Asian Research Consortium Asian Journal of Research in Social Sciences and Humanities Vol. 6, No. 11, November 2016, pp. 26-40. Asian Journal of Research in Social Sciences and Humanities ISSN 2249-7315 A Journal Indexed in Indian Citation Index DOI NUMBER:10.5958/2249-7315.2016.01173.4 Category:Science and Technology www.aijsh.com Fault Detection and Diagnosis of Spiral Type Heat Exchanger using ANN N. Bagyalakshmi*; Dr. M. Thirumarimurugan** Abstract *Assistant Professor, Department of EIE, Adhiyamaan College of Engineering, Hosur, India. **Professor, Department of Chemical Engineering, Coimbatore Institute of Technology, Coimbatore, India. Fault Detection and Diagnosis (FDD) is important in many industries to provide safe operation of a process. Actuator faults, sensor faults and process faults are the common faults occurring in chemical process. In this present work Sensor and Process faults of Spiral type heat exchanger is detected and diagnosed using ANN.NARX network (Nonlinear Autoregressive with External input) is used as ANN network structure. Network is trained using Levenberg, Bayesian and Scaled Conjugate Gradient Methods. To achieve FDD, a set of residuals is generated by ANN which indicates the state of the system and provide information about faults. Mean Square Error, Integral Absolute Error (IAE), Integral Time Absolute Error (ITAE) and Integral Square Error (ISE) is obtained for the above said methods which are shown in simulation results. Keywords: Fault, FDD, Error, Residual, NARX, Levenberg, Bayesian, Scaled Conjugate Gradient. 1. Introduction Spiral Heat Exchangers (SHE) refer to a helical tube configuration, which refers to a pair of flat surfaces that are coiled to form the two channels in a counter-flow arrangement. The main 26
advantage of the SHE is its highly efficient use of space. The term fault means that any unpermitted deviation occurring in a system.the faults present in the system affect the sensors, the actuators, or the system components [1]. Actuator fault, sensor fault and process fault are the common faults occurring in chemical process. Total actuator fault can occur as a result of a breakage, cut or burned wiring, shortcuts, or the presence of outer body in the actuator. Sensor faults represent incorrect reading from the sensors. Process faults of heat exchanger includes fouling, fault in volumetric flow rate etc. To identify and remove these types of faults in the system, Fault Detection and Diagnosis (FDD) techniques are proposed [1]. These techniques are generally classified as model-based approaches and data-driven approaches. Some of the model-based FDD techniques include observer-based approach, parity-space approach, and kalman based approach. Data driven approaches include Fuzzy logic, Artificial Neural Network (ANN) and Genetic Algorithm (GA). FDD techniques provide early warning to the system operators and prevents the system causing failures. ANN are used in various application areas such as fault detection and diagnosis, Pattern recognition, system identification[2],[3]. 2. Residual Generation To achieve FDD, a set of residuals need to be generated. The residual is defined as difference between the measured and estimated process output. To detect and diagnose the fault, FDD has to undergo two step process Residual generation and Residual evaluation as in fig 1. The Residual generator generates a residual and the Residual evaluator compares the residual to determine the occurrence of fault with a threshold [4]. In the ideal case, the residual will be equal to zero when no fault is present and different from zero when a fault is present. A well designed residual signal is defined such that it is equal to zero for fault free case and not equal to zero for faulty system[9]. Fig. 1. Block Diagram of Fault Detection and Isolation Techniques 27
3. Fault Detection and Diagnosis FDD performs two tasks Fault detection and Fault isolation. Fault detection is to determine whether the fault has occurred or not. The role of fault isolation is to locate and isolate the fault [4].In this work fault is detected and diagnosed by ANN. Artificial neural networks (ANNs) have the capability to learn the complex relationships between the inputs and the outputs of the system. The advantages of using a neural network in FDD is its ability to attain input-output mapping. Using input-output mapping a neural network is able to modify its weights by training samples. The training samples consist of an input signal and a desired response. During training the weights are modified in order to reduce the error between the desired response and actual response of the network. Fig 2 shows a general block diagram of ANN based Fault diagnosis. 4. Neural Network Configuration Fig. 2. ANN Based Fault Diagnosis ANN consists of number of interconnected units. The input characteristics and its interconnection with other units determines the output of ANN. ANN consists of Input layer, Output layer and hidden layer with a number of nodes in it. Input layer has no input weights and activation function. The output response for a given input is determined by the output layer. Hidden layer has no connection with outside world. Increasing the number of hidden layer increases the complexity of the network but it results in accurate results. For fault detection and diagnosis purposes, the ANN has to be trained first. Nonlinear Autoregressive with External Input(NARX) is used for training as it provide better results and predicts past values of input and output. NARX structure belongs to dynamic network which have feedback or recurrent connections with delay input. Implementation of NARX model is shown in Fig 3. 28
