Fuzzy learning vector quantization networks for power transformer condition assessment

Prediction of laser butt joint welding parameters using back propagation and learning vector quantization networks

Considered as the heart of electrical power transmission and distribution networks, power transformers are essential part of the electricity transmission grid. Among the condition monitoring and fault diagnosis tools for these machines, dissolved gas analysis (DGA) has proven its effectiveness in their early detection and classification of faults. Up to date, many methods have been proposed in the literature for the interpretation of DGA data, classified into traditional and intelligent methods. This paper proposes a two-steps hybrid method, which uses the strengths of both methods. The approach uses the evolutionary k-means clustering algorithm based on the genetic algorithm for subset formation and subset analysis by human expertise. In the diagnostic procedure, to determine the condition of a sample, the subset to which it belongs is first identified and then the corresponding diagnostic sub-model is applied. The proposed method has been implemented with 595 DGA data, tested on 254 DGA data and validated on the International Electrotechnical Commission (IEC) TC10 database. Their performances were evaluated and compared with existing traditional, intelligent and hybrid methods. From the results obtained with the IEC TC10 database, the newly proposed approach depicts the best overall diagnosis accuracies. Indeed, the best performance is achieved with the proposed method compared to other models in the literature, with diagnostic accuracy of 98.29% compared to 88.89% of the Gouda triangle method, to 88.03% of the Hyosun Corporation gas ratio method or to 86.32% of the three ratios technique.

Unnatural control chart patterns (CCPs) are associated with a particular set of assignable causes for process variation. Hence, effectively recognizing CCPs can substantially narrow down the set of possible causes to be examined, and accelerate the diagnostic search. Recently, machine-learning techniques, especially the artificial neural network (ANN), have been widely used as an effective tool for CCP recognition (CCPR) tasks. Most ANN applications in CCPR have been using static supervised ANNs, such as back propagation networks (BPNs) and learning vector quantization (LVQ) networks. The false recognition problem (i.e. the patterns are misclassified) commonly encountered for these ANN-based CCPR models is mainly due to the fact that the static ANNs cannot appropriately deal with dynamic patterns, such as CCPs. In this research, a dynamic training algorithm is designed to provide an LVQ network-based CCPR model the capability to on-line recognize the dynamic CCPs that vary over time. The numerical results using simulation show that the dynamically trained LVQ network-based model proposed in this research performs much better than other ANN-based models reported in literature with respective to recognition accuracy and speed. Although this research considers the specific application of a real-time CCPR model based on an LVQ network, the proposed dynamic training algorithm could be applied to CCPR systems based on other ANN architectures in general.

Stereology and stochastic geometry can be used as auxiliary tools for diagnostic purposes in tumour pathology. Whether first-order parameters or stochastic-geometric functions are more important for the classification of the texture of biological tissues is not known. In the present study, volume and surface area per unit reference volume, the pair correlation function and the centred quadratic contact density function of epithelium were estimated in three case series of benign and malignant lesions of glandular tissues. The information provided by the latter functions was summarized by the total absolute areas between the estimated curves and their horizontal reference lines. These areas are considered as indicators of deviation of the tissue texture from a completely uncorrelated volume process and from the Boolean model with convex grains, respectively. We used both areas and the first-order parameters for the classification of cases using artificial neural networks (ANNs). Learning vector quantization and multilayer feedforward networks with backpropagation were applied as neural paradigms. Applications included distinction between mastopathy and mammary cancer (40 cases), between benign prostatic hyperplasia and prostatic cancer (70 cases) and between chronic pancreatitis and pancreatic cancer (60 cases). The same data sets were also classified with linear discriminant analysis. The stereological estimates in combination with ANNs or discriminant analysis provided high accuracy in the classification of individual cases. The question of which category of estimator is the most informative cannot be answered globally, but must be explored empirically for each specific data set. Using learning vector quantization, better results could often be obtained than by multilayer feedforward networks with backpropagation.

