TMCnet News
Study on Multi-Equipment Failure Prediction Based on System Network [Sensors & Transducers (Canada)](Sensors & Transducers (Canada) Via Acquire Media NewsEdge) Abstract: For failure prediction and information interaction problems of multi-equipment health management, fault prediction technology of multi-equipment and multi-parameter was proposed based on the system network. The state judgment and fault diagnosis algorithm is adopted by SOM self-organizing feature map neural network, the network parameters and node structure are changed adaptively, and the real-time updating of running status of equipment fault detection is realized. The improved fault prediction algorithm based on Elman feedback artificial neural network promoted the characteristics of the approximate any nonlinear function with arbitrary precision. Referencing to historical data by the feedback for the health management of multi-equipment, the algorithm provided early detection, isolation, management and forecast for fault omen, incipient fault status and ancillary component failure state in multi-equipment health management. The self adaptive abilities and the robustness of fault prediction system are improved effectively. Copyright © 2013 IFSA. Keywords: System network, Failure diagnosis, Failure prediction, Artificial neural network, Self-organizing feature map. (ProQuest: ... denotes formulae omitted.) 1. Introduction Mechanical equipment fault diagnosis and forecast is an essential link in the process of development and use. With the wide application of new technology, the more sophisticated and smart modem mechanical system should be applied. The cost and difficulty also will be increased in equipments, storage, maintenance and other aspects. Some installation and debugging process requirements are relatively high in certain precision machinery equipment systems which adopt microelectronics technology, computer technology and sensitive components technologies. Some equipment failures probable have bred not only in use, but also during the process of production, transportation, and maintenance. The prediction parameter algorithm is the key of fault forecast. Traditional equipment fault forecast technology is usually based on expert experience. The limitation includes some aspects as follows [1], Firstly, the tradition fault detection and diagnosis are limited to "failure", lacked of the accumulation and analysis of the daily data processing. The traditional fault forecast just predicts only in the event of a failure judgment, but not track the whole health condition of the equipment entire life cycle. Secondly, the fault diagnosis methods based on internet is mainly remote data transmission through the network, lack of real-time, field data collection using the network technology. Thirdly, the traditional fault diagnosis methods only contrapose a single mechanical equipment and system, or contrapose only some local parameters of the equipment system in a certain stage. It is very effectively discovering and solving the single equipment failures using this method, but lacking of systematicness for the whole system life cycle. Prognostic and Health Management (PHM) [2] put forward a new solution of health status management by synthesized utilization of modem information technology and artificial intelligence technology. Mechanical equipment fault diagnosis and forecast are the important part of core content in health management. Fault prediction is the technology and means providing early detection, isolation capabilities, manage and forecast, according to the omen, incipient fault and fault condition of subsidiary component failure state [3]. This paper puts forward a new equipment fault prediction system based on system network in order to solve the problem of information interaction in PHM. The improved fault detection and prediction algorithms are more suitable for the mechanical equipment fault detection systems which need more test parameters, high real-time requirements and long running. The algorithms would facilitate system fault analysis and the comprehensive decision, achieve the whole life cycle health management in mechanical equipments. 2. Framework of Equipment Failure Prediction System Based on Internet of Things The System network technology is a network that makes any articles connect to the Internet with the specified protocol for information exchange and communication in order to achieve articles intelligent identification, positioning tracking and management by using RFID, infrared sensors, GPS, laser scanners and other information sensing device [4]. It is an extension and expansion of the Internet technology. There are 3 levels of management in the system structure. The Network Management Center (NMC) located in the top level is responsible for the fault management of whole network, multi-domain fault diagnosis, and addition and quiting of Domain Management Agency (DMA). NMC also charged with the task of maintenance and transmission of diagnostic model library, accepted DMA query and organization collaboration. DMA is the second level next to the NMC management level, corresponding to the deployment of the autonomous domain management node of system network and responsible for generating dynamically Mobile Agent (MA) according to the predetermined strategy, collecting the local field multi-equipment fault state and decision-making level of diagnosis. DMA generates fault prediction and diagnosis tasks of low level, and implements in local. The most basic fault management level is formed of MA in the form of code which is embedded in the key nodes. It performs basic monitoring, interactive, predictive and diagnostic operations. The system structure is shown in Fig. 1. 