10. Electrical Contact Failure Detection Based on Dynamic Resistance Principle Component Analysis and rbf Neural Network

Circuit Breakers (CBs) are extremely important components for power system reliability and protection. These critical components need a comprehensive maintenance plan to avoid failure and unplanned power outage [1]. Among the popular maintenance policies, the condition based maintenance (CBM) has been reported as the most efficient one for power system CBs, because of economic achievements and its availability in the smart environment [2-3]. Condition monitoring systems provides the information needed for CBM. The interpretation of these data are greatly depended on operator's knowledge which could result in improper description of CB condition. The monitoring of CBs is usually split into the monitoring of its separate components, i.e. gas mixture, mechanical parts, switching, contacts and control circuit [4]. Defects on electric contacts are responsible for many CB's failures. The design of modern high voltage puffer type SF6 gas CBs is based on switching of two parallel contact sets. The main contacts are designed to carry the load current without any excessive temperature rise, thus they do not degrade by the effects of the arcing. Following the main contacts part, arc contacts are finally opened, thus initiating arc quenching and current interruption. [5-7]. The energy released by the electric arcing has the ability to induce rapid material erosion at the arcing contacts and yields to change in contact geometry and causes a significant decrease in the breaking capacity and working condition of the CB and associated electrical circuit [6-9]. In addition to contact erosion, the other contact's prevalent defect is contact misalignment. Due to misalignment of arcing contact at the break assembly, damage to fixed arcing contact may be happened. This condition changes the contact geometry as well and yields to imperfect performance of CB.

A useful parameter to effectively diagnose the electrical contact failures is dynamic resistance measurement (DRM). This method is used to evaluate the contact condition without dismantling the CB. The original approach for dynamic contact resistance is to measure the resistance during an open operation of the breaker. This method has been deve loped in recent years. For instance, Ref [7] reports when the DRM test is carried out in rated contact speed, several spikes can be observed in arcing contact part which causes the DRM curve in two consecutive tests have some differences. So, it suggests that DRM should be done at low speed (0.002-0.2 m1s) in case of reproducible resistance curve to be obtained, these curves are easy to analyze and interpret. But authors of Ref [10] believe that this method is instructive for some breaker mechanisms. Since an adjustment to the operating mechanism is required, there is a potential risk of damaging the operating mechanism when restoring it back in service. Ref. [5] reports the presence of metallic fluorides deposited on breaker contacts. These nonconductive powders do not decrease the breaker's breaking capacity but cause an increase in contact resistance and thus may lead to wrong interpretation of contact condition. So, for this kind of breakers, it is proposed that DRM should carried on 2800 A DC. Large number of works done till now have been tried to modify the DRM test and none of them contributed to the comprehensive investigations on which correlates the data extracted from dynamic resistance curve to the type of contact failure.

The study presented here aims to develop a smart failure detection technique which is easy to understand and implement. DRM data are gathered from three similar electrical contacts. One of these contacts is in good condition, while two other contacts are exposed to intentional malfunctions, i.e. contact misalignment, and contact erosion. Using principal component analysis (PCA), the most important eigenvectors of DRM are selected. Having these selected eigenvectors and use of RBF neural network, an intelligent way to identity contact condition is suggested. This is to prepare a classifying method for distinguishing normal versus different faulty conditions.

A. Dynamic contact resistance measurement

To effectively diagnose the condition of main and arcing contacts DRM is carried on. For DRM two signals must be recorded: the injected current and the voltage drop across the breaker contacts. The experimental apparatus is designed to controllably open a pair of contacts. DRM is measured during close-open operation. During an opening operation, the main contacts separate first. The current then transferred to the arcing contacts which separate a few milliseconds later. The arc, which is formed after the separation of the arcing contacts, is cleared at the next current zero-crossing [11].

The test current is delivered by a regular 12V/220 Ah car battery. The current passing the contacts is measured via a 1 mn non-inductive shunt resistance. All signals are recorded with a digital storage oscilloscope with 50 GHz bandwidth. The contacts open at a constant velocity of 0.2 mis, controlled by the opening control mechanism to keep the system variables to a minimum. In this way, resistance curves are smoother, and reproducible from one test to another. Moreover, main contact part can be easily identified at low contact speed [7]. By injecting a high current through the contacts and simultaneous monitoring the voltage drop across the main contact during the opening operation, and by dividing instantaneous voltage to current, DRM curve will be acquired [6]. The DRM results are usually presented in contact resistance (ohm) versus displacement of moving contact (mm) or time of operation (s) [12, 13]. In order to investigate the impact of different failures on DRM characteristics, measurements are gathered from three similar contacts: Healthy, eroded and deformed ones.

