1 volume

International Telemetering Conference Proceedings / October 26-29, 1992 / Town and Country Hotel and Convention Center, San Diego, California / Our telemetry department has an application for a data categorization/compression of a high speed transient signal in a short period of time. Categorization of the signal reveals important system performance and compression is required because of the terminal nature of our telemetry testing. Until recently, the hardware for the system of this type did not exist. A new exploratory device from Intel has the capability to meet these extreme requirements. This integrated circuit is an analog neural network capable of performing 2 billion connections per second. The two main advantages of this chip over traditional hardware are the obvious computation speed of the device and the ability to compute a three layer feed-forward neural network classifier. The initial investigative development work using the Intel chip has been completed. The results from this proof of concept will show data categorization/compression performed on the neural network integrated circuit in real time. We will propose a preliminary design for a transient measurement system employing the Intel integrated circuit.

This thesis addresses the significance of speech features within the task of speaker recognition. Motivated by the perception of simple attributes like `loud', `smooth', `fast', more than 70 new speech features are developed. A set of basic speech features like pitch, loudness and speech speed are combined together with these new features in a feature set, one set per utterance. A neural network classifier is used to evaluate the significance of these features by creating a speaker recognition system and analysing the behaviour of successfully trained single-speaker networks. An in-depth analysis of network weights allows a rating of significance and feature contribution. A subjective listening experiment validates and confirms the results of the neural network analysis. The work starts with an extended sentence analysis; ten sentences are uttered by 630 speakers. The extraction of 100 speech features is outlined and a 100-element feature vector for each utterance is derived. Some features themselves and the methods of analysing them have been used elsewhere, for example pitch, sound pressure level, spectral envelope, loudness, speech speed and glottal-to-noise excitation. However, more than 70 of the 100 features are derivatives of these basic features and have not yet been described and used before in the speakerr ecognition research,e speciallyyn ot within a rating of feature significance. These derivatives include histogram, 3`d and 4 moments, function approximation, as well as other statistical analyses applied to the basic features. The first approach assessing the significance of features and their possible use in a recognition system is based on a probability analysis. The analysis is established on the assumption that within the speaker's ten utterances' single feature values have a small deviation and cluster around the mean value of one speaker. The presented features indeed cluster into groups and show significant differences between speakers, thus enabling a clear separation of voices when applied to a small database of < 20 speakers. The recognition and assessment of individual feature contribution jecomes impossible, when the database is extended to 200 speakers. To ensure continous vplidation of feature contribution it is necessary to consider a different type of classifier. These limitations are overcome with the introduction of neural network classifiers. A separate network is assigned to each speaker, resulting in the creation of 630 networks. All networks are of standard feed-forward backpropagation type and have a 100-input, 20- hidden-nodes, one-output architecture. The 6300 available feature vectors are split into a training, validation and test set in the ratio of 5-3-2. The networks are initially trained with the same 100-feature input database. Successful training was achieved within 30 to 100 epochs per network. The speaker related to the network with the highest output is declared as the speaker represented by the input. The achieved recognition rate for 630 speakers is -49%. A subsequent preclusion of features with minor significance raises the recognition rate to 57%. The analysis of the network weight behaviour reveals two major pointsA definite ranking order of significance exists between the 100 features. Many of the newly introduced derivatives of pitch, brightness, spectral voice patterns and speech speed contribute intensely to recognition, whereas feature groups related to glottal-to-noiseexcitation ratio and sound pressure level play a less important role. The significance of features is rated by the training, testing and validation behaviour of the networks under data sets with reduced information content, the post-trained weight distribution and the standard deviation of weight distribution within networks. The findings match with results of a subjective listening experiment. As a second major result the analysis shows that there are large differences between speakers and the significance of features, i. e. not all speakers use the same feature set to the same extent. The speaker-related networks exhibit key features, where they are uniquely identifiable and these key features vary from speaker to speaker. Some features like pitch are used by all networks; other features like sound pressure level and glottal-to-noise excitation ratio are used by only a few distinct classifiers. Again, the findings correspond with results of a subjective listening experiment. This thesis presents more than 70 new features which never have been used before in speaker recognition. A quantitative ranking order of 100 speech features is introduced. Such a ranking order has not been documented elsewhere and is comparatively new to the area of speaker recognition. This ranking order is further extended and describes the amount to which a classifier uses or omits single features, solely depending on the characteristics of the voice sample. Such a separation has not yet been documented and is a novel contribution. The close correspondence of the subjective listening experiment and the findings of the network classifiers show that it is plausible to model the behaviour of human speech recognition with an artificial neural network. Again such a validation is original in the area of speaker recognition