3. Fault Detection for Self-organizing Neural Network The Multi-equipment health management system based on the system network can continuous abtion testing data from sensors in state judgment function part, process and analysis, real-time monitoring the working condition of mechanical equipments. Therefore, the fault detection algorithm must be able to receive the detection data quickly, analysis and processing, and judge working status of the equipments immediately. Contraposing a certain type of machinery and equipment, failure detection algorithm improved in this paper must identify failure mode and type, determine the corresponding fault cause based on the numerical change of tested parameters, and also determine directly that the equipments are in normal working state or a failure condition. The algorithm must detect the corresponding failure mode, fault type, fault location, fault maintenance and recovery suggestion if a fault occurs. 3.1. Self-Organizing Neural Network The advantage of Self-organizing Feature Map (SOM) which adopts unsupervised learning method is that it can automatically search for the essential property and inherent law of samples, and selfadaptively change the network parameters and the node structure through self organizing [5]. SOM is able to conduct fault diagnosis of multi-equipment in a more intelligent way, also can be improved continuously to judge more accurately the condition and fault of equipments with the further use of equipment health management system. In this way, SOM provides the real-time data of the operation status of the equipment for the failure prediction [6]. The neural network of SOM is composed of input layer and mapping layer (output layer) (Fig. 2) [7]. The number of neuron nodes in input layer is m, while the mapping layer is a two-dimensional planar array containing a times b neuron nodes and each neuron node is linked to all neuron nodes which in the input layer. SOM network is capable of finding out the similarity automatically between input data, and allocating the similar input data in the network according to the principle of proximity, which will form a neural network that can respond to the input data selectively. SOM network learning algorithm can be divided into the following steps: 1) Initialization network. The initialization of SOM neural network is to set the initial value of the weight between mapping layer and input layer with random numbers, and set the smaller weight value to connection weight from input nodes to output nodes. Then choose j adjacent nodes from output neuron nodes to form the set Sj , in which Sj (/) indicates the set of j adjacent nodes at the time of t, and the set Sj (i) will be smaller as time goes. 2) Input vector. Input vector to the input layer. X = {_xx,x2,x"...,xm)T xcX xcn 3) Calculation Euclidean distance between the weight vectors of mapping layer and the vectors of input layer. Formula (1) can be adopted for calculating the Euclidean distance between the weight vector Wj of the j* neuron node in the mapping layer and the input vector X in the input layer. ... (1) where wtJ means the weight between the ith neuron node in the input layer and the j01 neuron node in the mapping layer. Then calculate to get the node with the shortest distance, noting as f so as to make sure that for any j, the node k can conform to the equation (2), and then get the set of neuron nodes adjacent to the node. ... (2) 4) Adjusting the weight. Calculate the weight of the output neuron node f and its adjacent nodes according to equation (3). ... (3) where 0 < rj < 1 , and rj is the constant, which decreases from 1 to 0 as times goes. ... (4) 5) Calculation the output. Calculate with equation (5), in which /(*) is a function between 0 to 1 or other nonlinear function. ... (5) 6) Judge if the result meets the preset requirement. Finish the algorithm if the result can meet the preset requirement; otherwise, return to Step (2), and continue the next round of learning. 3.2. Fault Detecting Algorithm Based on the requirement of equipment detecting and management, it searches for failure modes, failure variety, failure symptom, causes of failure, etc. And then show the failure symptom of the equipments with the eight basic detecting parameters of working current, working voltage, starting voltage, transmitting frequency, receiving frequency, sensibility, noise and temperature by taking into consideration of the found data and information. Therefore, there are eight neuron nodes in the input layer of SOM network model, which are signified withP\, Pi, ... P8? respectively. Based on the cause of failure corresponding to the failure symptom, design the competitive layer of SOM network as a 6 times 6 two-dimensional node arrays. Present a sample set containing various failure samples based on the collected failure samples, as shown in Table 1. Achieve the state judgment and fault diagnosis algorithm based on SOM with codes, in which the code of SOM network is achieved with MATLAB. To the established network, train for 10, 30, 50, 100, 200, 500 and 1000 times respectively. The result of the classification after training is shown in Fig. 3. Table 2 