Fig. 3. Implementation of NARX Model 5. Levenberg-Marquardt (LM) Training The Levenberg-Marquardt (LM) algorithm is the most widely used optimization algorithm[12]. Its performance is more better than conjugate gradient methods. For detection and diagnosis of sensor and process faults of heat exchanger, input and output data is loaded in the neural network toolbox. Nonlinear Autoregressive with External Input(NARX) is used as a network structure to perform fault detection. trainlm is the training function for LM method which automatically update the weight and bias value.residual are generated for both sensor and process faults. Fig 2 shows the general block diagram of LM method for sensor and process faults of heat exchanger with PID controller. The simulation results using LM method are shown in Fig 4 to Fig 9. Fig. 4.Mean Square Error Graph of LM Method for Sensor Fault 29
Fig. 5. Error Graph of LM Method for Sensor Faults Fig. 6. Simulation Results of LM Method for Sensor Faults 30
Fig. 7.Mean Square Error Graph of LM Method for Process Fault Fig. 8. Error Graph of LM Method for Process Faults 31
Fig. 9. Simulation Results of LM Method for Process Faults 6. Bayesian Regularization Training An extension of the Levenberg-Marquardt algorithm is the Bayesian regularization[13]. One of the main problems with regularizing a neural network is, it leads to over fitting of the data or poor generalization of the network.. The network was trained in MATLAB by using Neural Network Toolbox. trainbr is the training function of Bayesian Regularization. This training function updates the weight and bias value and it minimizes a combination of squared errors and determines the correct combination to produce a network. It provides an efficient criterion for stopping training process and prevents overtraining of the network. The simulation results using BR method is shown in Fig 10 to Fig 15. Fig. 10.Mean Square Error Graph of Bayesian Regularization Methods for Sensor Fault 32
Fig. 11. Error Graph of BR Method for Sensor Fault Fig. 12. Simulation Results of Bayesian Regularization Method for Sensor Fault 33
Fig. 13. Mean Square Error Graph of Bayesian Regularization Methods for Process Fault Fig. 14. Error Graph of BR Method for Process Fault 34
Fig. 15. Simulation Results of Bayesian Regularization Method for Process Fault 7. Scaled Conjugate Gradient(SCG) Method Scaled Conjugate Gradient algorithm is a supervised learning algorithm. SCG is a batch learning method and there will be no effect if parameters are shuffled. trainscg is the training function for scaled conjugate gradient method[38].this algorithm takes only a little memory. Training automatically stops when generalization stops improving when there is increase in the mean square error of the validation samples. The simulation results for above faults using scaled conjugate gradient method is shown in Fig 16 and Fig 20. Fig. 16. Mean Square Error Graph of Scaled Conjugate Gradient Methods for Sensor Fault 35
Fig. 17. Error Graph of SCG Method for Sensor Fault Fig. 18. Simulation Results of Scaled Conjugate Gradient Method for Sensor Fault 36
Fig. 19. Mean Square Error Graph of Scaled Conjugate Gradient Methods for Process Fault Fig. 20. Error Graph of SCG Method for Process Fault 37
8. Results and Discussion The best neural network architecture is determined by the number and the size of hidden layer. LM provides improved accuracy than other algorithms. Levenberg-Marquardt algorithm is used to reduce the computational overhead where as Bayesian regularization algorithm reduces the long cross-validation. Various parameters of Levenberg-Marquardt and Bayesian regularization for sensor and process faults are compared which are shown in table T-1&T-2. LM methods have least mean square error when compared to Bayesian and scaled conjugate methods.br methods also provide good results but the only disadvantage is it take more time to converge. Integral Absolute Error (IAE), Integral Square Error (ISE) and Integral of Time and Absolute Error (ITAE) is calculated for the both sensor & process faults with PID controller and without PID controller and their comparative results are shown in table T-3 and T-4. Table1 Comparative Results of Training Algorithm for Sensor Results Table 2 Comparative Results of Training Algorithm for Process Results Parameters Levenberg- Bayesian Marquardt regularization Scaled conjugate gradient Number of hidden