No description available.

No description available.

No description available.

Student Number : 9800233A - MSc dissertation - School of Electrical and Information Engineering - Faculty of Engineering and the Built Environment / Condition monitoring of machinery has increased in importance as more engineering processes are automated and the manpower required to operate and supervise plants is reduced. The monitoring of the condition of machinery can significantly reduce the cost of maintenance. Firstly, it can allow an early detection of potential catastrophic fault, which could be extremely expensive to repair. Secondly, it allows the implementation of conditions based maintenance rather than periodic or failure based maintenance [1]. In these cases, significant savings can be made by delaying schedule maintenance until convenient or necessary. Although there are numerous efficient methods for modeling of mechanical systems, they all suffer the disadvantage that they are only valid for a particular machine. Changes within the design or the operational mode of the machine normally require a manual adaptation. Using Neural Networks to model technical systems eliminates this major disadvantage. The basis for a successful model is an adequate knowledge base on which the network is "trained". Without prior knowledge of the machines systematic behavior or its history, training of a neural Network is not possible. Therefore, it is a pre-requisite that the knowledge base contains a complete behavior of the machine covering the respective operational modes whereby, not all rather the most important modes are required. Neural networks have a proven ability in the area of nonlinear pattern classification. After being trained, they contain expert knowledge and can correctly identify the different causes of bearing vibration. The capacity of artificial neural networks to mimic and automate human expertise is what makes them ideally suited for handling nonlinear systems. Neural networks are able to learn expert knowledge by being trained using a representative set of data [2]-[6]. At the beginning of a neural network’s training session, the neural network fault detector’s diagnosis of the motor’s condition will not be accurate. An error quantity is measured and used to adjust the neural network’s internal parameters in order to produce a more accurate output. This process is repeated until a suitable error is achieved. Once the network is sufficiently trained and the parameters have been saved, the neural network contains all the necessary knowledge to perform the fault detection. One of the most important aspects of achieving good neural network performance has proven to be the proper selection of training features. The curse of dimensionality states that, as a rule of thumb, the required cardinality of the training set for accurate training increases exponentially with the input dimension [7]. Thus feature selection which is a process of identifying those features that contribute most to the discrimination ability of the neural network is required. Proposed methods for selecting an appropriate subset of features are numerous [8]-[11]. Methods based on generating a single solution, such as the popular forward step wise approach, can fail to select features which do poorly alone but offer valuable information together. Approaches that maintain a population of solutions, such as genetic algorithms (GA) are more likely to speedily perform efficient searches in high dimensional spaces, with strong interdependencies among the features. The emphasis in using the genetic algorithm for feature selection is to reduce the computational load on the training system while still allowing near optimal results to be found relatively quickly. To obtain accurate measure of the condition of machinery, a wide range of approaches can be employed to select features indicative of condition. By comparing these features with features for known normal and probable fault conditions, the machine’s condition can be estimated. The most common approach is that of analysis in the frequency domain by applying a Fast Fourier Transform (FFT) to the time domain history data. The idea is simply to measure the energy (mean square value) of the vibrations. As the machine condition deteriorates, this measure is expected to increase. The method is able to reveal the harmonics around the fundamental frequency of the machine and other predominant frequency component (such as the cage frequency) [12]. Frequency analysis is well established and may be used to detect, diagnose and discriminate a variety of induction motor faults such as broken rotor bars, cage faults, phase imbalance, inner and outer race faults. However, as common in the monitoring of any industrial machine, background noise in recorded data can make spectra difficult to interpret. In addition, the accuracy of a spectrum is limited due to energy leakage [12- 14]. Like many of the new techniques now finding application in machinery condition monitoring, Higher Order Statistics was originally confined to the realms of non-linear structural dynamics. It has of recent however found successful application to the identification of abnormal operation of diesel engines and helicopter gearboxes [5, 7]. Higher Order Statistics provide convenient basis for comparison of data between different measurement instances and are sufficiently robust for on-line use. They are fast in computation compared with frequency or time-domain analysis. Furthermore, they give a more robust assessment than lower orders and can be used to calculate higher order spectra. This dissertation reports work which attempts to extend this capability to induction motors. The aim of this project is therefore to examine the use of Genetic Algorithms to select the most significant input features from a large set of possible features in machine condition monitoring contexts. The results show the effectiveness of the selected features from the acquired raw and preprocessed signals in diagnosis of machine condition. This project consists of the following tasks: #1; Using Fast Fourier transform and higher order signals techniques to preprocess data samples. #1; Create an intelligent engine using computational intelligence methods. The aim of this engine will be to recognize faulty bearings and assess the fault severity from sensor data. #1; Train the neural network using a back propagation algorithm. #1; Implement a feature selection algorithm using genetic algorithms to minimize the number of selected features and to maximize the performance of the neural network. #1; Retrain the neural network with the reduced set of features from genetic algorithm and compare the two approaches. #1; Investigate the effect of increasing the number of hidden nodes in the performance of the computational intelligence engine. #1; Evaluate the performance of the system using confusion matrices. The output of the design is the estimate of fault type and its severity, quantified on a scale between 0-3. Where, 0 corresponds to the absence of the specific fault and 3 the presence of a severe machine bearing fault. This research should make contribution to many sectors of industry such as electricity supply companies, and the railroad industry due to their need of techniques that are capable of accurately recognizing the development of a fault condition within a machine system component. Quality control of electric motors is an essential part of the manufacturing process as competition increases, the need for reliable and economical quality control becomes even more pressing. To this effect, this research project will contribute in the area of faults detection in the production line of electric motor.

