. Moreover, we utilize the approach of longest common subsequence (LCSS) in determining the similarity between different types of walking trajectories. In order to establish the position and speed boundaries required for the similarity measure, we compare the performance of a number of approaches, including fixed boundary values, variable boundary values, learning boundary by support vector regression, and learning boundary by cascade neural networks.

Sensor signals produced in industrial manufacturing processes contain valuable information about the condition of operations. However, extracting the appropriate feature for effective fault diagnosis is difficult. Moreover, the adaptability and flexibility of current fault diagnosis systems are often found wanting in real-world applications. Unfortunately, it is essential to rebuild most fault diagnosis systems when new fault types emerge. This paper presents an intelligent fault diagnosis system based on a hidden Markov model. Introducing the concepts of time marginal energy and frequency marginal energy, the features of which can be acquired by the wavelet packet technique satisfy the requirements for fault diagnosis. By utilizing the best tree principle, this method not only extracts the feature automatically without a priori experience but also compresses the data; both of which ensure a system that is practical for real-time application. The new diagnosis system developed here is efficient and effective, as demonstrated by the model developed and applied to a real-time sheet metal stamping process. Based on tests conducted during two experiments (one based on simple blanking, the other on progressive operations) and related comparisons, the proposed method is substantially more effective than other approaches. In addition, the new method requires that only related models be created for new fault types, which results show are ideal for shop floor applications.

In this research, computational intelligence techniques are applied towards the modeling of human sensations in virtual environments. We specifically focus on the following important questions: (1) how to efficiently model the relationship between human sensations and the physical stimuli presented to humans, (2) how to validate the human sensation models, and (3) how to reduce the size of the input data when it gets large and how to select the information which is most important to human sensation modeling. In order to provide an experimental testbed for the implementation of the proposed learning and analysis techniques, a full-body motion virtual reality interface capable of recording human sensations is developed. We propose using cascade neural networks with node-decoupled extended Kalman filter training for modeling human sensation in virtual environments. For the purpose of sensation model validation, we propose using a stochastic similarity measure based on hidden Markov models to calculate the relative similarity between model-generated sensation and actual human sensation. Next, we investigate a number of feature extraction and input selection techniques for reducing the input data size in human sensation modeling. We propose and develop a new input selection method based on independent component analysis, which is capable of reducing the data size and selecting the stimuli information that is most important to the human sensation.