neuron 80 50 60 Delay 1 1 1 Training Function Trainlm Trainbr trinscg Training Mean Square Error 1.68113e-1 1.95446 8.2165e-1 Validation Mean square Error 12.18606 0.0000 68.89413 Testing Mean Square Error 15.79149 2.60331 34.87527 Epoch 2 26 28 Parameters Levenberg- Marquardt Bayesian regularization Scaled conjugate gradient Number of hidden neurons 70 40 50 Delay 1 1 1 Training Function Trainlm trainbr trinscg Training Mean Square Error 3.7411e-1 7.3327e-1 1.2119e-1 Validation Mean square Error 4.38944 0.0000 60.2486 Testing Mean Square Error 238.9651 3.4623e-1 7.3678 Epoch 1 67 30 Table 3 Error Calculation for Sensor and Process Faults with PID Training method Sensor fault Process fault ITAE IAE ISE ITAE IAE ISE Levenberg-Marquardt 205.5 102.8 5123 3691 1846 1.365e+05 Bayesian regularization 218.3 1091 4.765e+04 3420 1710 1.17e+05 Scaled conjugate gradient 2612 1306 6.824e+04 3497 1748 1.224e+05 38
Table 4 Error Calculation for Sensor and Process Faults without PID Training method Sensor fault Process fault ITAE IAE ISE ITAE IAE ISE Levenberg-Marquardt 111 55.49 153.4 2924 1462 8.585e+04 Bayesian regularization 1433 716.4 2053e+04 2670 1335 7.131e+04 Scaled conjugate gradient 1860 930.2 3.463e+04 2747 1373 7.558e+04 9. Conclusion Sensor fault and process fault for Spiral heat exchanger was detected and diagnosed using ANN. Training of ANN is done by Levenberg-Marquardt, Bayesian regularization algorithm & Scaled conjugate gradient method. Various parameters of network such as Mean Square Error, Number of hidden layer, Epoch,Integral Absolute Error (IAE), Integral Square Error (ISE) and Integral of Time and Absolute Error (ITAE) was compared for the above methods. These errors are comparatively less in LM method for sensor and process faults. Levenberg-Marquardt reduces computational overhead and Training Mean Square Error and Testing Mean Square Error and number of iterations are lesser and provide accurate results during training. Levenberg-Marquardt method shows good results than Bayesian method & Scaled conjugate gradient method. References Khoukhi, A., Khalid, H., Doraiswami, R., Cheded, L.: Fault Detection And Classification using kalman filter and hybrid neuro-fuzzy systems. In: International journal of computer application, Vol 45(2012), No 22, May 2012. Baligh Mnassri., Mostafa, El., Adel El., Ouladsine M.: Reconstruction -Based contribution approaches for improved fault diagnosis using principal component analysis. In: Elsevier,Journal Of Process Contro, Vol 33 (2015), 2015, p. 60-75. El Harabi R., Ould Bouamama B., Ben Gayed, M.K., Abdelkrim M.N.: Robust Fault Diagnosis Of Chemical System By Using Uncertain Bond Graph Model. In: IEEE,8th International Multi-Conference on Systems, Signals & Devices, November 6, 2011. Asokan, A., Sivakumar, D.: Model Based Fault Detection and Diagnosis Using Structured Residual Approach in a Multi-Input Multi-Output System. In: Serbian Journal Of Electrical Engineering, Vol 4 (2007), No 2, November 2007, p. 133-145. Hossein, M., Sobhani., Poshtan, J.: Fault Detection And Isolation Using Unknown Input Observers With Structured Residual Generation.In: International Journal of Instrumentation and Control Systems (IJICS),Vol 2 (2012), No 2, April 2012. Thumathi B.T., Miles Feinstein A., James Fonda W., Turnbull A., Fay Weaver J., Mark Calkins E., Jagannathan S.: An online model based fault diagnosis scheme for HVAC System. In: IEEE International conference on control Application (CCA), September 28-30,2011, Denver,CO,USA, 39
Svärd, C., Nyberg, M., Frisk, E., Krysander, M.: Data-Driven and Adaptive Statistical Residual Evaluation for Fault Detection with an Automotive Application. In: ELSEVIER, Mechanical systems and signal processing, Vol 45, No 1, p. 170-192. Guo L., Zhang Y.M., Wang H., Fang J.C.: Observer-Based Optimal Fault Detection and Diagnosis Using Conditional Probability Distributions. In: IEEE Transactions On Signal Processing, October 2006, Vol 54, No 10. Shukri Zainal Abidin M., YusofF R.: Application of a Model-based Fault Detection and Diagnosis using Parameter Estimation and Fuzzy Inference to a DC Servomotor. In: IEEE,International Symposium on Intelligent Control, October 2002, Vancouver, Canada. Aitouche A., Maquin D., Busson F.: Multiple Sensor Fault Detection in Heat Exchanger Systems. In: IEEE International Conference on Control Applications, CCA'98, September 1998, Trieste, Italy, p. 741-745. Jiang B., Chowdhury B.:Observer-based Fault Diagnosis for a Class of Nonlinear Systems.In: Proceeding of the 2004 American Control Conference on Boston, Massachusetts, June 30- July 2,2004. Messai, A., Mellit,A., Abdellani,I., Massi Pavan, A.: On-line fault detection of a fuel rod temperature measurement sensor in a nuclear reactor core using ANNs.In: ELSEVIER,Progress in Nuclear Energy 79 (2015),p.8-21. Sheela, T., Naresh, R.,Rameshwar, J.:Comparative Study of Back propagation algorithms In Neural Network Based Identification Of Power System.In: International Journal of Computer Science & Information Technology (IJCSIT) Vol 5, No 4, August 2013. 40