is the results of seven training classification for SOM network. It can be seen that when the training reaches 10 times, the samples can only be classified into 3 types, in which Type 1, 2, 5, 6 and 8 belong to the first type, while Type 3 and 7 belong to the second type, and Type 4 forms the third type. In the same, when the training reaches 30 times, the samples are divided into 7 types; and when it reaches 50 times, the samples are divided into 6 types. When the training is less than 100 times, the samples cannot be fully classified, as several samples are overlap. However, when the training reaches 100 times, every sample can be classified as different types. With the frequency of training increasing to 500 or 1000, the result is classified clearly. Obviously, under this condition, it is none of significance to increase the frequency of training. Therefore, treat the network which after 100 times of training as the SOM network ultimately. The classification of each neuron in the SOM network after training is shown as Fig. 4, in which the blue nodes indicate the neuron nodes which win the in the competition. The topology structure of the network is shown as Fig. 5, in which the competitive layer of the SOM network is composed of 6 times 6 neuron nodes. The direct distance between the adjacent neuron nodes is shown as Fig. 6, in which the blue indicates neuron nodes, the red lines indicate the link between nodes, and the color of each diamond indicates the distance between nodes, as to which the deeper the color, the farther the distance. 4. Failure Prediction Based on Elman Artificial Neural Network Failure prediction is to predict the possible failure in the future by reading the detecting data about the working condition of the system or equipment in a recent period, analyzing comprehensively the detecting parameters during the period, stimulating the trend of the data about each detecting parameter, predicting the detecting parameters at a certain time in the future with the help of algorithms and then analyzing and processing the data got from the prediction [8, 9]. 4.1. Algorithm Choosing The input of feedback artificial neural network contains delayed feedback on input or output data; the learning process of this network is the changing process as the state of its neuron nodes [10]. A steady and constant state of a neuron node will be reached after the learning process. Elman neural network is a common feedback artificial neural network, which includes input layer, hidden layer, acceptor layer and output layer, as shown in Fig. 7. In this network, the links among input layer, hidden layer and output layer are similar to that feed forward artificial neural network. The nodes in the input layer are for transmitting signals; the transfer function of nodes in the hidden layer can adopt a linear function or a nonlinear function; the acceptor layer (competitive layer) is for caching the output value of the nodes in the hidden layer at the previous moment and returning the value to the input layer; and the nodes in the output layer are for node weighting. The space expression is as follows: ... (6) ... (7) ... (8) where, u is the R-dimension input vector of the input layer, x is the N-dimension unit vector of the hidden layer (middle layer), x. is the N-dimension feedback state vector, y is the N-dimension output vector of the output layer; w is the connection weight between the acceptor layer and the middle layer, w2 is the connection weight from the input layer to the middle layer; w3 is the connection weight from the middle layer to the output layer. g(*) is the transfer function for transmitting neuron nodes, and /(*) is the transfer function for transmitting the neuron nodes in the middle layer. The advantage of Elman neural network is its acceptor layer which feedback the output of the hidden layer to the input of the hidden layer through delay and storage [11, 12]. This self-connected mode makes the network sensitive to the history data. The internal feedback enhanced the network's capability of processing dynamic information, so as to achieve dynamic modeling and approach nonlinear mapping with any precision, during which process the influence of external noise on the system can even be neglected [13, 14]. 4.2. Failure Prediction of Elman Artificial Neural Network Predicting failure based on Elman artificial neural network is to improve the self-learning and selfadaptive capability of the system and promote the robustness of the prediction system by taking advantage of its delayed input as well as its features of approaching nonlinear functions with any precision and its ability of referring to the history learning data in terms of its feedback [15, 16]. The algorithm flow of failure prediction is shown in Fig. 8. First select the detecting data of a period of time as the test sample, and then design and test the algorithm for the detecting data obtained at the same moment of nine successive samples. According to the flow, read the data and load it into the program. Execute the program, take every three groups of detecting data as the input, and take the forth group of detecting data as the target output. It can get five training samples. Input the five groups of samples into Elman network in a circular way, and conduct training repeatedly. After finishing the training, take the ninth group of detecting data as the test sample to test the availability of Elman artificial neural