Weighted networks can give more detailed description of interaction between agents of corresponding systems. Link weight also provides another way to improve the properties and functions of networks. Based on the concept of network efficiency in binary networks, in this paper, the efficiency of weighted networks with similarity or dissimilarity weight is defined. The effect of weight distribution on the network efficiency are investigated. From the initial regular network with homogeneous link weights, a method is introduced to randomize the weight distribution over the links. The results demonstrate that the random redistribution of link weight can improve the network efficiency. Moreover, exponential distribution of link weight shows more significant improvement compared with the other common distributions, such as uniform, Poisson, Gauss, and power law distributions. Meanwhile, it is also found that the total weight of the corresponding minimum spanning tree is reduced with the randomization of link weight. That means the cost of transportation is decreased with the increase of link weight heterogeneity. All these results can help us get deeper understanding about the effect of link weight on the property and function of networks.

Support Vector Machines (SVMs) are being used extensively now days in the arena of pattern recognition and regression analysis. It has become a good choice for machine learning both for supervised and unsupervised learning purposes. The SVM is primarily based on the mapping the data to a hyperplane using some kernel function and then increasing the margin between the hype planes so this hyperplane classifies the data in the normal and fault state. Due to large amount of input data, it is computationally cumbersome to yield the desired results in shortest possible time by using SVM. To overcome this difficulty in this work, we have employed statistical Time-Domain Features like Root Mean Square (RMS), Variance, Skewness and Kurtosis as pre-processors to the input raw data. Then various combinations of these time-domains signals and features have been used as inputs and their effects on the optimal model selection have been investigated thoroughly and optimal one has been suggested. The procedure presented here is computational less expensive otherwise to process the input data for model selection we may have to use super computer. The implementation of proposed method for machine learning is not much complicated and by using this procedure, an impending fault/abnormal behavior of the machine can be detected beforehand.