In all these cases, there is an increased risk for excess heat within the CB, with the possibility of reaching higher temperatures than the CB is designed for, and thereby the probability of explosion of the breaker [6].

Realizing that the characteristic parameters of DRM have a statistical nature, a large number of measurements are to be done. For this purpose, DRM is carried out 100 times during each healthy and faulty operations. DRM data are saved in database. Using PCA method the most important parameters of waveform are extracted. These data are considered as inputs for RBF neural network to determine the class of contact as healthy, eroded and deformed.

B. Principal component analysis (PCA)

Principal component analysis is a mainstay technique for extracting structure from possibly high dimension datasets. PCA is an orthogonal transformation of the coordinate system in which the data are described. The new coordinate values, by which the data are represented, are called principal components.

It is often the case that a small number of principal components are sufficient to account for the main structure embedded in data. Listed below are the general steps for performing a principal component analysis:

• Standardize the data.

• Obtain the eigenvectors and eigenvalues from the covariance matrix or correlation matrix.

• Sort eigenvalues in descending order and choose the k eigenvectors that correspond to the k largest eigenvalues where k is the number of dimensions of the new feature subspace.

• Construct the projection matrix W from the selected k eigenvectors.

• Transform the original dataset X via W to obtain a k­ dimensional feature subspace Y [14, 15].

There are different methods for determining number of factors (k) in order to not lose useful data. For example the Kaiser rule is to retain components with eigenvalues more than 1.0. This rule assumes that any component with eigenvalue less than 1.0 contains less information than one of the original variables and so is not worth retaining. However, in practice smaller values like 0.7 are selected to consider less informative data [16].

Dynamic resistance waveform is a time-dependence signal which contains useful information about contact condition. Parameters like main and arcing contacts resistance values, overlap time, etc. can reveals contact abnormalities, but are not capable to find type of fault. Using all data hidden in dynamic resistance waveform is not also the case, because of time consumption, noise problem and memory issues. So, it is reasonable to extract the most useful data of a given DRM signal and to compare these characteristics to find contact condition.

Radial basis function neural network (RBFNN) (Радиальная базисная функция нейронной сети) has a feed­ forward structure. It typically has three layers of nodes: the input layer, the hidden layer with a non-linear RBF activation function and the output layer. The radial basis function is commonly taken to be Gaussian. RBF networks are typically trained by a two-step algorithm. In the first step, the center vectors C, of the RBF functions in the hidden layer are chosen. Centers can be randomly sampled from some set of examples, or they can be determined using k-means clustering. The second step simply fits a linear model with coefficients Wi to the hidden layers' outputs with respect to some objective function. A common objective function is the mean square error (MSE) [17, 18].

In this study after performing PCA on DRM signal, eigenvalues with variance more than 0.9 are chosen as inputs for RBFNN. So, an input matrix would have a size of 10x300. Which 10 returns to PCA output result and 300 is number of tests have been done on contact samples.

Since input has three classes, the output matrix would be a 1 x300 matrix. In which first 100 elements returns to healthy contact (number 1), second 100 elements are corresponded to contact with erosion (number 2) and third 100 elements are responsible for deformed contact (number 3). Mean square error is selected to evaluate the efficiency of RBFNN in classifying given data.

Techniques which are used to assess the condition of an equipment should be easy to measure, interpret and understand. DRM is an effective method to evaluate contact condition. In order to eliminate the difficulties of signal interpretation and develop an intelligent approach to precisely diagnose the contact condition, a new method based on principal component analysis and RBFNN is suggested.

Required data are gathered through DRM tests on three similar contacts: healthy, eroded and deformed ones. The effectiveness, feasibility and applicability of the proposed method were verified through case studies. As the mean square error which was chosen to evaluate the efficiency of RBFNN receive a number of 2.0753e-29. This approach can also be used to identify any other kind of contact failures. Use of this technique would decrease number of unnecessary maintenance activities and then reduce the power system costs.