The history of patient classification in nursing dates back to the period of Florence Nightingale. The first and the foremost condition for providing quality nursing care, which is measured by care standards, and determined by number of hours of actual care, is the appropriate number of nurses. Patient classification criteria are discussed in this paper. Hybrid classification model based on learning vector quantization (LVQ) networks and self-organising maps (SOM) are purposed. It is possible to discus three types of experimental results. First result could be assessment of Braden scale and Mors scale by LVQ. Second result, the time for nursing logistics activities. The third is possibility to predict appropriate number of nurses for providing quality nursing care. This research was conducted on patients from Institute of Neurology, Clinical Centre of Vojvodina.

This paper presents a new hybrid intelligent algorithm for assessing and enhancing voltage security. The problem is decomposed into two stages. Firstly, the security of current operating point is assessed by learning vector quantization (LVQ) network according to the pre-specified criterion. Throughout the paper, continuation power flow (CPF) is adopted as a tool for defining voltage security margin (VSM). If an insecure state is indicated by the LVQ, the power system operator is given the suggestion for appropriate control settings so as to maintain the specified security level. In the second stage, the optimal reactive power dispatch (ORPD) problem is formulated and VSM is considered as an additional constraint. To include VSM determined by CPF along the optimization process, feed-forward neural network (FFNN) is trained to learn and perform similarly to CPF to relieve the intensive computing requirement. Ant colony optimization (ACO) is applied to handle the VSM constrained ORPD problem. The proposed method was tested on IEEE 30-bus system and successful results were obtained.

The learning vector quantization (LVQ) network was used to classify the ECG ST segment into different morphological categories. Due to the lack of data in the ST elevation categories, the classifier was only trained to identify different types of ST depressions (horizontal, upsloping and downsloping). The accuracies were 91%, 85% and 65% respectively for the training, validation and testing data respectively. Despite the low accuracy for the testing data, most of the mis-classifications were downsloping ST depression being classified as horizontal ST depression. We concluded that more data and more training are needed in order to train the LVQ to recognize other morphological types of ST deviation and to improve the accuracy.

The use of neural networks to create adaptive models of users and communications channels for use in designing system response characteristics is discussed. Two types of neural networks that will be useful for this type of task are considered. One, the Kohonen learning vector quantization (LVQ) network, is a clustering network. It can adjust the vector element values of a set of quantizing vectors in order to create exemplar vectors which represent clusters in a set of data. The other, the Kohonen self-organizing topology-preserving map (SOTPM), is a more advanced and powerful network that uses similar principles as the LVQ. It can create a topographic mapping of a set of vector data which creates a data clustering visible in a reduced dimensionality space from the original data. This will facilitate interpretation of the data clusters describing different user types. The use of a rather highly evolved form of neural networks to create more powerful system models is discussed.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

The power transformer is an important equipment in the power system. Its running state is directly related to the safe and stable operation of the power grid. The volume fraction of the dissolved gas in the oil of the transformer body and its variation law are closely related to the fault mode of the transformer, therefore, the dissolved gas analysis (DGA) technology in transformer oil is widely used. Actually, the relationship between the volume fraction of the dissolved gas in transformer oil and the time is a multi-dimensional time series. The sequences are arranged in the same time interval or in different time intervals, which contains the external environment of the power transformer, the operation conditions and the inherent relationship between the gas content in the transformer oil. Therefore, the operation status of the power transformer can be revealed by the prediction of the time series. Based on this, the fault type of the transformer can be determined. At present, many literatures have been applied to a single time series forecasting method, such as BP neural network and radial basis function neural network. The single prediction method can obtained certain accuracy by adjusting the weights and thresholds of the network. To further improve the prediction accuracy, a combined forecasting model should be proposed. The research result of this paper is to predict the DGA data of transformer, and has a certain guiding significance to determine the transformer fault types. The proposed method for time series prediction can also be used in other time series prediction of the electric power system.

