Journal of Control Science and Engineering

Volume 2018, Article ID 6565737, 10 pages

https://doi.org/10.1155/2018/6565737

## Data Preprocessing Method and Fault Diagnosis Based on Evaluation Function of Information Contribution Degree

School of Automation, Hangzhou Dianzi University, Hangzhou 310018, China

Correspondence should be addressed to Chenglin Wen; nc.ude.udh@lcnew

Received 18 February 2018; Revised 14 April 2018; Accepted 30 April 2018; Published 2 July 2018

Academic Editor: Youqing Wang

Copyright © 2018 Siyu Ji and Chenglin Wen. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

#### Abstract

Neural network is a data-driven algorithm; the process established by the network model requires a large amount of training data, resulting in a significant amount of time spent in parameter training of the model. However, the system modal update occurs from time to time. Prediction using the original model parameters will cause the output of the model to deviate greatly from the true value. Traditional methods such as gradient descent and least squares methods are all centralized, making it difficult to adaptively update model parameters according to system changes. Firstly, in order to adaptively update the network parameters, this paper introduces the evaluation function and gives a new method to evaluate the parameters of the function. The new method without changing other parameters of the model updates some parameters in the model in real time to ensure the accuracy of the model. Then, based on the evaluation function, the Mean Impact Value (MIV) algorithm is used to calculate the weight of the feature, and the weighted data is brought into the established fault diagnosis model for fault diagnosis. Finally, the validity of this algorithm is verified by the example of UCI-Combined Cycle Power Plant (UCI-ccpp) simulation of standard data set.

#### 1. Introduction

In this paper, based on the evaluation function to calculate the weight of the data feature, the premise is to evaluate the accuracy of the function, which is the theoretical basis and guarantee for the validity of the follow-up work. At present, BP neural network is one of the most popular regression algorithms. It can effectively approximate complex nonlinear mappings based on inputting samples and has the advantages of simple structure and high operability and has been applied in many fields [1, 2]. On the other hand, there are some problems [3]. First of all, the parameter training method based on the gradient descent learning algorithm converges slowly and falls into the local optimum. Secondly, there are many parameters in the neural network that need to be trained, which can take a great time to deal with. In the actual operation of the system, the model data of the system may not have been completely collected but acquired one by one or block by block, while the neural network is a learning algorithm lacking the ability to update online. It is unacceptable for many practical situations to retrain the network in response to changes in the modalities of system. However, if the parameters in the neural network are not updated in time, the fitting output of the network will greatly deviate from the real value.

In order to solve the above problems, the industry has proposed Neural Network Incremental Learning algorithm to deal with them [4]. The incremental learning method can adjust the artificial neural network by analyzing the specific conditions and recognition results of each new sample, learn new knowledge based on the existing knowledge, and flexibly adapt to the dynamic changes of the environment [5]. Therefore, there are many researches on incremental learning algorithms [6, 7]. The main idea of incremental learning is mainly reflected in two aspects: (1) in the actual perception of data, the amount of data is often gradually increased. Therefore, in the face of new data, the learning method should only update the changes caused by the new data without changing the original knowledge base and then learn from the new data contained in the knowledge; (2) the cost of modifying a well-trained system is usually less than the cost of retraining one system. There are many frameworks for incremental learning. The core of each framework is to evaluate the similarity between new data and stored knowledge. The method thus determines the way in which new knowledge is perceived and the knowledge base is increased, which affects the growth of knowledge. Therefore, the judgment mechanism of new knowledge is the core part of incremental learning. In an Orthogonal Least Square (OLS) learning algorithm proposed by Chen, structure and parameter identification are performed simultaneously [8]. In [9], a dynamic fuzzy neural network (DFNN) based on RBF is proposed. The parameters are adjusted by Linear Least Square (LLS), and the structure can be adaptively added and deleted according to the rules. On the basis of these, a generalized DFNN (GDFNN) is proposed by the literature [10]. Based on the Ellipse Basis Function (EBF), an online parameter locating mechanism is proposed. The above algorithms are more difficult to achieve and at the same time cannot guarantee the real-time algorithm.

The Kalman filtering method was widely used in the fields of process control, communication, biomedical science, etc., once proposed in the 1960s. Because of its recursive nature, it does not need to store a large amount of historical information and reduces the amount of computer storage. Combining the system state equation and the observation equation directly, it can directly give the estimation accuracy when estimating system state parameters. Its concise way of thinking had become the theoretical basis for the development of such theories as estimation theory and emerging information fusion [11, 12]. Compared with Kalman, the least squares method uses all the observed data to estimate the value of the state quantity at the initial time. Due to the large number of observations and the statistical characteristics of the method, then this method has a higher accuracy. However, due to its centralized nature, it lacks real-time performance. Kalman filter after the observation data is updated, the state variables are improved with new observation data to obtain the state variables at this observation time. Therefore, Kalman filtering is suitable for real-time processing. In this paper, a method of real-time update of hidden layer output weights based on Kalman filter is proposed, which avoids the retraining of the model. At the same time, the deviation from the real value of the model output data caused by the failure of the model parameters to be updated due to the system modal change is eliminated.

The rest of this article is organized as follows: Section 2 gives a brief description of the problems to be solved. The formal description of BP neural network, the gradient descent method, least squares method, and Kalman filter method are used to update the global or local parameters of the network in Section 3. In Section 4, we give a detailed introduction to the process of weighting MIV algorithm. Section 5 simulates the proposed algorithm based on the UCI standard dataset and results of comparisons and analyses. Section 6 summarizes the related research contents, elaborates on the existing problems, and gives a prospect of the next work.

#### 2. Problem Description

Based on the data-driven fault diagnosis method, a series of data processing is often required before the fault diagnosis, such as data standardization, data dimension reduction, feature selection, feature weighting, etc. Some even need to map the data onto high-dimensional space, like we are using support vector machine (SVM) for data classification. The ultimate goal of all the above data preprocessing operations is to improve the diagnostic performance of the fault diagnosis method. Different data preprocessing operations are often adopted for different fault diagnosis methods. Feature weighting algorithm is relatively special, and no matter which kind of data preprocessing operation, you can then weight their features. The purpose of the weight of the feature is to amplify the difference between the feature variables of the data and to eliminate the feature redundancy to a certain extent.

is the data set composed of samples and is a sample of the data set , where is the sample at time and is the component at time , . Our purpose is to find a transformation matrix to perform a weighted transformation on the sample dataMaking the weighted transformed sample data can be more effective in fault diagnosis. The next question is how to model and find the transformation matrix : the traditional principal component analysis (PCA) algorithm can solve the transformation matrix to map the original data containing features onto the selected m-dimensional feature space and have , so PCA lost some information in the process of data transformation. The information entropy feature weighting algorithm weights the sample data based on the uncertainty of the feature set’s classification of the data set, but its calculation is more difficult; the data set also has higher requirements. In this paper, we want to obtain the weight of the feature through an evaluation function. The evaluation function is generated based on the BP neural network fitting function. In order to ensure the accuracy of the evaluation function, we generally need to retrain the network when the system modal changes. This is a time-consuming task. In order to avoid the retraining of the model, this paper proposes a Kalman filter-based algorithm for real-time update of hidden layer output weights of BP neural network and then establishes the evaluation function model to obtain the transformation matrix (i.e., feature weights). Following we are going to introduce the establishment of the evaluation function and the weight of the data feature in detail. The algorithm implementation process is shown in Figure 1.