Knowledge of the mass of the manipulated load (i.e. payload) in off-highway machines is useful information for a variety of reasons ranging from knowledge of machine stability to ensuring compliance with transportion regulations. This knowledge is difficult to ascertain however. This dissertation concerns itself with delineating the motivations for, and difficulties in development of a dynamic payload weighing algorithm. The dissertation will describe how the new type of dynamic payload weighing algorithm was developed and progressively overcame some of these difficulties.<p> The payload mass estimate is dependent upon many different variables within the off-highway vehicle. These variables include static variability such as machining tolerances of the revolute joints in the linkage, mass of the linkage members, etc as well as dynamic variability such as whole-machine accelerations, hydraulic cylinder friction, pin joint friction, etc. Some initial effort was undertaken to understand the static variables in this problem first by studying the effects of machining tolerances on the working linkage kinematics in a four-wheel-drive loader. This effort showed that if the linkage members were machined within the tolerances prescribed by the design of the linkage components, the tolerance stack-up of the machining variability had very little impact on overall linkage kinematics.<p> Once some of the static dependent variables were understood in greater detail significant effort was undertaken to understand and compensate for the dynamic dependent variables of the estimation problem. The first algorithm took a simple approach of using the kinematic linkage model coupled with hydraulic cylinder pressure information to calculate a payload estimate directly. This algorithm did not account for many of the aforementioned dynamic variables (joint friction, machine acceleration, etc) but was computationally expedient. This work however produced payload estimates with error far greater than the 1% full scale value being targeted. Since this initial simplistic effort met with failure, a second algorithm was needed. The second algorithm was developed upon the information known about the limitations of the first algorithm. A suitable method of compensating for the non-linear dependent dynamic variables was needed. To address this dilemma, an artificial neural network approach was taken for the second algorithm. The second algorithms construction was to utilise an artificial neural network to capture the kinematic linkage characteristics and all other dynamic dependent variable behaviour and estimate the payload information based upon the linkage position and hydraulic cylinder pressures. This algorithm was trained using emperically collected data and then subjected to actual use in the field. This experiment showed that that the dynamic complexity of the estimation problem was too large for a small (and computationally feasible) artificial neural network to characterize such that the error estimate was less than the 1% full scale requirement.<p> A third algorithm was required due to the failures of the first two. The third algorithm was constructed to ii take advantage of the kinematic model developed and utilise the artificial neural networks ability to perform nonlinear mapping. As such, the third algorithm developed uses the kinematic model output as an input to the artificial neural network. This change from the second algorithm keeps the network from having to characterize the linkage kinematics and only forces the network to compensate for the dependent dynamic variables excluded by the kinematic linkage model. This algorithm showed significant improvement over the previous two but still did not meet the required 1% full scale requirement. The promise shown by this algorithm however was convincing enough that further effort was spent in trying to refine it to improve the accuracy.<p> The fourth algorithm developed proceeded with improving the third algorithm. This was accomplished by adding additional inputs to the artificial neural network that allowed the network to better compensate for the variables present in the problem. This effort produced an algorithm that, when subjected to actual field use, produced results very near the 1% full scale accuracy requirement. This algorithm could be improved upon slightly with better input data filtering and possibly adding additional network inputs.<p> The final algorithm produced results very near the desired accuracy. This algorithm was also novel in that for this estimation, the artificial neural network was not used soley as the means to characterize the problem for estimation purposes. Instead, much of the responsibility for the mathematical characterization of the problem was placed upon a kinematic linkage model that then fed its own payload estimate into the neural network where the estimate was further refined during network training with calibration data and additional inputs. This method of nonlinear state estimation (i.e. utilising a neural network to compensate for nonlinear effects in conjunction with a first principles model) has not been seen previously in the literature.

