Soft Redundant Instrument for Metering Station in Gas Transportation System

Transcription

1 Soft Redundant Instrument for Metering Station in Gas Transportation System N.S. Rosli, R. Ibrahim, I. Ismail Electrical and Electronic Engineering Department, Universiti Teknologi PETRONAS, Bandar Seri Iskandar, Perak, Malaysia. Abstract- This study focuses on the development of a soft redundant instrument to replace secondary hard instrument for saving cost of installation. The model of the soft redundant instrument implements an artificial neural network (ANN) approach; providing the alternative for secondary measurement. Moreover, the reliability and quality of the metering station play a crucial part in the gas transportation system as it affects the billing integrity between the gas supplier and their customers. In this work, the problem with a primary instrument was analysed to diagnose the faulty condition of measurement in metering station. Accordingly, different ANN algorithms were investigated and compared to select a prediction model that provides the best performance. This suggested prediction technique can be considered as an additional verification system to be compared to the hard instrument that is installed on the pipeline. Hence, in order to determine accuracy in the rate of measurement from the proposed model, it is required that variation in measurement between hard and soft instrument should not exceed 1%. Based on this model, the ANN prediction model is able to reconstruct the unreliable data during faulty instrument. Index Terms Metering Station; Instrument Fault; Fault Diagnosis; Neural Network Model; measurement of gas metering station in Kapar Power Plant. This plant consists of a double piping system with records of two runs for each pipeline. The instruments used include temperature transmitter, flow transmitter, pressure transmitter, turbine meter and gas chromatography [4]. In addition, the plant is using ultrasonic meters to measure flow of natural gas. Then the measurements are used to compute the energy of gas supplied for customer billing. The implementation of a fault diagnosis system is proposed to replace the operation of the current system, which has a conflict during a distressed condition that affects billing integrity. In this study, the focus is on developing a soft redundant instrument in order to back up the hard instrument at the time of malfunctioning. Moreover, the secondary instrument is used as a benchmark for testing. This can also justify the model s performance, which is important in getting an accurate prediction to save costs by reducing secondary hard instrument installation cost and in increasing business credibility. Figure 1 below shows the diagram of developing a soft redundant instrument. I. INTRODUCTION Generally, the instrumentation and measurement are important components that cannot be separated in order for a system to work at normal operating conditions. However, in practical applications, these systems may be subjected to abnormal conditions. According to Korbiczin, in engineering terms, these abnormalities are referred to as faults [1]. Faults in an engineering system may lead to some measurement inaccuracies of a process. It may also cause the process to shut down. Beside, fault diagnosis refers to the prompt identification and analysis of system abnormalities. Nevertheless, the early detection of these faults is important to ensure reliability, safety and efficiency of the process [2]. There are many different types of fault diagnoses that have been applied in the industry. Besides that, the neural network also well known in numerous applications such as in automotive, banking, medical, transportation, and so on. In this study, the implementation of Artificial Neural Network (ANN) is explored for industrial process control. It is of utmost importance that gas meters are calibrated accurately, especially in the oil and gas industry [3]. The measurement of natural gas is not only important for billing purpose but also for control and indication of process condition. This research is investigating the accurate Figure 1: Development of Soft Redundant Instrument II. RESEARCH BACKGROUND The metering system has three essential transmitters; which are Flow meter (FT), Pressure Transmitter (PT) and Temperature Transmitter (TT). These instruments measure volume (V g(meas)), pressure (P meas) and temperature (T meas) of transmitting gas in the pipeline. The objective of this research is to focus on analyzing the behavior of pressure instruments during their operation as well as the relationship between them in order to detect the fault and to correct any faulty data reading obtained from the instruments. This study aims at solving the problem of only one faulty instrument. The worst case scenario happen when two or more instruments are detected to be faulty at the same time. Therefore, the action is taken to overcome only one faulty instrument. Table 1 below demonstrates the frequency of faults in an instrument /15/$ IEEE