Multi-filter semi-supervised transformer model for fault diagnosis

Power transformer is one of the most important and expensive equipment in a power system. Its reliability directly affects a power system. To ensure the reliable operation of a power transformer, its condition needs to be continuously monitored and evaluated. Over the past two decades, a number of diagnostic techniques have been developed for transformer condition assessment such as dissolved gas analysis (DGA), degree of polymerization (DP) measurement, polarization and depolarization current (PDC) measurement, frequency domain spectroscopy (FDS), frequency response analysis (FRA), and partial discharge (PD) detection. However, the interpretations of measurement results acquired from these diagnostics are usually based upon the empirical models, which are sometimes inaccurate and incomplete especially in abnormal transformer operation scenarios. Therefore, accurate interpreting on the measurement data obtained by the above techniques and subsequently making explicit condition assessment of transformers is still a challenge task. Nowadays, considerable efforts have been made in the field of transformer condition monitoring and assessment. Majority efforts are dedicated in developing accurate transformer models and reliable transformer fault diagnosis systems. After completing a comprehensive literature review on various diagnostic techniques for transformers condition assessment, this thesis focuses on three main aspects of transformer’s health condition, including oil characteristics and dissolved gases in transformers, moisture concentration of oil-cellulose insulation and hot spot temperature of transformer windings. Since there is a lack of common framework for applying pattern recognition algorithms (i.e. data centric approaches) to interpret oil characteristics and DGA data, this thesis firstly provides a critical review on various pattern recognition techniques for power transformer insulation diagnosis using DGA and oil characteristics datasets. A general pattern recognition application framework is then proposed. The important issues for improving the applicability of pattern recognition techniques for transformer insulation diagnosis are also discussed. To improve the data quality of training database and enhance the diagnostic accuracy of pattern recognition algorithms, a hybrid algorithm, SMOTEBoost is proposed. It adopts Synthetic Minority Over-sampling Technique (SMOTE) to handle the class imbalance problem, where samples belonging to different fault types (insulation conditions) are unevenly distributed in the training database. By using the boosting approach for reweighting and grouping data points in the training database, the SMOTEBoost can facilitate pattern recognition algorithms consistently attaining desirable diagnosis accuracies. To solve the intricate difficulties in moisture estimation of transformer oil-cellulose insulation system, this thesis introduces two modelling approaches, i.e., multi-physics finite element modelling (FEM) and particle tracing method, where the temperature dependent moisture dynamics in transformers’ insulation system is taken into account. In multi-physics approach, moisture dynamics is modelled by coupling the electromagnetic, thermal, fluid flow and moisture migration physics simultaneously. In particle tracing method, moisture diffusion is formulated from a microscopic view of water molecules’ motion. Especially, the transmission probability of water molecules (termed as particles in the paper) is employed to correlate the microscopic particles’ motion with the macroscopic moisture distribution. Extensive ageing and moisture diffusion experiments have been conducted on a prototype transformer to verify the proposed modelling approaches for an accurate estimation of moisture in transformers. This thesis also proposes a distributed parameter model to investigate the effect of moisture dynamics on dielectric response of a transformer’s cellulose insulation. The correlation between moisture distribution (under non-equilibrium conditions due to thermal transients) and dielectric response parameters (dielectric losses and permittivity) of transformer cellulose insulation is revealed. The proposed methodology can help the proper interpretation of dielectric response measurement of field transformers under thermal transients. To overcome the inaccuracy in empirical thermal dynamic models, in this thesis a moisture dependent thermal model (MDTM) is developed for estimating transformer’s hot spot temperature. In this model, nonlinear thermal resistance is formulated by considering both oil and cellulose (paper and pressboard) of the transformer. Especially, the effect of moisture concentration and hot spot temperature on the thermal resistance of cellulose is taken into account. The proposed MDTM is verified by using historical data of moisture-in-oil and temperature measurements on an in-service vegetable oil-filled transformer. To integrate every piece of data and information obtained from different transformer diagnostic measurements and subsequently evaluating the overall health condition of a transformer, this thesis proposes a data and information fusion framework based on Bayesian Network (BN). Within the Bayesian Network, Monte Carlo and Bootstrap methods are employed to extract the most informative characteristics regarding transformer condition from different diagnostic measurements. Results of case studies demonstrate that the proposed data and information fusion framework can evaluate the effectiveness of combinations of different diagnostic measurements and subsequently facilitate determining optimal diagnostic strategies involved in transformer condition assessment. It is expected that the data centric diagnostic approaches developed in this thesis can provide an accurate modelling and reliable assessment of transformer’s health condition.