Knowledge of the mass of the manipulated load (i.e. payload) in off-highway machines is useful information for a variety of reasons ranging from knowledge of machine stability to ensuring compliance with transportion regulations. This knowledge is difficult to ascertain however. This dissertation concerns itself with delineating the motivations for, and difficulties in development of a dynamic payload weighing algorithm. The dissertation will describe how the new type of dynamic payload weighing algorithm was developed and progressively overcame some of these difficulties.<p> The payload mass estimate is dependent upon many different variables within the off-highway vehicle. These variables include static variability such as machining tolerances of the revolute joints in the linkage, mass of the linkage members, etc as well as dynamic variability such as whole-machine accelerations, hydraulic cylinder friction, pin joint friction, etc. Some initial effort was undertaken to understand the static variables in this problem first by studying the effects of machining tolerances on the working linkage kinematics in a four-wheel-drive loader. This effort showed that if the linkage members were machined within the tolerances prescribed by the design of the linkage components, the tolerance stack-up of the machining variability had very little impact on overall linkage kinematics.<p> Once some of the static dependent variables were understood in greater detail significant effort was undertaken to understand and compensate for the dynamic dependent variables of the estimation problem. The first algorithm took a simple approach of using the kinematic linkage model coupled with hydraulic cylinder pressure information to calculate a payload estimate directly. This algorithm did not account for many of the aforementioned dynamic variables (joint friction, machine acceleration, etc) but was computationally expedient. This work however produced payload estimates with error far greater than the 1% full scale value being targeted. Since this initial simplistic effort met with failure, a second algorithm was needed. The second algorithm was developed upon the information known about the limitations of the first algorithm. A suitable method of compensating for the non-linear dependent dynamic variables was needed. To address this dilemma, an artificial neural network approach was taken for the second algorithm. The second algorithms construction was to utilise an artificial neural network to capture the kinematic linkage characteristics and all other dynamic dependent variable behaviour and estimate the payload information based upon the linkage position and hydraulic cylinder pressures. This algorithm was trained using emperically collected data and then subjected to actual use in the field. This experiment showed that that the dynamic complexity of the estimation problem was too large for a small (and computationally feasible) artificial neural network to characterize such that the error estimate was less than the 1% full scale requirement.<p> A third algorithm was required due to the failures of the first two. The third algorithm was constructed to ii take advantage of the kinematic model developed and utilise the artificial neural networks ability to perform nonlinear mapping. As such, the third algorithm developed uses the kinematic model output as an input to the artificial neural network. This change from the second algorithm keeps the network from having to characterize the linkage kinematics and only forces the network to compensate for the dependent dynamic variables excluded by the kinematic linkage model. This algorithm showed significant improvement over the previous two but still did not meet the required 1% full scale requirement. The promise shown by this algorithm however was convincing enough that further effort was spent in trying to refine it to improve the accuracy.<p> The fourth algorithm developed proceeded with improving the third algorithm. This was accomplished by adding additional inputs to the artificial neural network that allowed the network to better compensate for the variables present in the problem. This effort produced an algorithm that, when subjected to actual field use, produced results very near the 1% full scale accuracy requirement. This algorithm could be improved upon slightly with better input data filtering and possibly adding additional network inputs.<p> The final algorithm produced results very near the desired accuracy. This algorithm was also novel in that for this estimation, the artificial neural network was not used soley as the means to characterize the problem for estimation purposes. Instead, much of the responsibility for the mathematical characterization of the problem was placed upon a kinematic linkage model that then fed its own payload estimate into the neural network where the estimate was further refined during network training with calibration data and additional inputs. This method of nonlinear state estimation (i.e. utilising a neural network to compensate for nonlinear effects in conjunction with a first principles model) has not been seen previously in the literature.