2 TABLE 1: Percentage of fault data in the primary instrument Type of Fault Percentage of fault (%) One instrument fault only 9.40 Two instrument's fault 0.31 All three instruments faults 0.30 The total number of data samples used for testing was 6408 that is equivalent to one-year sampling data. These data have been divided into two categories, which are healthy data and fault data. The faulty data has been further categorized into 3 types of common faults in accordance with the frequency of fault data as shown in Table 2. Figure 3: Hang Fault TABLE 2: Percentages of fault data in the primary instrument Type Fault Power Cut-Off Hang Fault Missing Data of Description of Fault Supervision system reads zero value from instruments when the plant is shutdown The readings show a constant value for the next sampling data The data does not have any reading even the plant is in operation mode Percentage of fault data in one year (%) Based on the table above, the most serious case which happens mostly in the primary instrument is the hang fault. The data reliability is required even during the malfunctioning, for billing purpose. However, this research emphasize on solving the most common problem occurring during the operation of metering system. Whereas, the phenomenon shown above are common fault conditions occurred in metering system. For this research, the secondary value is considered as a good value for testing purpose. This is because the secondary instrument work as back up for main instrument when the primary instrument was fault. Therefore, the prediction model is used to verify the primary value. The secondary instrument (testing) is validated when the faulty condition occurs for a very small percentage. The prediction model of the soft redundant instrument is compared with the actual value of the secondary instrument. The Table 2 above also illustrate the reliability and availability of soft redundant instrument. For a clearer description of these fault conditions, Figure 2 to 4 shows the phenomenon that commonly occurs in instrumentation process control. Figure 4: Missing Data The Table 3 below shows the comparison of fault quantity in the primary and secondary instrument. Type of Fault TABLE 3: Percentage of fault data in gas metering system Primary Reading Secondary Reading Both Instrument Power Cutoff Hang Fault Missing Data The objectives of this research are listed as below: To design the soft redundant instrument for validation purpose of hard instrument. To verify and compare the soft redundant instrument with testing instrument so that it can be proven as a reliable verification tool. To develop a prediction model that will produce accurate reading of less than one percent error. III. DESIGN OF SOFT REDUNDANT INSTRUMENT Figure 2: Power Cut-Off For designing the prediction model, the system must have the capability to learn the provided historical data effectively. The main concern of faulty conditions is that the prediction model must be able to predict in short term load, in large processing system. Therefore, Artificial Neural Network (ANN) method is proposed as the intelligent prediction model to control the highly uncertain system of non-linear process or behavior. In developing ANN architecture that determines the performance of the model in term of its reliability and robustness of the system, the factors that affect the development of model are model inputs, data preprocessing, parameter estimation and model validation [5].

3 There are three classifications of network architectures; Single-Layer Feedforward Networks (SLN), Multilayer Feedforward Networks also known as Multilayer Perceptron (MLP) and Recurrent Networks (RN). The features, operation and application of these architectures are explained in detail in Reference [6]. Generally, the most popular network architecture is the MLP network. While, in load forecasting applications, the basic form of multilayer feed-forward architecture is still the most popular [7]. On the other hand, the architecture that is most suitable for the ANN model in engineers view is MLP, which is proven to be working fast, as time is the most important factor in producing good product quality. The simulation results of ANN model are further discussed in Section IV. First of all, the way of gathering the data is crucial to present the input to the network that can process the data. After that, all inputs are scaled or normalised to [-1, 1] range to have reasonable maximum and minimum values for each input type. This neural network architecture is used to predict the short-term instrument input, which is the hourly forecasting. Data was acquired for the duration of one year on the hourly basis for each day. The obtained data were then sorted on hourly basis and further data were filtered out to remove the faulty or irrelevant data. The average data obtained, after filtering was performed for every month, was 550 data samples. The data for one year (January December) consecutively, was trained and validated in this experiment. There are 6408 data samples. All combinations of inputs were investigated and the performance of the model was defined. 50% of the data was taken as the training set and the remaining 50% of the data was for the validation set. There were few methods utilised in choosing the number of inputs for ANN model. Many ANNs were trained using various inputs. This was done to determine which factors were the most significant to the ANN model [7]. Generally, the ways of determining the model inputs are based on prior knowledge of causal variables together with assessment of potential inputs and outputs in time series plots [8], cross-validation technique or principal component analysis technique that generalises the independent data set using statistical analysis [9], or the stepwise approach, where the selection of input is based on evaluation performance of the initial model with different combinations of input variables and by removing the input variables which are not beneficial to the model [10]. Besides that, there is a sequential forward method to identify the best input for the model [11]. In this research, the inputs are selected from a sequential forward