network, and judge whether it can predict the future detecting data correctly. The equipment test sample data are shown in Table 3, and the normalization data are shown in Table 4. For selecting the proper number of neuron nodes in the hidden layer, set the number of neuron nodes in the hidden layer into two groups, which are 10, 20, 30, 40 and 40, 50, 60, 70 respectively; then train and test the neural network and get the result of prediction, then analyze the prediction error of each network. The result of prediction comparison is shown in Fig. 9. From the comparison of the above simulation result, it can be seen that when the number of neuron nodes in the hidden layer reaches 50, the prediction error for each parameter is the smallest. Therefore, 50 is determined as the number of neuron nodes in the hidden layer while adopting Elman artificial neural network to conduct failure prediction for the system. Failure prediction includes the prediction about the parameter of the system or equipment at a certain moment in the future and the judgment on the further state of the system or equipment. The prediction algorithm based on Elman neural network can accomplish prediction about the value of parameters at a certain future moment, while the algorithm based on SOM can accomplish the judgment on the predicted state of equipment. 5. Conclusions Based on the actual needs and conditions of the system equipment, this paper has built the multiequipment health management system based on the system network, designed using adaptive neural network SOM fault detection algorithm and Elman neural network fault prediction algorithm based on the System network technology. This design is based on SOM network fault detection algorithm to analyze the collected test parameters and make the real-time operating status judgments. Its self-learning feature can adapt to the change of device parameters and improve detection accuracy. Fault prediction algorithm based on Elman neural network can predict parameters data in the lower range of error according to the detecting data and provide the basis for system failure prediction. References [1]. X. H. Shi, Y. C. Liang, X. Xu, An improved Elman model and recurrent back-propagation control neural networks, Journal of Software, Vol. 14, Issue 6, 2003, pp. 1110-1119. [2]. Sun Bo, Kang Rui, Xie Jinsong, Fault prognostics and health management system research and application overview, Systems Engineering and Electronics, Vol. 29, Issue 10,2007. [3]. Li, Chen Xi, Wu Fan, et al., Information aggregation technique for smart grid networking, Information and Communication Technology, Vol. 2,2010, pp. 21-28. [4]. M. J. Roemer, G. J. Kacprzynski, Advanced diagnostics and prognostics for gas turbine engine risk assessment, in Proceedings of the IEEE Aerospace Conference, Big Sky, MT, 18 -25 March 2000, Vol. 6,2000, pp. 345-353. [5]. Guangfan Zhang, Seungkoo Lee, Nicholas Propes, Yongsheng Zhao, George Vachtsevanos, Ash Thakker, Tom Galie, A novel architecture for an integrated fault diagnostic/prognostic system, in Proceedings of AAAI Spring Symposium, Palo Alto, CA, March 25-27,2002, pp. 2468-2479. [6]. B. Nickerson, R. Lally, Development of a smart wireless net-workable sensor for aircraft engine health management, in Proceedings of the IEEE Aerospace Conference, 10-17 March 2001, Vol. 7, pp. 3255-3262. [7]. L.P. Khoo, S. B. Tor, J. R. Li, A rough set approach to the ordering of basic events in a fault tree for fault diagnosis, The International Journal of Advanced Manufacturing Technology, Vol. 17, 2001, pp. 769-774. [8]. A. Hess, The joint strike fighter (JSF) prognostics and health management, in Proceedings of 4th Annual System Engineering Conference, 2001. [9]. Yang Wenping, Chen Guoding, Shi Boqiang, et al., Steyr automobile engine fault diagnosis based on fractal theory, Mechanical Engineering, Vol. 38, Issue 2,2002, pp. 49-52. [10]. Wang Mianyu, Process fault diagnosis based on graph theory, Zhejiang Science, Hangzhou, 2002. [11]. Wang Xue, Zuo Wei, Zhao Guohua, Agile virtual test diagnosis system based on multi-sensor integration fusion technology, China Mechanical Engineering, Vol. 12, Issue 8,2001, pp. 914-918. [12]. Zeng Shengkui, G. P, Michael, Wu Ji, Status and development of fault prognostics and health management (PHM) technology, Aeronautics, Vol. 26, Issue 5,2006, pp. 626-632. [13]. Zhou Yuyang, Chen Hanping, Wang Weizhe, et al., Troubleshooting gray mathematical model, Chinese CSEE, Vol. 22, Issue 6,2002, pp. 146-151. [14]. X. C. Cheng, X. Chen, P. Guo, Software reliability prediction with an improved elman network model, Journal on Communications, Vol. 32, No. 4, 2011, pp. 86-93. [15]. He Yong, Chen Gan, Predictive failure analysis based on BP neural network model, Microcomputer Information, Vol. 22, Issue 6,2006, pp. 29-34. [16]. Ma Chunyang, Li Guo, Equipment failure prediction based on frizzy neural network, Noise and Vibration Control, Vol. 6,2006, pp. 23-29. 1 Fayu SUN,1 Lei GAO,1 Jinlong ZOU,2 Tao WU, 3 Jing LI 1 Science and Technology on Electromechanical Dynamic Control Laboratory, Xian, Shaanxi 710065, China 2Computer Science School, Xi'an Polytechnic University, Xi'an, Shaanxi 710048, China 3School of Electronic Information Engineering, Xi'an Technological University, Xi'an, Shaanxi 710032, China 1 E-mail: [email protected] Received: 19 August 2013 /Accepted: 25 October 2013 /Published: 25 November 2013 (c) 2013 International Frequency Sensor Association |