Taiwan is often threaten by typhoon during summer and autumn. The surges brought by theses typhoons not only cause human lives in danger, but also cause severe floods in coastal area. Storm surge prediction remains still a complex coastal engineering problem to solve since lots of parameters may affect the predictions. The purpose of this study is to predict storm surges using an Artificial Neural Network (ANN). A non-linear hidden-layer forward feeding neural network using back-propagation learning algorithms was developed. The study included a detailed analysis the factors may affect the predictions. The factors were obtained from the formulation of storm surge discrepancies after Horikawa (1987). Storm surge behaviors may vary from different geographical locations and weather conditions. A correlation analysis of the parameters was carried out first to pick up those factors shown high correlations as input parameters for establishing the typhoon surge predictions. The applications started with collecting tide and meteorological data (wind speed, wind direction and pressure) of Dapeng Bay and Kaohsiung harbor. A harmonic analysis was utilized to identify surge deviations. The surge deviation recorded at Dapeng Bay was found higher then Kaohsiung harbor for the same typhoon events. Correlation analysis has shown positive correlations between wind field, both wind speed and direction, and the associated storm surge deviations at Dapeng Bay. Correlation coefficients (CC) 0.6702 and 0.58 were found respectively. The variation of atmospheric pressure during typhoons is found with positive correlation too (i.e. CC=0.3626). Whereas the analysis has shown that the surges at Kaohsiung harbor were only sensitive to wind speed (CC=0.3723), while the correlation coefficients of the wind direction (CC=-0.1559) and atmospheric pressure (CC= -0.0337) are low. The wind direction, wind speed and atmospheric pressure variation were then used as input parameters for the training and predictions. An optimum network structure was defined using the Dapeng Bay data. The best results were obtained by using wind speed, wind direction and pressure variation as input parameters. The ANN model can predict the surge deviation better if the empirical mode decomposition (EMD) method was used for training.