method which is based on the least error of the predicted value as compared to the actual value (secondary). By iteratively adding the input variable to the model, the Table 4 shows the input variables with other different input. This metering system is using the ultrasonic meter where measures pressure and temperature transmitter. In addition, gas chromatography measures the volume of gas and flow computer calculates the energy consumption of the gas produced that has been sold. Therefore, for this system, the input variables of the neural network prediction model is past one hour of pressure P(k-1), current temperature T(k-1) and current volume V(k) with the output pressure (), which represent the variable to predict the output layer. TABLE 4: Selection of Input Variable Categories of Input Variable Past input Combination of past input and other parameter Combination of several input Other parameter only Parameter of Input P(k-1) P(k-1)V(k), P(K-1) T(k) P(k-1)V(k)T(k) V(k)T(k) An optimal architecture was aimed at getting a better and efficient network to fit the model problem. When dealing with these kinds of problems, the important aspect to be considered is the time. Nevertheless, the speed of information learning and retrieved data processing can affect the ability to correlate and interconnect the source data [12]. A few optimization parameters are taken into account such as weightages, number of hidden layer and nodes and number of input variables. In this study, the two major techniques are considered to get the best model are pruning and constructive methods. The pruning method starts with a big network and gradually removes the undesired nodes or connections. While constructive method begins with a small network and gradually increases the number of nodes or connections as per requirement [13]. For the purpose of this research, the constructive method was employed. If a small number of neurons have a problem, slightly higher number of neurons was tried. The backpropagation learning method of the network was chosen as it is most frequently used in network training. A few network architectures were simulated and the results were compared to choose the best architecture that could be used for ANN model. The training of the network is a loop of calculations and will stop the process once the validation performance goal is achieved. The network will calculate the output, weight, error to repeatedly adjust until all targets are reached. Therefore, the successful performance of a network is presented in a minimum error percentage form, when no offset is added to the input. A slightly higher offset is produced by the model; the prediction result will also have higher error. MLP is the selected architecture to be simulated due to its good performance and fairly low root mean square error (). The formula that is used to compute is:

4 IV. RESULTS AND DISCUSSIONS A. Neural Network Model for Soft Redundant Instrument The soft redundant instrument that is integrated with the neural network is tested and simulated using MLP network. In this network, the parameter optimisation is done by selecting optimal number of hidden layers, number of hidden neurons, activation function and training algorithm. The number of neurons chosen is 10, which offers the best performance. Based on the result, the pressure of transmitting gas is estimated. The Table 5 below shows the selection of the best input for the instrument. TABLE 5: Development of Neural Network Prediction Model Input Selection (validation data) (training data) P(k-1) P(k-1),V(k) P(k-1),T(k) P(k-1),T(k),V(k) T(k),V(k) It is shown that the relationship of past pressure and current volume give the least as compared to other combination of inputs. This is because the pressure and volume are strongly interrelated with each other. While the other combinations have less impact on the model, since the model does not have the ability to learn the relationship and behavior of input and output variables among instruments. After the selection of the input, the parameters of the network are simulated to achieve the least error percentage. Furthermore, two training algorithms are compared for network training parameter i.e., Lavenberg-Marquardt and Bayesian Regularisation. Lavenberg-Marquardt has proven to be the faster training function, whereas Bayesian Regularisation improves generalization. These two training functions were selected to compare with each other as they are more popular and considered as good training algorithms. However, the default activation function for the output layer is pure linear because the transfer function is linear, so that the output will be equal to the target value. Table 6 shows the comparison of network parameters to choose the best performance of the model based on the best input selection of past pressure and current temperature from Table 5. Training Algorithm Levenberg- Marquardt Bayesian Regularization TABLE 6: Comparison of network parameter Activation Function (hidden layer) (validati on data) (training data) Tan-sigmoid Log-sigmoid Tan-sigmoid Log-sigmoid Based on the results shown, the best performance obtained when the Lavenberg-Marquardt training algorithm is applied together with tangent-sigmoid for hidden layer activation function. The choice of the transfer function for the hidden layer constraints the output to a nonlinear range of -1 to 1, according to the problem that needs to be solved. Table 7 below summarizes the parameter of neural network model. TABLE 7: Parameters of Neural Network Prediction Model Network Parameter Result No. of Hidden layer 2 No. of neurons 10 