A distributed diagnosis system combining the immune network (IN) and learning vector quantization (LVQ) is proposed for accurately detecting faulty sensor outputs in control plants. The system has two execution modes, namely, its training mode, where the LVQ extracts a correlation between each two sensors from their outputs when they work properly, and its diagnosis mode, where the LVQ contributes to testing each two sensors using the extracted correlation, and the IN contributes to determining faulty sensors by integrating the local testing results obtained from the LVQ. With the proposed method, faulty sensors, such as age deteriorated ones, which have been difficult to be detected only by checking each sensor output independently, can be specified.

An improved Back Propagation (BP) artificial neural network is utilized to assess the insulation condition of a large oil immersed electric power transformer in this paper. After a complete comparison of performances between a few different network architectures, a new kind of BP network structure with a promoted learning algorithm is chosen to train the diagnostic network. Furthermore, some techniques in the reliability analysis of data is introduced into the BP network so as to realize pre-treatment of the data acquired through Dissolved Gas Analysis (DGA), as it is a useful tool for assessing oil-paper insulation. It is verified by the DGA data from substations, that the improved BP algorithm bound with the technique of data pre-treating obtained much higher accuracy. So, it is worthy of being applied for insulation diagnosis in utilities.

To allow utilities to fulfill self-imposed and regulative performance targets the demand for new optimized tools and techniques to Estimate the performance of modern Transformers has increased. The modern power transformers has subjected to different types of faults, which affect the continuity of power supply which in turn causes serious economic losses. To avoid the interruption of power supply, various fault diagnosis approaches are adopted to detect faults in the power transformer and has to eliminate the impacts of the faults at the initial stage. Among the fault diagnosis methods, the hybrid technique of Particle Swarm Optimization (PSO) with Support Vector Machine (SVM) learning algorithm is simple conceptually derived and its implementation process is faster with better scaling properties for complex problems with non linearity and load variations but performance factor related to accuracy has a declined value in case of correlations implicit . In order to obtain better fault diagnosis to improve the service of the power transformer, SVM is optimized with Improved PSO technique to achieve high interpretation accuracy for Dissolved Gas Analysis (DGA) of power transformer through the extracting positive features from both the techniques. Primary SVM is applied to establish classification features for faults in the transformer through DGA. The features are applied as input data to Autonomous optimized Technique for faults analysis. The proposed methodology obtains the DGA data set from diagnostic gas in oil of 500 KV main transformers of Pingguo Substation in South China Electric Power Company. The simulations are carried out in MATLAB software with an Intel core 3 processor with speed of 3 GHZ and 4 GB RAM PC. The result obtained by Autonomous optimized Technique (IPSO-SVM) is compared against PSO-SVM to estimate the performance of the classifiers in terms of execution time and quality of classification for precision. The test results indicate that the Autonomous optimization of IPSO-SVM approach has significantly improved the classification accuracy and computational time for power transformer fault classification. Keywords: Transformer Fault Analysis, Improved Particle Swarm Optimization, Hybrid Optimization, Dissolved Gas Analysis, Support Vector Machine

The concept of Concealed Weapon and Explosive (CWE) detection by the analysis of the Late Time Response (LTR) of the complex human-CWE object in UWB Radar, has been presented in [1,2]. As the overall reflected human signal depends on the human stance and orientation with respect to the radar system, this paper investigates whether the resonant frequencies can be classified according to the illuminated simple i.e. human or complex i.e. human-CWE object. This classification yields that the human frequencies do not overlap with the CWE signature frequencies therefore the CWE frequencies can be obtained and the body-worn CWE detection is realised. The resonant frequency classification is achieved via a Learning Vector Quantization (LVQ) network.