Data Division 50/50 Learning Algorithm Lavenberg-Marquardt Activation Function Tangent Sigmoid function Based on the best performance of the model development, the soft redundant instrument is used to validate the accuracy of the model. The interface for training the best model is shown in Figure 5 while Figure 6 and 7 show the graph of actual and predicted pressure for training and validation data, respectively. Figure 5: Graphical User Interface for ANN Training Model Pressure Prediction during Fault Input Instrument (Training Data) Predicted Pressure Actual Pressure No of Data Figure 6: Performance of Neural Network Prediction Model (Training Data) Pressure Prediction during Fault Input Instrument (Validation Data) Predicted Pressure Actual Pressure No of Data Figure 7: Performance of Neural Network Prediction Model (Validation Data)

5 After training and validating the soft redundant instrument, modelling is done using Simulink to show that the hypothesis is accepted since the predicted value is close to the secondary reading when the primary reading instrument was in faulty condition. Figure 8 shows the Simulink block diagram for testing the neural network prediction model, while Figure 8 shows the graph of a comparison between predicted value, primary and secondary reading when the primary reading experienced hang fault. probability of occurrence of these problems, the model should be trained again to get the best model that could handle these situations. V. CONCLUSIONS ANN architecture is considered as good model when the prediction provides the least error once compared to the actual value. The soft redundant instrument is designed using neural network prediction models, and also verified in order to use for validation purpose of the hard instrument due to the fact that the prediction model produces accurate readings of less than 1% error. Therefore, it can be applied to replace the redundancy hard instrument, which has the same function. It can save cost with respect to billing and installation, when there is only one hard instrument without the other unnecessary hard verification tool. REFERENCES Figure 7: Simulink Block Diagram of Neural Network Prediction Comparison Between Predicted Value, Primary and Secondary Reading Predicted Value Primary Reading Secondary Reading Time (H) Figure 8: Comparison of Soft Redundant Software Prediction with Primary and Secondary Value during Hang Fault By using this soft redundant instrument, the prediction errors are found to be higher since the uncertainties were higher as well. This happens when the model predicts on the basis of the previously predicted output. Therefore, the prediction would be better when the model predicts multistep ahead and has a good historical training data. Hence, more data around 50% added data should be trained so that the model could experience in advance a pattern of input as well as the trending of the input. This is also because the inputs of the model are related to the time of daily hours. On the other hand, the faulty data could happen because of weather or instrument failure. Therefore, the model should be tested for its robustness in order to ensure the reliability of the model during an upset condition. For power cut-off and missing data problem, the frequency of the problems is very small; hence the prediction still can encounter the value and do not have much impact on the instrument failure. On the occasion of the [1] Korbicz, J., Koscielny, J. M., Kowalczuk, Z., & Cholewa, W. (2004). Fault Diagnosis Models, Artificial Intelligence, Applications. New York: Springer. [2] Isermann, R. (2005). Model-based Fault-detection and Diagnosis Status and Applications. Annual Reviews in Control, 29, pp [3] Mohsin, R., & Nasri, N. S. (1995). Calibration concept and equipment drifting factors: its application in gas metering facilities. Proceedings of The Eleventh Symposium of Malaysia Chemical Engineers. C15-1. [4] Rosli, N.S.; Ibrahim, R.; Nguyen Tuan Hung; Ismail, I, "Intelligent fault diagnosis for instrument in gas transportation system," Intelligent and Advanced Systems (ICIAS), th International Conference on, vol., no., pp.1,6, 3-5 June 2014 [5] H.R. Maier and G.C. Dandy, Neural network for the prediction and forecasting of water resources variables: a review of modelling issues and applications, Environmental Modelling and Software, vol. 15, 2000, pp [6] S. Haykin, Neural Network A Comprehensive Foundation, Hamilton, Ontario, Canada: Prentice Hall International. Inc., [7] H. Hippert, C. Pedreira, and R. Souza, Neural networks for short-term load forecasting: a review and evaluation, Power Systems, IEEE Transactions on, vol. 16, 2001, pp [8] R. Brown and I. Matin, Development of artificial neural network models to predict daily gas consumption, Industrial Electronics, Control, and Instrumentation, 1995., Proceedings of the 1995 IEEE IECON 21st International Conference on, 1995, pp vol.2. [9] U. Anders and O. Korn, Model selection in neural networks, Neural Networks, vol. 12, Mar. 1999, pp [10] D.A Linkens and Min-You Chen, Input selection and partition validation for fuzzy modelling using neural network, Fuzzy Sets and Systems, vol. 107, 1999, pp [11] D. Ververidis, C. Kotropoulos. "Sequential forward feature selection with low computational cost." Proceedings of the 8th European Signal Processing Conference [12] Z. Reitermanova, Feedforwad Neural Networks-Architecture Optimization and Knowledge Extraction, WDS'08 Proceedings of Contributed Papers, vol. Part I, 2008, pp [13] G. Bebis and M. Georgiopoulos, Feed-forward neural networks, Potentials, IEEE, vol. 13, 1994, pp